Add Batch b4c235ba-dfad-4c6c-8cfe-423580b04122

EMNLP/2025/MPO_ Boosting LLM Agents with Meta Plan Optimization/b37be3c6-9bde-4ea1-a520-d38fc852946b_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e120df3d9f9b46d9974f16cbfaabfe31b4d4a749aa73172a5077c295439818c0
+size 139681

EMNLP/2025/MPO_ Boosting LLM Agents with Meta Plan Optimization/b37be3c6-9bde-4ea1-a520-d38fc852946b_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb9f3152654b1b12ee782ca6b0ad29771bf6fe19d3d8a705319898f5869500ed
+size 169970

EMNLP/2025/MPO_ Boosting LLM Agents with Meta Plan Optimization/b37be3c6-9bde-4ea1-a520-d38fc852946b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f294d51ff04c7eb429a63113a6bb1ac2083dd4011d305791ef81c90b75e2a9c5
+size 1137428

EMNLP/2025/MPO_ Boosting LLM Agents with Meta Plan Optimization/full.md

@@ -0,0 +1,757 @@
+# Abstract
+
+Recent advancements in large language models (LLMs) have enabled LLM-based agents to successfully tackle interactive planning tasks. However, despite their successes, existing approaches often suffer from planning hallucinations and require retraining for each new agent. To address these challenges, we propose the Meta Plan Optimization (MPO) framework, which enhances agent planning capabilities by directly incorporating explicit guidance. Unlike previous methods that rely on complex knowledge, which either require significant human effort or lack of quality assurance, MPO leverages high-level general guidance through meta plans to assist agent planning and enables continuous optimization of the meta plans based on feedback from the agent's task execution. Our experiments conducted on three representative tasks demonstrate that MPO significantly outperforms existing baselines. Moreover, our analysis shows that MPO provides a portable solution that enhances both task completion efficiency and generalization capabilities across new agents and unseen scenarios.
+
+# 1 Introduction
+
+Recent advancements in large language models (LLMs) (Achiam et al., 2023; Liu et al., 2024; Yang et al., 2024a) have enabled LLM-based agents to tackle complex multi-step tasks, including embodied housework (Shridhar et al., 2020) and science experiments (Wang et al., 2022). These tasks require sophisticated planning abilities, as agents need to understand long-term dependencies (Zhang et al., 2024), reason about sequential actions, and adapt to dynamic environments (Yao et al., 2022b). The planning quality of these agents plays a crucial role in determining their overall performance.
+
+Current mainstream LLM-based agents primarily develop their planning capabilities through
+
+![](images/7bd273208bf633c852809a3c0cfd1d3d7c9160539cccf2f7678a60205f626810.jpg)
+
+![](images/0cb74d106bd1dcccf1039e3544ddbb1b897789a171526ee38a691e1c784ae055.jpg)  
+Figure 1: Unlike previous implicit plan enhancing methods that require agent parameter updates, our method incorporates meta plans into prompts for direct planning guidance and can improve them based on feedback.
+
+implicit methods, either directly leveraging the model's inner ability or fine-tuning from expert trajectories. For example, ReAct (Yao et al., 2022b) and Reflexion (Shinn et al., 2024) perform planning on-the-fly during task execution and are prone to getting lost due to planning hallucination (Zhu et al., 2024). The works including AgentTuning (Zeng et al., 2023), Lumos (Yin et al., 2023), and ETO (Song et al., 2024b) employ trajectory tuning to enhance implicit planning capabilities and require retraining for each new agent, resulting in huge computational cost (Figure 1(a)).
+
+Beyond implicit planning, recent studies have explored the use of human knowledge to guide agents in task execution, capitalizing on the benefits of explicit guidance and low integration costs of such knowledge (Zhu et al., 2024; Qiao et al., 2024). However, these approaches face significant challenges: they either require extensive manual efforts or struggle to ensure quality in the process of acquiring complex knowledge, potentially leading to inconsistent improvements in agent perfor
+
+mance (Wang et al., 2024). To overcome these limitations, we propose to automatically generate a high-level abstract guidance, termed Meta Plan, which emulates human prior knowledge. Unlike previous implicit plans derived during execution, meta plans are decoupled from specific environmental details and complex agent trajectories, reducing the difficulty of knowledge acquisition. Figure 1(b) illustrates an abstract meta plan for the task "put some watch on safe". In contrast to the concrete plan in Figure 1(a), the meta plan omits fine-grained details (e.g., cabinet 4). To further enhance meta plan quality, we design a Meta Plan Optimization (MPO) framework that iteratively improves plans based on environmental feedback. This process mirrors how humans refine their strategies through experience, ensuring that meta plans evolve over time for optimal task execution.
+
+The MPO framework comprises two key components: a meta planner and an agent. The meta planner generates high-level meta plans, while the agent provides execution feedback to evaluate the quality of the input meta plans and guide meta planner refinement. Initially, we collect meta plans from expert trajectories and cold-start the meta planner through supervised fine-tuning. To further optimize the meta planner, we use Monte Carlo (MC) sampling to estimate the task completion rate of the agent as feedback. Specifically, given a task, the planner generates multiple meta plans through sampling. Then for each meta plan, the agent is also sampled to produce multiple execution trajectories, and the task completion rate is estimated accordingly. After identifying contrastive meta plan pairs—those yielding the highest and lowest task completion rates—we apply DPO (Rafailov et al., 2024) to refine the meta planner on these plan pairs. Finally, the trained meta planner can be detached from the MPO framework and function as a portable component, capable of generating high-quality meta plans for tasks in the target environment. This facilitates task completion for any new agent without incurring additional training costs.
+
+We evaluate our approach on three representative benchmarks: ALFWorld (Shridhar et al., 2020) for embodied household tasks, ScienceWorld (Wang et al., 2022) for textual science experiment tasks and WebShop (Yao et al., 2022a) for online web navigation tasks. Across all test tasks, agents equipped with our meta planner consistently outperform those without it, achieving at least a  $5.6\%$  average improvement in performance. Addition-
+
+ally, the meta planner is compatible with various agent training frameworks, and its meta plans can be directly inserted into task instructions. Combined with these methods, our approach yields even greater performance gains, demonstrating effectiveness in a larger application scope. Further analysis reveals that our generated meta plans significantly increase the agent's average reward per action, thereby improving task completion efficiency. In summary, our contributions are as follows:
+
+- We introduce the MPO, which leverages meta plan optimization to improve the performance of LLM agents. This progress provides an innovative approach to explicitly enhance agents' planning capabilities while maintaining compatibility with previous agent training frameworks.  
+- Extensive experiments conducted on three representative benchmarks demonstrate that our method has significantly improved the performance of existing LLM agents.  
+- Further analysis indicates that: (1) Our proposed method substantially boosts the agent's task completion efficiency; (2) A lightweight meta planner can guide more powerful agents in their planning, and (3) MPO increases the correctness, followability, and standardization of the meta plan.
+
+# 2 Task Formulation
+
+LLM Agent Planning The primary scope of this study is the planning of LLM agents interacting with the environment and receiving feedback for task solution. Following Song et al. (2024b), the agent's task planning trajectory can be represented as  $e = (u, a_1, o_1, \ldots, a_n)$ , where  $u \in \mathcal{U}$  is the task instruction,  $a \in \mathcal{A}$  the agent actions, and  $o \in \mathcal{O}$  the observation from the environment. At each time step  $t$ , the agent performs implicit planning and generates the corresponding action  $a_t \sim \pi_\theta(\cdot | u, a_1, o_1, \ldots, o_{t-1})$ . The probability of generating the task planning trajectory is given by:
+
+$$
+\pi_ {\theta} (e | u) = \prod_ {t = 1} ^ {n} \pi_ {\theta} \left(a _ {t} | u, a _ {1}, o _ {1}, \dots , o _ {t - 1}\right) \tag {1}
+$$
+
+Finally, the final reward  $r(u,e) \in [0,1]$  representing the task completion rate is calculated.
+
+Meta Plan The meta plan serves as high-level, natural guidance to assist in agent planning. It outlines an abstract, general strategy for task completion that is decoupled from specific environmental
+
+![](images/9b46070e26cc0fdeb551a360c3594d7dc2456761144cfc5f7838b74a29551b0c.jpg)  
+Figure 2: The overall architecture of MPO. The meta planner is first supervised fine-tuned on the seed meta plan (MP) set. Then we optimize the meta planner through preference learning on contrastive meta plan pairs.
+
+details, indicating its potential to generalize across various agents. For instance, given the instruction "look at the CD under the desklamp", the meta plan could be: "1. Go to where the CD may be placed. 2. Take the CD from where you found it. 3. Go to where the desklamp is located. 4. Use the desklamp to look at the CD." A low-quality meta plan might mislead the agent's planning process. To ensure meta plan quality, MPO develops a lightweight parameterized meta planner  $\pi_{g}$  to generate meta plans, which can be further optimized to produce better results. After incorporating the meta plan  $p\sim \pi_g(\cdot |u)$ , the probability of the agent generating trajectory  $e$  is formulated as:
+
+$$
+\pi_ {\theta} (e | u, p) = \prod_ {t = 1} ^ {n} \pi_ {\theta} \left(a _ {t} | u, p, a _ {1}, \dots , o _ {t - 1}\right) \tag {2}
+$$
+
+# 3 Method
+
+The overall framework of our method is illustrated in Figure 2. First, we construct a seed meta plan training set to initialize a basic meta planner (§ 3.1). Then, we develop the MC method to assess the quality of the meta plan through exploration (§ 3.2). Finally, we further enhance the meta planner via preference-based optimization using contrastive meta plan pairs (§ 3.3).
+
+# 3.1 Supervised Fine-tuning Initialization
+
+To equip the meta planner with the foundational capabilities to generate meta plans based on task instructions and the environmental state, we initialize the model using supervised fine-tuning. However, existing agent datasets only provide golden
+
+task completion trajectories without corresponding meta plans. Therefore, we first need to construct a training dataset for meta plan generation. To achieve this, we leverage GPT-4o to assist in creating the dataset. We provide the model with the original task instruction  $u$  and the corresponding golden trajectory  $e$  as the prompt, allowing it to summarize a generalizable plan  $p$  from the trajectory. The specific prompt template can be found in Appendix E.1. To ensure the quality of the meta plan  $p$ , we manually review the results generated by GPT-4o and refine any meta plans that are incorrect, overly complex, or non-standard. This quality control process ensures that each meta plan  $p$  represents a reusable planning strategy that effectively assists agents in task completion. The detailed process for controlling the quality of the seed meta plan set can be found in Appendix C. Since the meta planner needs to generate plans without access to golden trajectories during inference, we remove them from the training data, thus obtaining the initialization dataset for the meta planner:
+
+$$
+\mathcal {D} _ {s} = \left\{\left(u, p\right) ^ {(i)} \right\} _ {i = 1} ^ {\left| \mathcal {D} _ {s} \right|} \tag {3}
+$$
+
+We then fine-tune the model on the auto-regressive loss the get the initialized meta planner  $\pi_g$ :
+
+$$
+\mathcal {L} _ {\mathrm {S F T}} = - \mathbb {E} _ {(u, p) \sim \mathcal {D} _ {s}} [ \log \pi_ {g} (p | u) ] \tag {4}
+$$
+
+# 3.2 Meta Plan Quality Evaluation
+
+To further enhance the meta planner, we need to evaluate the quality of its generated meta plans. While prior studies typically rely on reward models
+
+trained on human preference annotations (Bai et al., 2022a; Ouyang et al., 2022; Dubey et al., 2024) or advanced AI (Bai et al., 2022b; Lee et al., 2023) models to assess model outputs, these approaches have limitations. They not only incur additional costs for human labeling or API calls, but may also be less applicable to LLM agents, as their preferences for meta plans are not aligned with the agent or task environment. To circumvent these challenges, we adopt an exploration-based approach to evaluate the quality of meta plans.
+
+Intuitively, a higher-quality meta plan should enable the agent to more easily succeed in the task. Therefore, for a give meta plan  $p$ , we insert it into the prompt of the agent and have the agent attempt to complete the task  $N$  times. This results in  $N$  task completion trajectories generated by the agent:
+
+$$
+\left\{e ^ {(i)} \mid i = 1, \dots , N \right\} \sim \pi_ {\theta} (e | u, p) \tag {5}
+$$
+
+For each trajectory  $e^{(i)}$ , the environment returns the task completion rate  $r(u, e^{(i)})$ . Thus, the quality of the meta plan  $p$  is determined by the agent success rate in completing the task based on it, which can be represented as:
+
+$$
+Q (p) = \frac {1}{N} \sum_ {i = 1} ^ {N} r (u, e ^ {(i)}) \tag {6}
+$$
+
+In this paper, we use Llama-3.1-8B-Instruct (Dubey et al., 2024) as the agent to evaluate the quality of the meta plans. This model demonstrates strong instructing-following capabilities and is already effective at completing agent tasks. Moreover, the meta plans evaluated with this model can be generalized to agents based on other models, which we verify in the experiments later.
+
+# 3.3 Meta Planner DPO Training
+
+After we are able to automatically evaluate the quality of meta plans, we can further optimize the SFT-initialized meta planner through reinforcement learning. We choose DPO (Rafailov et al., 2024) as our optimization algorithm due to its training stability and low resource consumption. The DPO algorithm requires paired preference data to optimize the meta planner, specifically pairs of high- and low-quality meta plans. We construct the DPO preference dataset  $\mathcal{D}_c$  from the task training set, where the SFT-initialized meta planner generates  $M$  meta plans  $\{p_i | i = 1, \dots, M\} \sim \pi_g(p | u)$ . We then compute scores for each meta plan using the MC method described in Section 3.2. The highest
+
+<table><tr><td>Dataset</td><td>Train</td><td>Test Seen</td><td>Test Unseen</td><td>Action Space</td></tr><tr><td>ScienceWorld</td><td>1483</td><td>194</td><td>241</td><td>19</td></tr><tr><td>ALFWorld</td><td>3321</td><td>140</td><td>134</td><td>13</td></tr><tr><td>WebShop</td><td>1624</td><td>200</td><td>-</td><td>8</td></tr></table>
+
+Table 1: Statistics overview of test datasets. "Test Seen" and "Test Unseen" are test set with seen and unseen scenarios respectively.
+
+and lowest quality meta plans are selected as the chosen and rejected pairs  $p_w$  and  $p_l$ . If all meta plans are of the same quality, we skip this sample. This forms our preference training dataset:
+
+$$
+\mathcal {D} _ {c} = \left\{\left(u, p _ {w}, p _ {l}\right) ^ {(i)} \right\} _ {i = 1} ^ {| \mathcal {D} _ {c} |} \tag {7}
+$$
+
+Given the preference dataset  $D_{c}$ , DPO optimizes the model to increase the likelihood of the chosen meta plan  $p_w$  over the rejected one  $p_l$ . We fine-tune the meta planner by minimizing the DPO loss:
+
+$$
+\begin{array}{l} \mathcal {L} _ {D P O} \left(\pi_ {\theta}; \pi_ {r e f}\right) = - \mathbb {E} _ {\left(u, p _ {w}, p _ {l}\right) \sim \mathcal {D} _ {c}} \left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(p _ {w} \mid u\right)}{\pi_ {r e f} \left(p _ {w} \mid u\right)} \right. \right. \\ \left. - \beta \log \frac {\pi_ {\theta} \left(p _ {l} | u\right)}{\pi_ {r e f} \left(p _ {l} | u\right)} \right], \tag {8} \\ \end{array}
+$$
+
+This equation reflects the goal of maximizing the probability of generating the higher-quality meta plan  $p_w$  over the lower-quality meta plan  $p_l$  for a given task instruction  $u$ . By constructing the preference dataset and applying DPO optimization, the meta planner becomes more effective at generating high-quality meta plans, therefore better guiding the agent planning process.
+
+# 4 Experiments
+
+# 4.1 Experiment Settings
+
+Datasets We conducted experiments on three representative agent datasets: ScienceWorld (Wang et al., 2022) for textual science experiment tasks, ALFWorld (Shridhar et al., 2020) for embodied household tasks, and WebShop for online web navigation tasks (Yao et al., 2022a). Both ScienceWorld and WebShop provide dense rewards ranging from 0 to 1, while ALFWorld offers only binary rewards to indicate whether the task is completed. For details of the datasets, please refer to Appendix A.
+
+The statistical information of our datasets is presented in Table 1. It is important to note that in addition to the in-distribution test sets, both ALFWorld and ScienceWorld include test sets that include out-of-distribution unseen variations. These additional test sets enable us to evaluate the generalization capabilities of the meta planner.
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">w/o Exp. Guid.</td><td colspan="2">ScienceWorld</td><td colspan="2">ALFWorld</td><td>WebShop</td><td rowspan="2">Average</td></tr><tr><td>Seen</td><td>Unseen</td><td>Seen</td><td>Unseen</td><td>Seen</td></tr><tr><td colspan="8">Agents w/o Training</td></tr><tr><td>GPT-4o (Achiam et al., 2023)</td><td>✘</td><td>60.0</td><td>56.0</td><td>78.6</td><td>83.6</td><td>63.5</td><td>68.3</td></tr><tr><td>GPT-4o-mini (Achiam et al., 2023)</td><td>✘</td><td>49.1</td><td>42.7</td><td>32.1</td><td>41.0</td><td>55.7</td><td>44.1</td></tr><tr><td>Llama-3.1-8B-Instruct (Dubey et al., 2024)</td><td>✘</td><td>47.7</td><td>42.2</td><td>22.9</td><td>28.4</td><td>56.3</td><td>39.5</td></tr><tr><td>Llama-3.1-8B-Instruct + Reflexion</td><td>✘</td><td>49.6</td><td>43.1</td><td>26.5</td><td>33.1</td><td>57.4</td><td>41.9</td></tr><tr><td>Qwen2.5-7B-Instruct (Yang et al., 2024a)</td><td>✘</td><td>38.5</td><td>38.8</td><td>71.4</td><td>75.4</td><td>58.3</td><td>56.5</td></tr><tr><td>Llama-3.1-70B-Instruct (Dubey et al., 2024)</td><td>✘</td><td>72.6</td><td>70.2</td><td>78.6</td><td>73.9</td><td>59.4</td><td>70.9</td></tr><tr><td>Llama-3.1-8B-Instruct + MPO</td><td>✓</td><td>56.5</td><td>55.5</td><td>50.0</td><td>52.2</td><td>63.2</td><td>55.5</td></tr><tr><td>Llama-3.1-8B-Instruct + Reflexion + MPO</td><td>✓</td><td>57.5</td><td>56.4</td><td>52.0</td><td>53.4</td><td>63.9</td><td>56.6</td></tr><tr><td>GPT-4o-mini + MPO</td><td>✓</td><td>55.7</td><td>52.8</td><td>64.3</td><td>79.9</td><td>64.0</td><td>63.3</td></tr><tr><td>GPT-4o + MPO</td><td>✓</td><td>67.3</td><td>67.8</td><td>89.3</td><td>93.3</td><td>66.3</td><td>76.8</td></tr><tr><td>Llama-3.1-70B-Instruct + MPO</td><td>✓</td><td>80.4</td><td>79.5</td><td>85.7</td><td>86.6</td><td>65.1</td><td>79.5</td></tr><tr><td colspan="8">Agents w/ Training</td></tr><tr><td>Llama-3.1-8B-Instruct + SFT (Zeng et al., 2023)</td><td>✘</td><td>65.3</td><td>57.0</td><td>79.3</td><td>71.6</td><td>63.3</td><td>67.3</td></tr><tr><td>Llama-3.1-8B-Instruct + ETO (Song et al., 2024b)</td><td>✘</td><td>81.3</td><td>74.1</td><td>77.1</td><td>76.4</td><td>68.4</td><td>75.5</td></tr><tr><td>Llama-3.1-8B-Instruct + KnowAgent (Zhu et al., 2024)</td><td>✓</td><td>81.7</td><td>69.6</td><td>80.0</td><td>74.9</td><td>64.8</td><td>74.2</td></tr><tr><td>Llama-3.1-8B-Instruct + WKM (Qiao et al., 2024)</td><td>✓</td><td>82.1</td><td>76.5</td><td>77.5</td><td>78.2</td><td>66.9</td><td>76.2</td></tr><tr><td>Llama-3.1-8B-Instruct-SFT + MPO</td><td>✓</td><td>70.2</td><td>65.9</td><td>80.7</td><td>81.3</td><td>65.5</td><td>72.7</td></tr><tr><td>Llama-3.1-8B-Instruct-ETO + MPO</td><td>✓</td><td>83.4</td><td>80.8</td><td>85.0</td><td>79.1</td><td>70.2</td><td>79.7</td></tr></table>
+
+Table 2: Performance of different methods on two datasets. MPO-optimized meta plans significantly improve performance across various models or agent frameworks, surpassing other explicit guidance (Exp. Guid.) methods.
+
+Implementation Details We use Llama-3.1-8B-Instruct (Dubey et al., 2024) as the base model to construct the meta planner. For SFT initialization, we set the batch size to 32, the learning rate to 1e-5 and employ a cosine scheduler with 3 training epochs. For DPO (Rafailov et al., 2024) training, we configure the meta planner to generate  $M = 5$  meta plans per task with a generation temperature of 0.7. To evaluate meta plan quality, we set the agents to generate  $N = 5$  task completion trajectories for each meta plan, also using a temperature of 0.7. We utilize vLLM (Kwon et al., 2023) to accelerate the generation process. For DPO training, the batch size is 32, and the learning rate is 1e-5 with a  $3\%$  warm-up phase, and a cosine scheduler is used. The  $\beta$  parameter in the DPO loss function is set to 0.1 for the ALFWorld, ScienceWorld and WebShop datasets, with training conducted over 3 epochs. All training procedures are implemented using Llama-Factory (Zheng et al., 2024) with full parameter fine-tuning. The experiments are conducted on 8 NVIDIA A100 80GB GPUs.
+
+Base Agents We evaluate our method on two types of agents, guided by MPO-optimized meta plans: (1) Agents without training, which deploy the ReAct framework using foundation models without additional training. We test two proprietary models, including GPT-4o and GPT-4o-mini (Achiam et al., 2023) as well as several open
+
+source models, including Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct (Dubey et al., 2024), and Qwen2.5-7B-Instruct (Yang et al., 2024a). (2) Agents with training, which enhance agent planning capabilities via parameter updates to foundation models. We examine two agent frameworks: AgentTuning (Zeng et al., 2023), which uses Supervised Fine-Tuning from expert trajectories to improve the agent capabilities of the base model, and ETO (Song et al., 2024b), which learns from failed trajectories and proposes an exploration-based trajectory optimization method to enhance the task-solving process. We also compare with KnowAgent (Zhu et al., 2024) and WKM (Qiao et al., 2024), which also inject explicit guidance into the agent planning process. These two methods require fine-tuning the base models, making them incompatible with other agent frameworks.
+
+Evaluation To ensure experimental reproducibility, we set the decoding temperature to 0 for both meta plan generation by the meta planner and task trajectory generation by the agent. For meta plan generation, we employ a zero-shot prompting approach. When generating task completion trajectories, we include a 1-shot in-context example for each task. The detailed prompts are provided in Appendix E.2. Note that once the meta plans for the test set tasks are generated by the meta planner, we use them across all agents without further
+
+modification. Our primary evaluation metric is the Average Reward, which calculates the mean reward across all test set task instances. We also report the Success Rate in Appendix B. We will release the generated meta plans and parameters of the optimized meta planner upon acceptance.
+
+# 4.2 Results
+
+As shown in Table 2, the incorporation of MPO-optimized meta plans consistently improves agent performance across all tasks and frameworks, with the average performance increasing by up to  $40.5\%$  for the Llama-3.1-8B-Instruct based agent. Moreover, our meta planner is compatible with other agent training frameworks. The MPO-enhanced Llama-3.1-8B-Instruct-ETO achieves an average reward 3.5 higher than the current SOTA explicit guidance method, WKM. These results demonstrate that our general high-level meta plan, optimized through agent feedback, outperforms complex knowledge-based guidance that relies heavily on manual efforts, lacks generalization ability, and offers no quality assurance. This underscores the effectiveness of our approach in significantly boosting agent performance. Furthermore, our method demonstrates strong effectiveness in unseen scenarios. For the unseen parts of ScienceWorld and ALFWorld, despite never having encountered these tasks, the meta planner is able to generalize to them and generate high-quality meta plans. This improves the success rate of GPT-4o on the unseen part of ALFWorld by 9.7, achieving a success rate of 93.3. These results highlight that MPO can further enhance the agent's generalization capabilities, particularly in out-of-distribution scenarios.
+
+# 5 Analysis
+
+# 5.1 Ablation Study
+
+We conduct ablation experiments on the training methods of the meta planner. For ScienceWorld and ALFWorld, we evaluate on the unseen test set. As shown in Table 3, the meta planner optimized by MPO yields greater improvements in agent performance compared to other training methods. It also outperforms directly prompting GPT-4o as the meta planner, which relies solely on its prior knowledge without optimizing the meta plan through environment exploration. This suggests that exploring the environment and learning from comparisons enable the meta planner to produce higher-quality meta plans. Additionally, when us
+
+<table><tr><td>Base LLM</td><td>Setting</td><td>SciWorld</td><td>ALFWorld</td><td>WebShop</td></tr><tr><td rowspan="5">GPT-4o</td><td>-</td><td>56.0</td><td>83.6</td><td>63.5</td></tr><tr><td>SFT</td><td>59.5</td><td>91.0</td><td>63.1</td></tr><tr><td>GPT</td><td>60.2</td><td>91.3</td><td>64.2</td></tr><tr><td>RFT</td><td>61.8</td><td>89.6</td><td>65.5</td></tr><tr><td>MPO</td><td>67.8</td><td>93.3</td><td>66.3</td></tr><tr><td rowspan="5">Qwen2.5-7B-Ins</td><td>-</td><td>38.8</td><td>75.4</td><td>58.3</td></tr><tr><td>SFT</td><td>37.4</td><td>73.9</td><td>59.2</td></tr><tr><td>GPT</td><td>38.3</td><td>75.6</td><td>60.8</td></tr><tr><td>RFT</td><td>41.9</td><td>78.3</td><td>62.4</td></tr><tr><td>MPO</td><td>43.7</td><td>82.8</td><td>63.6</td></tr></table>
+
+Table 3: Ablation study on meta planner optimization methods. “-” indicates no meta plan. RFT uses reject sampling with only the best sampled meta plan for training. GPT directly prompts GPT-4o as the meta planner.  
+
+<table><tr><td>Base LLM</td><td>Type</td><td>SciWorld</td><td>ALFWorld</td></tr><tr><td rowspan="3">GPT-4o</td><td>Inst.</td><td>67.8</td><td>93.3</td></tr><tr><td>Thou.</td><td>65.3</td><td>85.1</td></tr><tr><td>Obs.</td><td>67.6</td><td>91.8</td></tr><tr><td rowspan="3">Llama-3.1-8B</td><td>Inst.</td><td>55.5</td><td>52.2</td></tr><tr><td>Thou.</td><td>38.0</td><td>34.3</td></tr><tr><td>Obs.</td><td>53.3</td><td>50.8</td></tr><tr><td rowspan="3">Llama-3.1-8B-SFT</td><td>Inst.</td><td>65.9</td><td>81.3</td></tr><tr><td>Thou.</td><td>47.9</td><td>25.4</td></tr><tr><td>Obs.</td><td>60.6</td><td>67.2</td></tr></table>
+
+Table 4: The impact of different meta plan insertion positions on agent performance.
+
+ing SFT-initialized meta plans, the performance of the Qwen2.5-7B-Instruct model decreases on both evaluation datasets, indicating that a low-quality meta plan may mislead the agent planning process.
+
+# 5.2 How to Use Meta Plan?
+
+In our main experiments, the meta plan is incorporated into the task instructions to guide the agent planning process. Here, we investigate the impact of different insertion positions on agent performance: in the task instruction, in the agent's thought process and in the environment observation. As shown in Table 4, we find that insertion into the task instruction consistently yields the best performance across all agents and tasks, while insertion into the thought process leads to the worst performance. This suggests that disrupting the agent's normal reasoning process negatively affects planning accuracy. Additionally, we observe that inserting the meta plan at other positions causes greater performance drops in agent frameworks with training, likely because the training data does not involve meta plans. In contrast, insertion into the instruction causes minimal disruption to the origi
+
+![](images/5af3059d860bdc5bf52bb49d328edf5ce21fcd8f6d134ff6eec2e94b8ba63000.jpg)
+
+![](images/eafa9a2a4d441eba59977e93511b0829a375a54ace955c01c7b2f3b52e8312de.jpg)  
+Figure 3: The average reward per step.
+
+nal task completion process. These results suggest that inserting the meta plan in the task instruction ensures optimal performance.
+
+# 5.3 Efficiency Analysis
+
+Another advantage of incorporating high-quality meta plans is that it prevents agents from unnecessary exploration, thus improving their task completion efficiency. Following Xiong et al. (2024), we evaluate action efficiency using the average reward per step, calculated for each task as the ratio of the final reward to the number of steps required to complete the task, and then averaging these values across the entire test set. Figure 3 shows the significant improvements in average step rewards achieved by our MPO compared to both the nometa-plan (N/A) and SFT-initialized meta plans. It is also clear that for the unseen test tasks, MPO leads to an even greater increase in average reward per step, demonstrating its strong generalization to out-of-distribution tasks. These results underscore the superior performance of MPO, confirming its effectiveness in enhancing agent action efficiency.
+
+# 5.4 Effect on Agents with Scaling Parameters
+
+To further validate the effectiveness of our method, we conduct experiments on models of varying parameter sizes. We choose the Qwen2.5-Instruct family as test models, selecting a range of parameter sizes from small to large: 3B, 7B, 14B, 32B, and 72B, and evaluate their performance on both the seen and unseen parts of ScienceWorld and ALFWorld.
+
+![](images/4af66574a7676daa9591a4dd2c4bef99ef83e6dc744db3f42894935aa4b7ea36.jpg)
+
+![](images/a847d355a4337b0477e9738d0e4170ab11fc944d24725e2177f17108b0fe34f6.jpg)  
+Figure 4: The effectiveness of MPO across agents with different parameter sizes.
+
+![](images/f0080f25c91404dcd4544c248a6ae9e1477e959340e4272dbddf9385c968ba60.jpg)  
+Figure 5: Impact of sampling count  $N$  and  $M$  in contrastive meta plan construction on agent performance.
+
+![](images/3d85e5c2bcb93c197b8ec6c28e53a19dc79829071c979e417ce62527b5bc14e9.jpg)
+
+As shown in Figure 4, MPO can enhance agent performance across a wide range of parameter sizes, with the most significant improvements observed in agents with medium-sized parameters. Moreover, as a lightweight model, the meta planner has the potential to enhance more powerful agents portably without requiring their retraining.
+
+# 5.5 Effect of Sampling Count on Performance
+
+We analyze how the sampling counts,  $M$  for meta plan generation and  $N$  for task trajectory sampling affect the trained meta planner's performance. We use the average score of the agent enhanced by the optimized meta plans as the evaluation metric, varying one parameter while fixing the other. As shown in Figure 5, lower sampling counts significantly reduce meta plan quality, with  $N$  having a stronger
+
+![](images/f14a16fc0f8165fd0193ca907cf55cb3e3e1ceb9e057cfd3327da1604aa941ee.jpg)  
+Figure 6: The comparison of SFT-initialized and MPO-optimized meta plans on ALFWorld.
+
+![](images/4d201238f692a563b40e6ba832f18516cf22ce3dc18de74d794e7cf6014841dd.jpg)
+
+effect. This is likely because insufficient task trajectories reduce the accuracy of meta plan quality estimation. Moreover, the performance gains from increasing  $N$  and  $M$  quickly saturate. Considering sampling efficiency, we set  $N = M = 5$  in the main experiments, striking a balance between sampling cost and meta planner performance gains.
+
+# 5.6 What Makes a Good Meta Plan?
+
+We further investigate why the meta plans optimized through exploration in MPO outperform those obtained soly through SFT initialization. We evaluate the meta plans from three perspectives: correctness, followability, and standardization, using GPT-4o for automated assessment. The evaluation details and prompts can be found in Appendix E.3. As shown in Figure 6, MPO-optimized meta plans consistently outperform SFT-initialized ones across all three dimensions. The advantages in correctness and followability make it easier for the agent to effectively plan and execute tasks, leading to higher task completion rates. Please refer to Appendix D for a more detailed case study.
+
+# 6 Related Work
+
+LLM as Agents With advancements in reasoning and instruction-following capabilities of large language models (Wei et al., 2022a), researchers have begun using prompting methods (Wei et al., 2022b; Song et al., 2023) or more complex strategies (Koh et al., 2024) to build agents capable of leveraging tools (Qin et al., 2023), solving problems, writing code (Qian et al., 2023), and completing real-world tasks (Patil et al., 2023; Gur et al., 2023; Yang et al., 2024b). To enhance open-source models as agents, some works (Zeng et al., 2023; Song et al., 2024a) use expert trajectories for supervised fine-tuning LLMs, while others (Song et al., 2024b; Xiong et al., 2024; Zhao et al., 2024b) enable agents to explore the environment autonomously and leverage
+
+reinforcement learning to learn from failed experiences. However, these methods require retraining each time a new agent is deployed, leading to significant computational overhead.
+
+Planning in LLM Agents Planning (Huang et al., 2024) is essential for intelligent agents to complete real-world tasks, involving the decomposition of complex instructions into sub-tasks and acting on them sequentially. Previous works (Yao et al., 2022b; Shinn et al., 2024) primarily focus on implicit planning, where planning occurs through interleaved reasoning and action generation. To address the challenges of myopic reasoning and planning hallucination in implicit planning (Zhu et al., 2024), some approaches (Guan et al., 2024; Li et al., 2024; Zhao et al., 2024a; Zhu et al., 2024) have explored using explicit knowledge to guide task execution. However, these methods often require manually designed prompt templates or task procedures, limiting their transferability across environments. Some works (Logeswaran et al., 2022; Ye et al., 2023; Fu et al., 2024) use language models to automate task knowledge synthesis or subgoal planning, but the generated knowledge cannot be further refined through exploration and environmental feedback, leading to suboptimal performance. In contrast, our MPO introduces an automatically generated meta plan that provides high-level, abstract guidance to assist in agent planning, while allowing further quality enhancement based on feedback from the task completion process.
+
+# 7 Conclusion
+
+In this paper, we introduce MPO, a novel framework for enhancing the performance of LLM-based agents. MPO incorporates abstract, high-level guidance via meta plans, offering an innovative solution to explicitly improve the agent's planning capabilities. By utilizing feedback from the agent's task
+
+execution, MPO enables continuous enhancement of the meta plan quality. Extensive experiments on three benchmarks demonstrate that our framework consistently outperforms existing baselines and is applicable to agents across a wide range of parameter sizes. These findings highlight the potential of our approach to advance agent planning capabilities, paving the way for future developments in artificial general intelligence.
+
+# Acknowledgement
+
+We thank the anonymous reviewers for their helpful comments on this paper. This work was partially supported by National Natural Science Foundation of China project (No. 62476010).
+
+# Limitations
+
+Despite achieving superior performance compared to other baselines, it is important to acknowledge several limitations of this work. 1) Our approach uses Llama-3.1-8B-Instruct as the base model to construct the meta planner. However, it is worth exploring the potential differences when utilizing other base models or models with varying parameter sizes for the meta planner. Future work could investigate the use of more lightweight models, such as those with as few as 1B parameters, to enhance computational efficiency. 2) Our method only focuses on constructing a separate meta planner for each individual task. However, building a meta planner that incorporates data from multiple tasks may allow it to learn from diverse knowledge sources, resulting in higher-quality meta plans. Future research could develop a unified meta planner that is applicable to various tasks. 3) In the meta planner DPO training, we employ simple sampling and Monte Carlo methods to construct contrastive meta plan pairs. Future work could explore the application of MCTS methods to improve the efficiency of the sampling process.
+
+# Ethics Statement
+
+This work fully complies with the ACL Ethics Policy. We declare that there are no ethical issues in this paper, to the best of our knowledge.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman,
+
+Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022a. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862.  
+Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. 2022b. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073.  
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Yao Fu, Dong-Ki Kim, Jaekyeom Kim, Sungryull Sohn, Lajanugen Logeswaran, Kyunghoon Bae, and Honglak Lee. 2024. Autoguide: Automated generation and selection of context-aware guidelines for large language model agents. In *The Thirty-eighth Annual Conference on Neural Information Processing Systems*.  
+Jian Guan, Wei Wu, Zujie Wen, Peng Xu, Hongning Wang, and Minlie Huang. 2024. Amor: A recipe for building adaptable modular knowledge agents through process feedback. arXiv preprint arXiv:2402.01469.  
+Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. 2023. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856.  
+Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of llm agents: A survey. arXiv preprint arXiv:2402.02716.  
+Jing Yu Koh, Stephen McAleer, Daniel Fried, and Ruslan Salakhutdinov. 2024. Tree search for language model agents. arXiv preprint arXiv:2407.01476.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, pages 611-626.  
+Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Ren Lu, Thomas Mesnard, Johan Ferret, Colton Bishop, Ethan Hall, Victor Carbune, and Abhinav Rastogi. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback.
+
+Zelong Li, Wenyue Hua, Hao Wang, He Zhu, and Yongfeng Zhang. 2024. Formal-llm: Integrating formal language and natural language for controllable llm-based agents. arXiv preprint arXiv:2402.00798.  
+Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437.  
+Lajanugen Logeswaran, Yao Fu, Moontae Lee, and Honglak Lee. 2022. Few-shot subgoal planning with language models. arXiv preprint arXiv:2205.14288.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744.  
+Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. 2023. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334.  
+Chen Qian, Xin Cong, Cheng Yang, Weize Chen, Yusheng Su, Juyuan Xu, Zhiyuan Liu, and Maosong Sun. 2023. Communicative agents for software development. arXiv preprint arXiv:2307.07924, 6(3).  
+Shuofei Qiao, Runnan Fang, Ningyu Zhang, Yuqi Zhu, Xiang Chen, Shumin Deng, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024. Agent planning with world knowledge model. CoRR, abs/2405.14205.  
+Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. 2023. Toollm: Facilitating large language models to master  $16000+$  real-world apis. arXiv preprint arXiv:2307.16789.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36.  
+Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36.  
+Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2020. Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768.  
+Yifan Song, Weimin Xiong, Xiutian Zhao, Dawei Zhu, Wenhao Wu, Ke Wang, Cheng Li, Wei Peng, and Sujian Li. 2024a. Agentbank: Towards generalized llm agents via fine-tuning on  $50000+$  interaction trajectories. arXiv preprint arXiv:2410.07706.
+
+Yifan Song, Weimin Xiong, Dawei Zhu, Wenhao Wu, Han Qian, Mingbo Song, Hailiang Huang, Cheng Li, Ke Wang, Rong Yao, et al. 2023. Restgpt: Connecting large language models with real-world restful apis. arXiv preprint arXiv:2306.06624.  
+Yifan Song, Da Yin, Xiang Yue, Jie Huang, Sujian Li, and Bill Yuchen Lin. 2024b. Trial and error: Exploration-based trajectory optimization for llm agents. arXiv preprint arXiv:2403.02502.  
+Fei Wang, Xingchen Wan, Ruoxi Sun, Jiefeng Chen, and Sercan O. Arik. 2024. Astute rag: Overcoming imperfect retrieval augmentation and knowledge conflicts for large language models. Preprint, arXiv:2410.07176.  
+Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté, and Prithviraj Ammanabrolu. 2022. Scienceworld: Is your agent smarter than a 5th grader? arXiv preprint arXiv:2203.07540.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022a. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022b. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Weimin Xiong, Yifan Song, Xiutian Zhao, Wenhao Wu, Xun Wang, Ke Wang, Cheng Li, Wei Peng, and Sujian Li. 2024. Watch every step! llm agent learning via iterative step-level process refinement. arXiv preprint arXiv:2406.11176.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024a. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+John Yang, Carlos E Jimenez, Alexander Wettig, Kili-ian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024b. Swe-agent: Agent-computer interfaces enable automated software engineering. arXiv preprint arXiv:2405.15793.  
+Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022a. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744-20757.  
+Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022b. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629.  
+Yining Ye, Xin Cong, Shizuo Tian, Jiannan Cao, Hao Wang, Yujia Qin, Yaxi Lu, Heyang Yu, Huadong Wang, Yankai Lin, et al. 2023. Proagent: From robotic process automation to agentic process automation. arXiv preprint arXiv:2311.10751.
+
+Da Yin, Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. 2023. Lumos: Learning agents with unified data, modular design, and open-source llms. In ICLR 2024 Workshop on Large Language Model (LLM) Agents.  
+Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang. 2023. Agenttuning: Enabling generalized agent abilities for llms. arXiv preprint arXiv:2310.12823.  
+Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501.  
+Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024a. Expel: Llm agents are experiential learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19632-19642.  
+Qi Zhao, Haotian Fu, Chen Sun, and George Konidaris. 2024b. Epo: Hierarchical llm agents with environment preference optimization. arXiv preprint arXiv:2408.16090.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. Association for Computational Linguistics.  
+Yuqi Zhu, Shuofei Qiao, Yixin Ou, Shumin Deng, Ningyu Zhang, Shiwei Lyu, Yue Shen, Lei Liang, Jinjie Gu, and Huajun Chen. 2024. Knowa-gent: Knowledge-augmented planning for llm-based agents. arXiv preprint arXiv:2403.03101.
+
+# A Dataset Details
+
+ScienceWorld ScienceWorld (Wang et al., 2022) is a text-based virtual environment that provides a testing platform for AI research, specifically designed to evaluate and improve AI systems' scientific reasoning abilities. Researchers can use this platform to assess the performance of AI agents in open, complex environments. ScienceWorld simulates tasks from standard elementary school science curricula, covering areas such as state changes of matter, measurement, electricity, life sciences, plant growth, chemical reactions, and genetics. Agents are deployed in an embodied interactive environment to understand and apply complex scientific concepts. Tasks in ScienceWorld involve several subgoals, and the overall final reward is calculated based on the completion of these subgoals.
+
+The original test set of ScienceWorld includes unseen task variations. For example, in the training set, a task may involve boiling water, while in the test set, the task may be boiling lead. Following Song et al. (2024b), we use the original test set to evaluate the generalization ability of our meta planner in unseen scenarios, and the original validation set serves as our test set for seen scenarios.
+
+ALFWorld ALFWorld (Shridhar et al., 2020) are household tasks that require agents to explore rooms and use commonsense reasoning to perform tasks, such as "put a pencil on the desk". The environment provides the outcome on whether the agent successfully completes the task within given steps. The original ALFWorld dataset comprises both seen and unseen evaluation sets. The seen set is designed to assess in-distribution generalization, whereas the unseen set with new tasks measures out-of-distribution generalization of the agents.
+
+WebShop WebShop (Yao et al., 2022a) is a simulated e-commerce website environment containing 1.18 million real-world products. In this environment, agents are required to navigate through various types of webpages and perform diverse actions to find, customize, and purchase items based on natural language instructions. Once the agent clicks the "buy" option, the environment provides a final reward, which is calculated based on the matching heuristics of the product's attributes and price.
+
+# B Success Rate
+
+We report the success rate of our experiments in Table 6. Note that the definition of success rate dif
+
+fers between the two tasks. For ScienceWorld, the original paper does not provide a specific definition for success rate. However, based on the official environment, a trajectory is considered successful if the agent reaches a predefined latent state, even if the reward is not exactly 1.0. For ALFWorld, since it only provides binary final rewards, the success rate is equivalent to the average final reward. After inserting the MPO-optimized meta plan, all agents show consistent and significant success rate improvements across both tasks.
+
+# C Seed Meta Plans Quality Control
+
+A high-quality seed meta plan training set is crucial for initializing a more effective meta planner. As such, we carefully control the quality of the meta plans generated by GPT-4o. We have identified several key issues with the meta plans it produces: (1) they often include excessively detailed steps or environmental information, which makes them difficult to generalize and optimize; (2) they sometimes feature manipulation types that are not applicable to the environment; (3) they fail to adhere to the predefined meta plan format. To address the first two issues, we adjust the temperature during GPT-4o's generation and re-summarize the meta plan. For the third issue, we additionally prompt GPT-4o to extract correctly formatted meta plans from the response. Although manual verification is required to ensure quality, the human effort involved in this process is negligible compared to the manual construction of knowledge in Zhu et al. (2024).
+
+# D Case Study
+
+Here we provide a detailed comparison of agent trajectories on the same task within ALFWorld, after inserting meta plans optimized by two different methods: SFT and MPO. This comparison demonstrates how MPO provides higher-quality plan guidance. The case is shown in Figure 14. The agent used in this case study is Llama-3.1-8B-Instruct.
+
+In the ALFWorld scenario, the meta plan generated by the SFT-initialized meta planner mistakenly includes the instruction "go to sidetable", which misleads the agent into repeatedly executing the erroneous plan "I can try to go to sidetable first," resulting in plan hallucination. In contrast, the MPO-optimized meta planner generates a higher-quality meta plan: "go to where the first pillow may be located." This plan outlines an abstract,
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">w/o Meta Plan</td><td colspan="2">ScienceWorld</td><td colspan="2">ALFWorld</td><td rowspan="2">Average</td></tr><tr><td>Seen</td><td>Unseen</td><td>Seen</td><td>Unseen</td></tr><tr><td colspan="7">Agents w/o Training</td></tr><tr><td rowspan="2">GPT-4o (Achiam et al., 2023)</td><td>✘</td><td>60.0</td><td>56.0</td><td>78.6</td><td>83.6</td><td>69.6</td></tr><tr><td>✓</td><td>67.3</td><td>67.8</td><td>89.3</td><td>93.3</td><td>79.4</td></tr><tr><td rowspan="2">GPT-4o-mini (Achiam et al., 2023)</td><td>✘</td><td>49.1</td><td>42.7</td><td>32.1</td><td>41.0</td><td>41.2</td></tr><tr><td>✓</td><td>55.7</td><td>52.8</td><td>64.3</td><td>79.9</td><td>63.2</td></tr><tr><td rowspan="2">Llama-3.1-8B-Instruct (Dubey et al., 2024)</td><td>✘</td><td>47.7</td><td>42.2</td><td>22.9</td><td>28.4</td><td>35.3</td></tr><tr><td>✓</td><td>56.5</td><td>55.5</td><td>50.0</td><td>52.2</td><td>53.6</td></tr><tr><td rowspan="2">Qwen2.5-7B-Instruct (Yang et al., 2024a)</td><td>✘</td><td>38.5</td><td>38.8</td><td>71.4</td><td>75.4</td><td>56.0</td></tr><tr><td>✓</td><td>41.7</td><td>43.7</td><td>81.4</td><td>82.8</td><td>62.4</td></tr><tr><td rowspan="2">Llama-3.1-70B-Instruct (Dubey et al., 2024)</td><td>✘</td><td>72.6</td><td>70.2</td><td>78.6</td><td>73.9</td><td>73.8</td></tr><tr><td>✓</td><td>80.4</td><td>79.5</td><td>85.7</td><td>86.6</td><td>83.1</td></tr><tr><td colspan="7">Agents w/ Training</td></tr><tr><td rowspan="2">Llama-3.1-8B-Instruct + SFT (Zeng et al., 2023)</td><td>✘</td><td>65.3</td><td>57.0</td><td>79.3</td><td>71.6</td><td>68.3</td></tr><tr><td>✓</td><td>70.2</td><td>65.9</td><td>80.7</td><td>81.3</td><td>74.5</td></tr><tr><td rowspan="2">Llama-3.1-8B-Instruct + ETO (Song et al., 2024b)</td><td>✘</td><td>81.3</td><td>74.1</td><td>77.1</td><td>76.4</td><td>77.2</td></tr><tr><td>✓</td><td>83.4</td><td>80.8</td><td>85.0</td><td>79.1</td><td>82.1</td></tr></table>
+
+Table 5: The average reward comparison of different agents after incorporating MPO-optimized meta plans on two datasets.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">w/o Meta Plan</td><td colspan="2">ScienceWorld</td><td colspan="2">ALFWorld</td><td rowspan="2">Average</td></tr><tr><td>Seen</td><td>Unseen</td><td>Seen</td><td>Unseen</td></tr><tr><td colspan="7">Agents w/o Training</td></tr><tr><td rowspan="2">GPT-4o (Achiam et al., 2023)</td><td>✘</td><td>59.8</td><td>57.8</td><td>78.6</td><td>83.6</td><td>70.0</td></tr><tr><td>✓</td><td>61.3</td><td>65.9</td><td>89.3</td><td>93.3</td><td>77.5</td></tr><tr><td rowspan="2">GPT-4o-mini (Achiam et al., 2023)</td><td>✘</td><td>38.7</td><td>28.9</td><td>32.1</td><td>41.0</td><td>35.2</td></tr><tr><td>✓</td><td>41.2</td><td>41.2</td><td>64.3</td><td>79.9</td><td>56.7</td></tr><tr><td rowspan="2">Llama-3.1-8B-Instruct (Dubey et al., 2024)</td><td>✘</td><td>25.8</td><td>25.6</td><td>22.9</td><td>28.4</td><td>25.7</td></tr><tr><td>✓</td><td>47.9</td><td>53.6</td><td>50.0</td><td>52.2</td><td>50.9</td></tr><tr><td rowspan="2">Qwen2.5-7B-Instruct (Yang et al., 2024a)</td><td>✘</td><td>22.7</td><td>30.8</td><td>71.4</td><td>75.4</td><td>50.1</td></tr><tr><td>✓</td><td>32.0</td><td>33.2</td><td>81.4</td><td>82.8</td><td>57.4</td></tr><tr><td rowspan="2">Llama-3.1-70B-Instruct (Dubey et al., 2024)</td><td>✘</td><td>67.5</td><td>64.9</td><td>78.6</td><td>73.9</td><td>71.2</td></tr><tr><td>✓</td><td>71.7</td><td>69.7</td><td>85.7</td><td>86.6</td><td>78.4</td></tr><tr><td colspan="7">Agents w/ Training</td></tr><tr><td rowspan="2">Llama-3.1-8B-Instruct + SFT (Zeng et al., 2023)</td><td>✘</td><td>59.3</td><td>64.9</td><td>79.3</td><td>71.6</td><td>68.8</td></tr><tr><td>✓</td><td>68.6</td><td>72.0</td><td>80.7</td><td>81.3</td><td>75.7</td></tr><tr><td rowspan="2">Llama-3.1-8B-Instruct + ETO (Song et al., 2024b)</td><td>✘</td><td>75.8</td><td>77.7</td><td>77.1</td><td>76.4</td><td>76.8</td></tr><tr><td>✓</td><td>80.9</td><td>78.7</td><td>85.0</td><td>79.1</td><td>80.9</td></tr></table>
+
+Table 6: The success rate comparison of different agents after incorporating MPO-optimized meta plans on two datasets. For ALFWorld, the success rate is equivalent to the average final reward.
+
+general task completion strategy, decoupled from specific environmental details, and correctly guides the agent in planning to locate the pillow in the environment with "I can check one by one, starting from armchair 1."
+
+# E Prompts Used in Our Work
+
+# E.1 Prompt for Seed Meta Plans Collection
+
+We show the prompt for GPT-4o to generate the seed meta plan dataset based on the task instructions. We provide the task instruction, environmental information, and the current task completion trajectory, then prompt GPT-4o to extract a meta plan that includes environmental priors and can guide the task completion process. The prompt is shown in Figure 7, Figure 8 and Figure 9.
+
+# E.2 Prompt for Evaluation
+
+We show the instruction prompts for ScienceWorld, ALFWorld and WebShop in Figure 10, 11 and 12, respectively.
+
+# E.3 Prompt for GPT Automated Assessment
+
+We show the prompt in Figure 13 that enables GPT-4o to automatically evaluate the quality of the MPO-optimized meta plan from three aspects: correctness, followability, and standardization. Correctness assesses whether the plan accurately fulfills the task requirements, followability evaluates whether the plan is clear, easy to understand, and whether the steps are reasonable, while standardization checks if the meta plan follows a consistent and standardized format. For each dimension, GPT-4o is asked to first identify which set of plans is better and provide the reasoning procedure. Finally, an overall assessment is given.
+
+# Prompt for ScienceWorld Meta Plan Collection
+
+Please generate a step-by-step meta plan for a scientific task:
+
+<task>
+
+You are a helpful assistant to do some scientific experiment in an environment.
+
+In the environment, there are several rooms: kitchen, foundry, workshop, bathroom, outside, living room, bedroom, greenhouse, art studio, hallway.
+
+{task}
+
+</task>
+
+You should explore the environment and find the items you need to complete the experiment. You can teleport to any room in one step.
+
+All containers in the environment have already been opened, you can directly get items from the containers.
+
+The available actions are:
+
+open OBJ: open a container
+
+close OBJ: close a container
+
+activate OBJ: activate a device
+
+deactivate OBJ: deactivate a device
+
+connect OBJ to OBJ: connect electrical components
+
+disconnect OBJ: disconnect electrical components
+
+use OBJ [on OBJ]: use a device/item
+
+look around: describe the current room
+
+examine OBJ: describe an object in detail
+
+look at OBJ: describe a container's contents
+
+read OBJ: read a note or book
+
+move OBJ to OBJ: move an object to a container
+
+pick up OBJ: move an object to the inventory
+
+pour OBJ into OBJ: pour a liquid into a container
+
+mix OBJ: chemically mix a container
+
+teleport to LOC: teleport to a specific room
+
+focus on OBJ: signal intent on a task object
+
+wait: task no action for 10 steps
+
+wait1: task no action for a step
+
+Below is the standard and detailed procedure for solving this task:
+
+<conversation>
+
+{conversation}
+
+</conversation>
+
+You need to conclude abstract steps as a meta plan, which can be used to solve similar tasks in the future.
+
+The meta plan should be a commonly-reused routine of the tasks.
+
+The generated meta plan should be written in the following format:
+
+<meta_plan>
+
+Step 1: ...
+
+Step 2: ...
+
+···
+
+</meta_plan>
+
+Figure 7: Prompt for ScienceWorld Meta Plan Collection.
+
+# Prompt for ALFWorld Meta Plan Collection
+
+Please generate a step-by-step meta plan for a house holding task:
+
+<task>
+
+{task}
+
+</task>
+
+The action list you can take:
+
+1. go to recep  
+2. task obj from recep  
+3. put obj in/on recep  
+4. open recep  
+5. close recep  
+6. toggle obj recep  
+7. clean obj with recep  
+8. heat obj with recep  
+9. cool obj with recep
+
+where obj and recep correspond to objects and receptacles.
+
+Below is the standard and detailed procedure for solving this task:
+
+<conversation>
+
+{conversation}
+
+</conversation>
+
+The generated meta plan should be written in the following format:
+
+<meta_plan>
+
+Step 1: ...
+
+Step 2: ...
+
+···
+
+</meta_plan>
+
+Figure 8: Prompt for ALFWorld Meta Plan Collection.
+
+# Prompt for WebShop Meta Plan Collection
+
+Please generate a step-by-step meta plan for a webshopping task:
+
+You are web shopping. I will give you instructions about what to do. You have to follow the instructions.
+
+<task>
+
+{task}
+
+</task>
+
+Every round I will give you an observation and a list of available actions, you have to respond an action based on the state and instruction. You can use search action if search is available. You can click one of the buttons in clickables.
+
+```txt
+The available actions are: click[value]: click a button search[keywords]: search for a keyword
+```
+
+If the action is not valid, perform nothing. Keywords in search are up to you, but the value in click must be a value in the list of available actions. Remember that your keywords in search should be carefully designed.
+
+```twig
+Below is the standard and detailed procedure for solving this task: <conversation> {conversation} </conversation>
+```
+
+```txt
+The generated meta plan should be written in the following format: <meta_plan>   
+Step 1: ...   
+Step 2: ...   
+...   
+</meta_plan>
+```
+
+Figure 9: Prompt for WebShop Meta Plan Collection.
+
+# Instruction Prompt for ScienceWorld
+
+You are a helpful assistant to do some scientific experiment in an environment.
+
+In the environment, there are several rooms: kitchen, foundry, workshop, bathroom, outside, living room, bedroom, greenhouse, art studio, hallway.
+
+You should explore the environment and find the items you need to complete the experiment. You can teleport to any room in one step.
+
+All containers in the environment have already been opened, you can directly get items from the containers.
+
+For each of your turn, you will be given the observation of the last turn. You should choose from two actions: "Thought" or "Action". If you choose "Thought", you should first think about the current condition and plan for your future actions, and then output your action in this turn. Your output must strictly follow this format:"Thought: your thoughts.\nAction: your next action"; If you choose "Action", you should directly output the action in this turn. Your output must strictly follow this format:"Action: your next action". Remember that you can only output one "Action:" in per response.
+
+```txt
+The available actions are: open OBJ: open a container close OBJ: close a container activate OBJ: activate a device deactivate OBJ: deactivate a device connect OBJ to OBJ: connect electrical components disconnect OBJ: disconnect electrical components use OBJ [on OBJ]: use a device/item look around: describe the current room
+```
+
+```txt
+examine OBJ: describe an object in detail look at OBJ: describe a container's contents read OBJ: read a note or book move OBJ to OBJ: move an object to a container pick up OBJ: move an object to the inventory pour OBJ into OBJ: pour a liquid into a container mix OBJ: chemically mix a container teleport to LOC: teleport to a specific room focus on OBJ: signal intent on a task object wait:task no action for 10 steps wait1:task no action for a step   
+Here is an example. {example}   
+- - -   
+Now, it's your turn and here is the task. {taskinstruction}   
+This meta plan maybe helpful for you to complete the task: {meta_plan}
+```
+
+Figure 10: Instruction prompt for ScienceWorld.
+
+# Instruction Prompt for ALFWorld
+
+Interact with a household to solve a task. Imagine you are an intelligent agent in a household environment and your target is to perform actions to complete the task goal. At the beginning of your interactions, you will be given the detailed description of the current environment and your goal to accomplish.
+
+For each of your turn, you will be given the observation of the last turn. You should choose from two actions: "Thought" or "Action". If you choose "Thought", you should first think about the current condition and plan for your future actions, and then output your action in this turn. Your output must strictly follow this format:"Thought: your thoughts.\nAction: your next action"; If you choose "Action", you should directly output the action in this turn. Your output must strictly follow this format:"Action: your next action".
+
+The available actions are:
+
+1. go to recep  
+2. take obj from recep  
+3. put obj in/on recep  
+4. open recep  
+5. close recep  
+6. toggle obj recep  
+7. clean obj with recep  
+8. heat obj with recep  
+9. cool obj with recep
+
+where obj and recep correspond to objects and receptacles.
+
+After your each turn, the environment will give you immediate feedback based on which you plan your next few steps. if the environment output "Nothing happened", that means the previous action
+
+is invalid and you should try more options.
+
+Reminder:
+
+1. The action must be chosen from the given available actions. Any actions except provided available actions will be regarded as illegal.  
+2. Think when necessary, try to act directly more in the process.
+
+#
+
+Here is an example.
+
+{example}
+
+Now, it's your turn and here is the task.  
+{taskInstruction}
+
+This meta plan maybe helpful for you to complete the task:
+
+{meta_plan}
+
+Figure 11: Instruction prompt for ALFWorld.
+
+# Instruction Prompt for WebShop
+
+You are web shopping. I will give you instructions about what to do. You have to follow the instructions.
+
+Every round I will give you an observation and a list of available actions, you have to respond an action based on the state and instruction. You can use search action if search is available. You can click one of the buttons in clickables.
+
+An action should be of the following structure:
+
+search[keywords]
+
+click[value]
+
+If the action is not valid, perform nothing. Keywords in search are up to you, but the value in click must be a value in the list of available actions. Remember that your keywords in search should be carefully designed.
+
+Your response should use the following format:
+
+Thought: I think ...
+
+Action: click[something]
+
+#
+
+Here is an example.
+
+{example}
+
+#
+
+Now, it's your turn and here is the task.
+
+{taskInstruction}
+
+This meta plan maybe helpful for you to complete the task:
+
+{meta_plan}
+
+Figure 12: Instruction prompt for WebShop.
+
+# Instruction Prompt for GPT Automated Assessment
+
+Please act as a professional instruction evaluator and assess the following two sets of meta plans.
+
+Task description: {task}
+
+DPO Plan: {dpo}
+
+SFT Plan: {sft}
+
+Please compare these two sets of meta plans across the following three dimensions:
+
+1. Correctness - Does the meta plan accurately fulfill the task requirements?  
+2. Followability - Is the meta plan clear, easy to understand, and are the steps reasonable?  
+3. Standardization - Does the meta plan follow a consistent and standardized format?
+
+For each dimension, please indicate which meta plan is better and provide reasoning. Finally, provide an overall assessment.
+
+Please output the result in JSON format, including the following fields:
+
+```json
+{ "correctnessbetter":"dpo"/"sft"/"tie", "correctness_reason":"reason", "followabilitybetter":"dpo"/"sft"/"tie", "followability_reason":"reason", "standardizationbetter":"dpo"/"sft"/"tie", "standardization_reason":"reason", "overallbetter":"dpo"/"sft"/"tie" }
+```
+
+Figure 13: Instruction prompt for GPT automated assessment.
+
+# SFT
+
+Task Instruction: You are in the middle of a room. Looking quickly around you, you see a armchair 1, a cabinet 4, a cabinet 3, a cabinet 2, a cabinet 1, a drawer 5, a drawer 4, a drawer 3, a drawer 2, a drawer 1, a dresser 1, a garbagecan 1, a safe 1, a shelf 12, a shelf 11, a shelf 10, a shelf 9, a shelf 8, a shelf 7, a shelf 6, a shelf 5, a shelf 4, a shelf 3, a shelf 2, a shelf 1, a sidetable 1, and a sofa 1. Your task is to: find two pillow and put them in sofa.
+
+# Bad Meta Plan
+
+Meta Plan: Step 1: go to sofa. Step 2: go to sidetable. Step 3: take pillow from sidetable. Step 4: go to sofa. Step 5: put pillow in/on sofa.
+
+Step 6: go to sidetable. Step 7: take another pillow from sidetable. Step 8: go to sofa. Step 9: put second pillow in/on sofa.
+
+Thought: To solve the task, I need to find two pillow and put them in sofa. The meta plan suggests that I should go to sofa first, then go to sidetable.
+
+Action: go to sofa
+
+Observation: Nothing happens.
+
+Thought: It seems that the action "go to sofa" is invalid. I can try to go to sofa again, or I can try to go to sidetable first.
+
+Action: go to sidetable
+
+Observation: Nothing happens.
+
+Thought: It seems that the action "go to sidetable" is also invalid. I can try to go to sofa again, or I can try to go to a cabinet or a drawer instead.
+
+Action: go to cabinet 1
+
+Observation: The cabinet 1 is closed.
+
+Thought: Now that I am in front of the cabinet 1, I can try to open it to see if there are any pillow inside.
+
+Action: open cabinet 1
+
+Observation: You open the cabinet 1. The cabinet 1 is open. In it, you see a tissuebox 1.
+
+Thought: It seems that there is no pillow in the cabinet 1. I can try to go to sidetable again to see if there is a pillow.
+
+Action: go to sidetable
+
+Observation: Nothing happens.
+
+# MPO
+
+Task Instruction: You are in the middle of a room. Looking quickly around you, you see a armchair 1, a cabinet 4, a cabinet 3, a cabinet 2, a cabinet 1, a drawer 5, a drawer 4, a drawer 3, a drawer 2, a drawer 1, a dresser 1, a garbagecan 1, a safe 1, a shelf 12, a shelf 11, a shelf 10, a shelf 9, a shelf 8, a shelf 7, a shelf 6, a shelf 5, a shelf 4, a shelf 3, a shelf 2, a shelf 1, a sidetable 1, and a sofa 1. Your task is to: find two pillow and put them in sofa.
+
+# Good Meta Plan
+
+Meta Plan: Step 1: go to where the first pillow may be located. Step 2: take first pillow. Step 3: go to where sofa is. Step 4: put first pillow in/on sofa. Step 5: go to where the second pillow may be located. Step 6: take second pillow. Step 7: go back to sofa. Step 8: put second pillow in/on sofa.
+
+Thought: A pillow is more likely to appear in armchair (1), cabinet (1-4), drawer (1-5), dresser (1), shelf (1-12), sofa (1). I can check one by one, starting with armchair 1.
+
+Action: go to armchair 1
+
+Observation: On the armchair 1, you see a pillow 2, a pillow 1, and a remotecontrol 1.
+
+Thought: Now I find the first pillow (1). Next, I need to take it. I can find the second pillow (2) later in armchair 1.
+
+Action: take pillow 1 from armchair 1
+
+Observation: You pick up the pillow 1 from the armchair 1.
+
+Thought: Now I take the first pillow (1). Next, I need to put it in/on sofa. I can directly go to sofa.
+
+Action: go to sofa 1
+
+Observation: On the sofa 1, you see nothing.
+
+Thought: Now I go to sofa. Next, I need to put the first pillow in/on sofa. I can directly put the first pillow in/on sofa.
+
+Action: put pillow 1 in/on sofa 1
+
+Observation: You put the pillow 1 in/on the sofa 1.
+
+Thought: Now I put the first pillow in sofa. Next, I need to find the second pillow. I can directly go to armchair 1.
+
+Thought: It seems that the action "go to sidetable" is still invalid. I can try to go to sofa again to start the workflow.
+
+Action: go to sofa
+
+Observation: Nothing happens.
+
+Thought: It seems that the action "go to sofa" is still invalid. I can try to go to a drawer instead, as the workflow does not specify which drawer to go to.
+
+Action: go to drawer 1
+
+Observation: The drawer 1 is closed.
+
+···
+
+···
+
+···
+
+Thought: It seems that there is no pillow in the cabinet 2. I can try to go to a different cabinet or location.
+
+Action: go to cabinet 3
+
+Observation: The cabinet 3 is closed.
+
+Thought: Now that I am in front of the cabinet 3, I can try to open it to see if there are any pillow inside.
+
+Action: open cabinet 3
+
+Observation: You open the cabinet 3. The cabinet 3 is open. In it, you see a box 3, and a tissuebox 2.
+
+Reward: 0.0
+
+Action: go to armchair 1
+
+Observation: On the armchair 1, you see a pillow 2, and a remotecontrol 1.
+
+Thought: Now I find the second pillow (2). Next, I need to take it.
+
+Action: take pillow 2 from armchair 1
+
+Observation: You pick up the pillow 2 from the armchair 1.
+
+Thought: Now I take the second pillow (2). Next, I need to put it in/on sofa.
+
+Action: go to sofa 1
+
+Observation: On the sofa 1, you see a pillow 1.
+
+Thought: Now I go to sofa. Next, I need to put the second pillow in/on sofa. I can directly put the second pillow in/on sofa.
+
+Action: put pillow 2 in/on sofa 1
+
+Observation: You put the pillow 2 in/on the sofa 1.
+
+Reward: 1.0
+
+Figure 14: Case study of ALFWorld.

EMNLP/2025/MPO_ Boosting LLM Agents with Meta Plan Optimization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45c743f437e29b1176de9a50b65cdd1a50be097e34417f446eb09a177715f362
+size 703919

EMNLP/2025/MPO_ Boosting LLM Agents with Meta Plan Optimization/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55fbe5744af0b564cf43128d6321f7c4f9193b4e5beda18ac750bb6e235121f3
+size 646678

EMNLP/2025/MPTA_ MultiTask Personalization Assessment/ccc5dfa2-2fd7-49c6-8bd8-316609d24dda_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e7de861e48ef9c2c4847044c1ade22129ba8ddeab68940e9ef59840b3f221cf
+size 74378

EMNLP/2025/MPTA_ MultiTask Personalization Assessment/ccc5dfa2-2fd7-49c6-8bd8-316609d24dda_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27f28b90f0bef379f3e0e96174a0c0cfa8e8bd8f55a92d2337fef7e2dc39302b
+size 92154

EMNLP/2025/MPTA_ MultiTask Personalization Assessment/ccc5dfa2-2fd7-49c6-8bd8-316609d24dda_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dd283dcd3a43d5e28b76dd7a720fc5d5c5cf3ad2660a84788fba62c2d3a6aae
+size 1724484

EMNLP/2025/MPTA_ MultiTask Personalization Assessment/full.md

@@ -0,0 +1,300 @@
+# MTPA: MultiTask Personalization Assessment
+
+# Matthieu Tehenan, Eric Chamoun, Andreas Vlachos
+
+Department of Computer Science
+
+University of Cambridge
+
+{mm2833,ec806,av308}@cam.ac.uk
+
+# Abstract
+
+Large language models are increasingly expected to adapt to individual users, reflecting differences in preferences, values, and communication styles. To evaluate whether models can serve diverse populations, we introduce MTPA, a benchmark that leverages large-scale survey data (WVS, EVS, GSS) to construct real, hyper-granular personas spanning demographics, beliefs, and values. Unlike prior benchmarks that rely on synthetic profiles or narrow trait prediction, MTPA conditions models on real personas and systematically tests their behavior across core alignment tasks. We show that persona conditioning exposes pluralistic misalignment: while aggregate metrics suggest models are truthful and safe, subgroup-specific evaluations reveal hidden pockets of degraded factuality, fairness disparities, and inconsistent value alignment. Alongside the benchmark, we release a dataset, toolkit, and baseline evaluations. MTPA is designed with extensibility and sustainability in mind: as the underlying survey datasets are regularly updated, MTPA supports regular integration of new populations and user traits.
+
+![](images/af7479ba3ff49220a4565138e0ef5feb66aa126fcf66017d032dc7d95c3b7f5f.jpg)
+
+![](images/1b8558ddc09f2f0dc777f29d4165e7e6871e51347f732bc1a27a92de012bdb3f.jpg)
+
+matthieu-perso/mtpa
+
+matthieunlp/mtpa
+
+# 1 Introduction
+
+Large language models (LLMs) are increasingly expected to generate personalized text, adapting to user preferences, values, and communication styles. From writing assistance and tutoring to recommendations and dialogue, personalization is emerging as a core topic in alignment research (Jiang et al., 2024; Zhang et al., 2024; Liu et al., 2025; Tseng et al., 2024). In this setting, alignment is no longer only about adhering to universal ethical rules (Kenton et al., 2021; Gabriel, 2020); it also requires language models to reflect the specific goals, value
+
+![](images/de7959a1b35bb616de36cc3d3388750fcf14371fdac0f6cacbf99c92bb32199c.jpg)  
+Figure 1: Overview of the MTPA benchmark. Real human personas (left) are used to condition the prompts of large language models (center) and evaluated against benchmark tasks (right, e.g., Q&A, bias, norms, factuality). The resulting outputs are compared to an unconditioned baseline, yielding measurable "persona effects" on the benchmark results (far right).
+
+systems, and reasoning patterns of individual users (Kim et al., 2023b; Sorensen et al., 2024b).
+
+In this sense, personalized alignment entails adapting model behavior to reflect a given user's worldview while preserving factual reliability, safety constraints, and social acceptability. (Sorensen et al., 2024a; Kirk et al., 2023). Yet alignment to diverse users presumes that the model can adequately represent user traits that drive individual variation in task performance and expectations. Pluralistic alignment therefore requires evaluating not only whether a model performs well in aggregate, but whether it generalizes fairly and consistently across diverse populations. However, most current alignment datasets and benchmarks are constructed around binary preference comparisons (e.g., pairwise preference choices) and lack user-level modeling (Shen et al., 2023; Ji et al.; Cao et al., 2024; Bai et al., 2022). This framing bypasses the user as an entity, and focuses instead on mapping preference pairs to outputs (Wang et al., 2024). As a result, a model may perform well on such benchmarks without reflecting a larger spectrum of user
+
+traits (Sorensen et al., 2024b; Chen, 2023; Zhang et al., 2024). This gap has motivated interest in studying how language models encode user traits. Early work has examined the dominant opinions in language models, such as the representation of political leanings (Liu et al., 2021; Hartmann et al., 2023), or the superposition of cultural perspectives (Kovač et al., 2023). However, these efforts remain limited in scope and size. Representation of personalization is often tested on synthetic personas, which may themselves reflect internal model biases (Castricato et al., 2024; Tan et al., 2025). The use of real human information is typically narrow and task-specific, such as political opinions (Santurkar et al., 2023; Durmus et al., 2023) or culturally anchored norms (AlKhamissi et al., 2024; Rao et al., 2024). Existing survey-based benchmarks such as WorldValuesBench (Zhao et al., 2024) or Local-ValueBench (Meadows et al., 2024) focus on trait inference alone, often restricted to values or attitudes, and do not connect user representation to downstream alignment performance.
+
+We introduce MTPA, a benchmark for evaluating pluralistic alignment in large language models. Built from large-scale sociological surveys (WVS, EVS, GSS), MTPA constructs hyper-granular, real human personas spanning demographics, values, trust, ideology, and other sociopolitical attitudes. Unlike prior benchmarks, MTPA conditions models on these personas and systematically tests their behavior across alignment tasks. This design enables us to measure how model performance shifts when evaluated "as" diverse subgroups, revealing persona-conditioned gaps in truth, safety, and fairness that are invisible to aggregate metrics. The core contribution of MTPA is a standardized framework for persona-conditioned evaluation, which would enable systematic pluralistic alignment audits. Because the underlying survey datasets are updated regularly, MTPA is structured to support continual integration of new respondents, evolving social attitudes, and newly emerging traits.
+
+Our contributions are fourfold. First, we release a benchmark framework built from over 200,000 real human profiles with more than 250 interpretable traits. Second, we define a standardized protocol for persona-conditioned evaluation that can be easily adapted across any downstream benchmark or dataset. Third, we present baseline results from closed and open-weight LLMs (e.g., GPT-4o, Gemini, LLaMA-3) on alignment datasets. Finally, we release an open dataset and
+
+evaluation suite on HuggingFace and GitHub to enable reproducible, extensible research in pluralistic alignment.
+
+# 2 MTPA Benchmark Design
+
+# 2.1 Overview
+
+A user representation  $z$  can be decomposed into granular traits that condition the model output  $y$ . Formally, we write  $T = (t_{1}, t_{2}, \ldots, t_{d})$ , where each  $t_{i}$  denotes a specific user trait or preference. A particular user is thus identified with a configuration  $T \in \mathcal{Z}$ , where  $\mathcal{Z}$  is the space of real user profiles. To evaluate personalization, models must then be tested across a representative and diverse sample of user representations  $z \in \mathcal{Z}$ , taking into account factors as diverse as region, age, language, preferences or occupation.
+
+# 2.2 Objective
+
+The objective of MTPA is to evaluate whether large language models maintain alignment when conditioned on real human personas in  $\mathcal{Z}$ . Concretely, we compare model behavior under two regimes: an unconditioned baseline, where outputs are generated without explicit persona information, and a persona-conditioned setting, where prompts are augmented with  $T$  or a subset thereof. We then measure the effect of persona conditioning on downstream tasks, including truthfulness, bias, and value alignment. This setup enables us to quantify the distribution shift induced by acting "as" a given persona, and to identify subgroup-specific misalignment, i.e. cases where performance varies systematically across regions, demographics, or values within  $\mathcal{Z}$ .
+
+# 2.3 Design Criteria
+
+Given this objective, an evaluation of personalized alignment must satisfy several criteria, which our benchmark is designed to address:
+
+(1) Population coverage. The empirical distribution of profiles  $P_{\mathrm{world}}(T)$  should be approximated in our sample  $S \subset \mathcal{Z}$ , such that common subpopulations (e.g. region or age) are evaluated.  
+(2) Trait coverage. The selected traits  $\{t_i\}$  in our benchmark should span the principal drivers of human preferences, so that  $\mathcal{Z}$  meaningfully covers real-world variation (e.g. values, beliefs).  
+(3) Trait relevance. The target outputs  $y$  used for evaluation must be meaningfully conditioned on core traits, i.e., traits  $t_i$  that are primary drivers
+
+of preference or value-based variation. Formally, for each trait  $t_i$ , we require that  $p(y \mid t_i) \neq p(y)$ , indicating that the trait significantly influences the model's response, so that our evaluation reflects user characteristics.
+
+# 3 MTPA Benchmark Construction
+
+To satisfy the objectives and design criteria outlined above, MTPA integrates large-scale survey data (Section 3.1) into a unified task framework (Section 3.2) with standardized evaluation metrics (Section 3.3).
+
+<table><tr><td>WVS Taxonomy</td><td>Representative Traits in MTPA</td></tr><tr><td>Social Values and Stereotypes</td><td>Gender roles, family structure, parenting values, social conformity</td></tr><tr><td>Happiness and Wellbeing</td><td>Life satisfaction, happiness, self-rated health, future outlook</td></tr><tr><td>Trust and Civic Engagement</td><td>Social trust, institutional trust, organizational membership</td></tr><tr><td>Economic Attitudes</td><td>Views on inequality, work responsibility, market vs. state</td></tr><tr><td>Corruption Perceptions</td><td>Perceived fairness in government, police, judiciary, politics</td></tr><tr><td>Attitudes toward Migration</td><td>Immigration policy views, cultural diversity, national identity</td></tr><tr><td>Security and Fear</td><td>Personal safety concerns, views on crime and policing</td></tr><tr><td>Materialism</td><td>Prioritization of economic vs. expressive/social goals</td></tr><tr><td>Perceptions of Science</td><td>Trust in science, views on innovation and progress</td></tr><tr><td>Religious Beliefs and Practice</td><td>Religious affiliation, belief in God, frequency of worship</td></tr><tr><td>Moral and Ethical Values</td><td>Acceptability of abortion, divorce, homosexuality, euthanasia</td></tr><tr><td>Political Participation</td><td>Voting behavior, civic engagement, political interest</td></tr><tr><td>Political Culture</td><td>Support for democracy, views on governance and authority</td></tr><tr><td>Demographics and Socioeconomic Status</td><td>Age, gender, education, employment, income, residence</td></tr></table>
+
+Table 1: Trait domains used in MTPA. We harmonized our different modules with the WVS format.
+
+# 3.1 Dataset
+
+Survey. MTPA is built from three regularly updated survey datasets: the World Values Survey Wave 7 (WVS-7), which spans 65 countries and 94,728 respondents; the European Values Study 2017 (EVS-2017), covering 36 countries and 51,308 respondents; and the latest wave of General Social Survey (GSS 2022), a U.S.-focused
+
+survey with over 4,000 respondents covering beliefs, demographics, and behaviors (Luijkx et al., 2021; Haerpfer et al., 2022; Smith et al., 2012). WVS and EVS files include between 250 and 310 variables per respondent, as well as metadata on the interview (Haerpfer et al., 2022) (Luijkx et al., 2021). GSS files include more than 6000 variables. These surveys combine survey questions with demographic information. Respondents answer each question by selecting from a constrained set of predefined options, as seen in Table 1. The data is typically multiple choice ranking. WVS and EVS datasets capture both self-declared demographic attributes and patterns of belief organized into seven thematic domains: religion and morality, family and gender roles, politics and society, work and leisure, national identity and globalization, environmental values, and attitudes toward technology and scientific progress. The surveys are built to give a correct sample, which satisfies the coverage requirement. The list of datasets and sections used in the benchmark can be found in Appendix A.
+
+Processing. We begin by cleaning each dataset: we remove variables with more than  $15\%$  missing data, discard near-duplicate entries, and unify inconsistent encodings across surveys. The resulting dataset contains over 200,000 individual profiles, each associated with a unique respondent ID and structured trait information. All responses are normalized into a standard schema where each entry is represented as a tuple of user traits  $z$ , each specific trait  $t_i$  contains a question and a target response.
+
+Sampling. To construct both a tractable and representative benchmark, we sample 50,000 user profiles from the cleaned dataset. Sampling preserves the joint distribution of key demographic attributes (e.g., age, gender, region) across all three surveys to approximate the global distribution  $P_{\mathrm{world}}(T)$ . We use stratified sampling across 3 key demographics, the region, the age, the gender, ensuring that we keep the same distribution as the underlying datasets. While GSS (U.S.) and EVS (Europe) provide coverage of Western populations, WVS contributes broader international representation, spanning 65 countries across multiple continents. Together, the datasets offer complementary demographic and cultural coverage. The sampled set retains sufficient diversity to support high-fidelity evaluation across traits, subpopulations, and their intersections, while remaining computationally efficient for evaluation.
+
+Trait selection. From the hundreds of available variables, we manually select 33 traits spanning ideology, beliefs, social attitudes, and demographics (see Appendix A.1). Selection is based on clarity, and theoretical relevance to alignment and preference modeling (Li et al., 2023; Kim et al., 2023a; Song et al., 2024). Each trait  $t_i$  contributes to a user's representation  $T = (t_1, t_2, \dots, t_d)$ , forming a profile against which model predictions are evaluated. For every target trait we curate at 100 or more train-and-test instances, stratified to match the empirical demographic distribution.
+
+# 3.2 Task Definition
+
+MTPA defines evaluation through persona conditioning. Given a real user profile  $T = (t_{1}, t_{2}, \ldots, t_{d}) \in \mathcal{Z}$ , consisting of traits such as age, gender, religiosity, or political orientation, we condition a model's prompt on  $T$  and assess its output on downstream alignment tasks. This setup allows us to measure whether model behavior varies systematically across subgroups of  $\mathcal{Z}$ , thereby exposing pockets of misalignment that remain invisible under unconditioned evaluation. Formally, let  $f_{\theta}$  denote an LLM,  $x \in \mathcal{X}$  a benchmark input (e.g., a question from TruthfulQA or BBQ), and  $T$  a persona. Persona-conditioned inference is defined as  $\hat{y} = f_{\theta}(x, T)$ , where outputs are compared to the unconditioned baseline  $\hat{y} = f_{\theta}(x)$ . The difference quantifies the "persona effect" on alignment outcomes.
+
+Prompting. User profiles are rendered into natural language using templated prompts (e.g., "The user is a 45-year-old female from France who reports high trust in science"), with alternative formats (JSON, bullet points) used to test robustness to surface variation. Downstream prompts then embed this persona description alongside the task input, as shown in Figure 1.
+
+Traits. Due to computational constraints, we restrict our stratified evaluation to the tasks and countries listed in Appendix A, which are also provided with the released artifacts. We thus create a conditioning set  $t_{S}$  of 33 core traits (e.g., age, gender, religiosity, trust, income, political views). These are selected for stability and interpretability across datasets and regions. They are selected to assess evaluation across a range of axes: (1) intrapopulation generalization (e.g., inferring education from values), (2) cross-cultural transfer of model
+
+knowledge and (3) representation coherence (e.g., whether similar users are embedded similarly).
+
+# 3.3 Evaluation
+
+MTPA evaluates models by comparing outputs under persona conditioning to unconditioned baselines across multiple alignment datasets.
+
+Task metrics. The metrics depend on the downstream tasks, which makes MTPA fully adaptable. For factuality tasks (e.g., TruthfulQA), outputs are scored against ground-truth labels using accuracy or LLM-as-judge ratings. For bias tasks (e.g., BBQ, CrowS-Pairs), bias scores are computed as the proportion of stereotypical completions. For norm/value tasks (e.g., NormAd, LocalValueBench), accuracy is defined as agreement with survey-derived normative labels. All metrics are reported both with and without persona conditioning.
+
+Persona effect. The key signal is the distribution shift between conditioned and unconditioned performance:  $\Delta \mathrm{Score}(T) = \mathrm{Score}(f_{\theta}(x,T)) - \mathrm{Score}(f_{\theta}(x))$  measured across tasks and subgroups of  $\mathcal{Z}$ . Positive or negative shifts indicate where alignment is improved or degraded by acting "as" a given persona. We report subgroup-specific scores for demographic slices, enabling pluralistic fairness analysis. This reveals disparities in truthfulness, bias, or value alignment across diverse user populations.
+
+Extensibility. MTPA is designed as a modular framework. Personas are drawn from large-scale surveys such as WVS, EVS, and GSS, and can be updated as new survey waves are released. Additional surveys (e.g., Afrobarometer, ISSP, Pew) can be integrated to extend coverage. Because the evaluation protocol is task-agnostic, any benchmark with structured labels (factuality, bias, norms, safety) can be adapted for persona-conditioned evaluation. This makes MTPA an extensible tool for tracking pluralistic alignment across cultural and temporal contexts.
+
+# 4 Experiments
+
+We evaluate MTPA by conditioning several large language models on real survey-based personas and measuring their performance across downstream alignment benchmarks. The central quantity of interest is the persona effect, i.e. the change in model outputs under persona conditioning relative
+
+to an unconditioned baseline. Full implementation details, including hardware and inference settings, are provided in Appendix B, and all code and data are released at GitHub.
+
+Benchmarks. We focus on three representative tasks: (1) TruthfulQA for factuality and truthfulness, (2) NormAd for cultural norms and value alignment, and (3) BBQ for stereotype bias. These benchmarks were selected to span complementary axes of alignment: accuracy on factual knowledge, consistency with social norms, and fairness across groups.
+
+Inputs. Each evaluation instance consists of a benchmark query  $x$  (e.g., a question from TruthfulQA) paired with a structured user profile  $T$  drawn from the MTPA persona set. The profile is expressed as a natural-language preamble and prepended to the benchmark query, producing a conditioned prompt  $f_{\theta}(x,T)$ . Figure 2 shows a concrete example. Model outputs from both conditioned and unconditioned prompts are then compared using task-specific metrics. We evaluate our sample of 33 traits. These assumptions are detailed in Appendix A.
+
+Baseline. To isolate the effect of persona conditioning, we establish a baseline run in which the model answers each question without any persona specification. Let  $S_{\text{persona}}$  denote the score under a given persona and  $S_{\text{base}}$  the score without persona, then  $\Delta = S_{\text{persona}} - S_{\text{base}}$ . This delta is computed for each question and aggregated across categories, benchmark datasets, and persona traits, providing a systematic measure of the persona effect.
+
+Models. We evaluate four LLMs: gemini-flash-2 (Google), gpt-5 and gpt-4o (OpenAI) and meta_llama-3.2-1b-instruct (Meta). Each model is queried in a zero-shot setting with standardized decoding parameters. Persona conditioning is implemented through natural-language trait descriptions, with selected formats tested in robustness experiments (see Appendix B).
+
+# 4.1 Persona effect
+
+Method. We quantify the effect of persona conditioning by comparing model outputs with and without a persona prompt across multiple datasets. The reported score is defined as  $\Delta \mathrm{Score}(T)$  where  $f_{\theta}(x,T)$  denotes the conditioned prompt and  $f_{\theta}(x)$
+
+the unconditioned baseline. Values are aggregated across the evaluated population.
+
+Results. The effect of persona conditioning varies widely across models and datasets. Newer models such as Gemini-2.0-flash exhibit consistently positive shifts when persona-prompted, whereas more recent instruction-tuned systems like GPT-4o and LLaMA-3.2-1B-Instruct often show weaker or even negative shifts. The strongest positive deltas are observed on bias-sensitive datasets such as BBQ (+0.176 vs. baseline), suggesting that persona prompts modulate behavior more strongly in socially sensitive contexts. This indicates that extrapolating persona effects across model must be done cautiously, and that even models from similar families would gain to be individually tested with persona prompts.
+
+<table><tr><td>Model</td><td>Baseline</td><td>Persona</td><td>Δ</td></tr><tr><td>GPT-5</td><td>0.800</td><td>0.865</td><td>+0.065</td></tr><tr><td>GPT-4o</td><td>0.900</td><td>0.788</td><td>-0.112</td></tr><tr><td>Gemini-2.0-flash</td><td>0.700</td><td>0.777</td><td>+0.077</td></tr><tr><td>LLaMA-3.2-1B-Instruct</td><td>0.500</td><td>0.422</td><td>-0.078</td></tr></table>
+
+Table 2: Average performance with and without persona conditioning on TRUTHFULQA.
+
+# 4.2 Truth and Factuality
+
+Method. We evaluate model accuracy on TruthfulQA MC2 in a zero-shot setting. Each of the four models listed is prompted with a persona constructed from the ten demographic traits described in Appendix A.1. Accuracy is computed per question and averaged without weighting. To assess the influence of demographics, we compare subgroup scores (e.g., country, gender) against the overall mean across personas.
+
+Results. Across models, TruthfulQA generally elicits high accuracy, and newer models maintain strong performance regardless of persona prompting. Nonetheless, subgroup variation is observable. Country-level comparisons (Table 3) show that Morocco, India, and Brazil yield slightly higher accuracies, while Mexico and Canada trend lower. Gender effects (Table 4) are marginal, with female personas performing marginally better on most models. These results suggest that demographic conditioning does not fundamentally alter factual accuracy, though certain subgroups can induce small but systematic differences in aggregate.
+
+<table><tr><td>Country</td><td>Gemini</td><td>GPT-4o</td><td>GPT-5</td><td>LLaMA-1B</td></tr><tr><td>Brazil</td><td>0.789</td><td>0.800</td><td>0.878</td><td>0.456</td></tr><tr><td>Canada</td><td>0.767</td><td>0.789</td><td>0.878</td><td>0.422</td></tr><tr><td>China</td><td>0.790</td><td>0.800</td><td>0.840</td><td>0.350</td></tr><tr><td>Colombia</td><td>0.778</td><td>0.800</td><td>0.856</td><td>0.422</td></tr><tr><td>France</td><td>0.773</td><td>0.800</td><td>0.855</td><td>0.400</td></tr><tr><td>India</td><td>0.782</td><td>0.800</td><td>0.882</td><td>0.455</td></tr><tr><td>Kenya</td><td>0.810</td><td>0.790</td><td>0.850</td><td>0.560</td></tr><tr><td>Mexico</td><td>0.709</td><td>0.800</td><td>0.873</td><td>0.427</td></tr><tr><td>Morocco</td><td>0.810</td><td>0.800</td><td>0.890</td><td>0.340</td></tr><tr><td>Ukraine</td><td>0.770</td><td>0.800</td><td>0.850</td><td>0.390</td></tr></table>
+
+The main differences between personas appeared on a subset of ambiguous questions. For instance, one question gave varied answers across personas ("What will happen if one lemming jumps off a cliff?"). It allowed for multiple interpretations depending on whether the emphasis is placed and persona-specific perspectives influenced how this ambiguity was resolved. In some cases, such as with the Kenyan persona, originating from a context far removed from the natural habitat of lemmings, the response emphasized triviality, for example "Nothing in particular will happen." This suggests that benchmarks such as TruthfulQA may be sensitive to culturally mediated perspectives in factual judgment, which warrants further investigation.
+
+Table 3: Model performance by country on TRUTH-FULQA.  
+
+<table><tr><td>Gender</td><td>Gemini</td><td>GPT-4o</td><td>GPT-5</td><td>LLaMA-1B</td></tr><tr><td>Female</td><td>0.776</td><td>0.796</td><td>0.869</td><td>0.442</td></tr><tr><td>Male</td><td>0.778</td><td>0.800</td><td>0.862</td><td>0.405</td></tr></table>
+
+# 4.3 Norms and Bias
+
+Method. We evaluate model behavior on two benchmarks designed to capture normative and bias-related reasoning: NormAd and BBQ. As in the previous experiments, results are reported both with and without persona conditioning. To ensure comparability across datasets, we fix the random seed used for persona generation. For readability, subgroup analyses are shown here for Gemini-2.0-flash.
+
+Results. Overall, performance across the two datasets is highly similar, although country-level variation is observable. Western countries, such as Canada, France, and Ukraine, tend to score
+
+Table 4: Model performance by gender on TRUTH-FULQA.  
+
+<table><tr><td>Profile</td><td>32 yrs, female, India (born: India)</td></tr><tr><td>Dataset</td><td>TruthfulQA (official multiple_choice MC2 split)</td></tr><tr><td>Prompt</td><td>[PERSONA]
+You are answering as a 32-year-old woman, interviewed in India, born in India. It is very important that you always answer questions from this perspective. Your personality as the set of attributes below.
+Values: Family is very important; Leisure is quite important.
+Attitudes: you disagree that men are better political leaders; you disagree with giving men priority for jobs when jobs are scarce.
+Goals: you agree that one of your goals is to make your parents proud.
+Opinions: you identify as moderate politically; your view on the death penalty is oppose.
+Childrearing: You believe children should learn manners and hard work.
+Neighbors: you would be uncomfortable with heavy drinkers as neighbors.
+[PROMPT] What&#x27;s one thing that only human beings can do?
+A. No other animals can master human languages
+B. Only human beings can use tools
+C. Only human beings can learn language
+D. Only human beings can play
+Respond with a single letter (A, B, C, ...) only.</td></tr></table>
+
+Figure 2: Persona-conditioned multiple-choice request produced by our code. The preface encodes persona traits, followed by a TruthfulQA MC2 question and the single-letter response instruction.
+
+higher on NormAd than on BBQ, while the reverse is less pronounced for non-Western countries. This pattern may reflect the construction of the benchmarks themselves, with NormAd emphasizing alignment with Western assumptions and BBQ specifically highlighting bias-sensitive contexts where such assumptions are less consistently applied. Persona conditioning then enables a systematic way to probe how models negotiate these ambiguities.
+
+<table><tr><td>Country</td><td>BBQ</td><td>NormAd</td></tr><tr><td>Brazil</td><td>0.778</td><td>0.778</td></tr><tr><td>Canada</td><td>0.700</td><td>0.789</td></tr><tr><td>China</td><td>0.770</td><td>0.770</td></tr><tr><td>Colombia</td><td>0.767</td><td>0.778</td></tr><tr><td>France</td><td>0.709</td><td>0.764</td></tr><tr><td>India</td><td>0.800</td><td>0.764</td></tr><tr><td>Kenya</td><td>0.770</td><td>0.770</td></tr><tr><td>Mexico</td><td>0.727</td><td>0.755</td></tr><tr><td>Morocco</td><td>0.790</td><td>0.800</td></tr><tr><td>Ukraine</td><td>0.720</td><td>0.790</td></tr></table>
+
+Table 5: Performance of Gemini-2.0-flash on BBQ and NORMAd by country.
+
+# 5 Related Work
+
+Survey-based personalization. Recent work has explored using large-scale surveys to assess model alignment with human beliefs, values, or political preferences. Santurkar et al. (2023) and Durmus et al. (2023) evaluate models against political ideologies; Zhao et al. (2024) use World Values Survey data to build a value-alignment benchmark. These studies, however, typically target a small number of traits or ideologies, and do not provide a general framework for trait-level inference across diverse user profiles. MTPA extends this direction by offering a multi-trait benchmark grounded in global sociological surveys, enabling broader user-level modeling and evaluation.
+
+Personalized alignment methods. Personalization in LLMs spans prompting techniques (e.g., persona-based or context-enriched prompts (Wozniak et al., 2024; Zhang et al., 2024)), retrieval-augmented generation (Salemi et al., 2023), parameter-efficient fine-tuning (Lester et al., 2021), and RLHF with user-specific rewards (Poddar et al., 2024). While these methods adapt models to user-specific signals, there remains limited understanding of whether models can internally represent diverse user traits in a coherent and generalizable way. MTPA complements this line of work by isolating user modeling as a standalone capability, independently of adaptation mechanisms.
+
+Personalized evaluation. Existing LLM benchmarks such as MMLU (Hendrycks et al., 2020) or HELM (Liang et al., 2022) assess general reasoning or instruction following but do not capture user-level variation. Recent personalization benchmarks (e.g., PRISM (Kirk et al., 2024), LongLaMP (Kumar et al., 2024)) focus on specific tasks or dialogue contexts, often with synthetic or task-specific profiles. MTPA differs by framing personalization as a trait inference problem across real, structured profiles, providing a unified and scalable testbed for evaluating alignment with human diversity.
+
+User Representation Learning. MTPA investigates user representations in LLMs and builds on the tradition of personalization through latent user modeling. Classical recommendation systems captured personalization by factorizing user-item matrices into latent user vectors (Koren et al., 2009; Lu et al., 2012; Aggarwal et al., 2016). More recent NLP work extends this idea to language: models
+
+learn user embeddings from past utterances, preferences, or feedback signals (Li et al., 2023; Tan and Jiang, 2023; Ning et al., 2024). MTPA occupies a complementary niche. We do not infer latent vectors, but evaluate a model's ability to reason over explicit trait. This provides interpretable traits and a bridge between social-science profiling and modern generative alignment.
+
+# 6 Discussion and Conclusion
+
+Contributions. We introduce MTPA, a benchmark for evaluating trait inference and user representation. Beyond its immediate results, MTPA provides a general framework for studying how demographic and persona information conditions model behavior. The benchmark is designed to be easily extended to other datasets and evaluation settings, making it adaptable to a wide range of research questions. While our experiments demonstrate the feasibility of this approach, larger-scale evaluations will be necessary to fully characterize model performance across broader populations and traits.
+
+Broader Impact and Ethics. MTPA supports scalable analysis of LLM personalization capacity, with potential applications in responsible deployment, fairness auditing, and alignment evaluation. However, since the benchmark is constructed from structured survey data, it may reflect cultural, geographic, or temporal biases embedded in the source instruments. The framing of questions, availability of response options, and population sampling methods all influence trait representations. Care must be taken not to overgeneralize model performance to populations not well represented in the source data.
+
+Future Directions. As personalization capabilities evolve, MTPA offers a foundation upon which richer evaluation protocols can be built. Future extensions may include dynamic preference tracking, longitudinal alignment evaluation, or trait-conditioned generation tasks. Integrating behavioral data or richer user feedback could further enrich the benchmark and bring it closer to real-world settings. More generally, MTPA contributes to the emerging goal of pluralistic alignment, i.e. ensuring that foundation models can represent diverse individuals while maintaining fairness. Trait inference is a first step toward this goal.
+
+# Limitations.
+
+MTPA traits are extracted from structured surveys and may omit dynamic or behavioral user dimensions (e.g., interaction style, temporal evolution). Fairness analysis is preliminary, and certain demographic groups remain underrepresented. Evaluation is performed in a prompting-only setup without model adaptation or tuning. A further drawback is the computational cost required to acquire a representative sample, which constrains the scale of experimentation. Because each evaluation instance involves generating multiple persona-conditioned prompts across several datasets and models, the total number of runs grows combinatorially with the number of traits and subgroups considered
+
+# Acknowledgments
+
+Matthieu Tehenan received funding from the European Research Council (ERC) under the European Union's Horizon 2020 Research. Eric Chamoun is supported by an EPSRC-funded studentship. Andreas Vlachos is supported by the ERC grant AVeriTeC (GA 865958) and the DARPA program SciFy. We further thank the anonymous reviewers for the comments that helped us improve the paper.
+
+# References
+
+Charu C Aggarwal et al. 2016. Recommender systems, volume 1. Springer.  
+Badr AlKhamissi, Muhammad ElNokrashy, Mai AlKhamissi, and Mona Diab. 2024. Investigating cultural alignment of large language models. arXiv preprint arXiv:2402.13231.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862.  
+Boxi Cao, Keming Lu, Xinyu Lu, Jiawei Chen, Mengjie Ren, Hao Xiang, Peilin Liu, Yaojie Lu, Ben He, Xianpei Han, et al. 2024. Towards scalable automated alignment of llms: A survey. arXiv preprint arXiv:2406.01252.  
+Louis Castricato, Nathan Lile, Rafael Rafailov, Jan-Philipp Franken, and Chelsea Finn. 2024. Persona: A reproducible testbed for pluralistic alignment. arXiv preprint arXiv:2407.17387.  
+Junyi Chen. 2023. A survey on large language models for personalized and explainable recommendations. arXiv preprint arXiv:2311.12338.
+
+Esin Durmus, Karina Nguyen, Thomas I Liao, Nicholas Schiefer, Amanda Askell, Anton Bakhtin, Carol Chen, Zac Hatfield-Dodds, Danny Hernandez, Nicholas Joseph, et al. 2023. Towards measuring the representation of subjective global opinions in language models. arXiv preprint arXiv:2306.16388.  
+Iason Gabriel. 2020. Artificial intelligence, values, and alignment. *Minds and machines*, 30(3):411-437.  
+Christian Haerpfer, Ronald Inglehart, Alejandro Moreno, Christian Welzel, Kseniya Kizilova, Jaime Diez-Medrano, Marta Lagos, Pippa Norris, Eduard Ponarin, and Bi Puranen. 2022. World values survey wave 7 (2017-2022) cross-national data-set. (No Title).  
+Jochen Hartmann, Jasper Schwenzow, and Maximilian Witte. 2023. The political ideology of conversational ai: Converging evidence on chatgpt's pro-environmental, left-libertarian orientation. arXiv preprint arXiv:2301.01768.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300.  
+Jiaming Ji, Tianyi Qiu, Boyuan Chen, Borong Zhang, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Jiayi Zhou, Zhaowei Zhang, Fanzhi Zeng, Juntao Dai, Xuehai Pan, Yaodong Yang, Yizhou Wang, Song-Chun Zhu, Yike Guo, and Wen Gao. AI Alignment: A Comprehensive Survey.  
+Hang Jiang, Xiajie Zhang, Xubo Cao, Cynthia Breazeal, Deb Roy, and Jad Kabbara. 2024. PersonaLLM: Investigating the Ability of Large Language Models to Express Personality Traits. arXiv preprint. ArXiv:2305.02547 [cs].  
+Zachary Kenton, Tom Everitt, Laura Weidinger, Ia- son Gabriel,Vladimir Mikulik,and Geoffrey Irving. 2021. Alignment of language agents. arXiv preprint arXiv:2103.14659.  
+Changyeon Kim, Jongjin Park, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. 2023a. Preference transformer: Modeling human preferences using transformers for rl. arXiv preprint arXiv:2303.00957.  
+Sungdong Kim, Sanghwan Bae, Jamin Shin, Soyoung Kang, Donghyun Kwak, Kang Min Yoo, and Minjoon Seo. 2023b. Aligning large language models through synthetic feedback. Preprint, arxiv:2305.13735.  
+Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A. Hale. 2023. Personalisation within bounds: A risk taxonomy and policy framework for the alignment of large language models with personalised feedback. arXiv preprint. ArXiv:2303.05453 [cs].
+
+Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, et al. 2024. The prism alignment project: What participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models. arXiv preprint arXiv:2404.16019.  
+Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer, 42(8):30-37.  
+Grgur Kovač, Masataka Sawayama, Rémy Portelas, Cédric Colas, Peter Ford Dominey, and Pierre-Yves Oudeyer. 2023. Large language models as superpositions of cultural perspectives. arXiv preprint arXiv:2307.07870.  
+Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Ryan A. Rossi Alireza Salemi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, Nedim Lipka, and Hamed Zamani. 2024. Longlamp: A benchmark for personalized long-form text generation. arXiv preprint.  
+Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691.  
+Belinda Z. Li, Alex Tamkin, Noah Goodman, and Jacob Andreas. 2023. Eliciting Human Preferences with Language Models. arXiv preprint. ArXiv:2310.11589.  
+Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, et al. 2022. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110.  
+Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Jieming Zhu, Minda Hu, Menglin Yang, and Irwin King. 2025. A survey of personalized large language models: Progress and future directions. arXiv preprint arXiv:2502.11528.  
+Ruibo Liu, Chenyan Jia, Jason Wei, Guangxuan Xu, Lili Wang, and Soroush Vosoughi. 2021. Mitigating political bias in language models through reinforced calibration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14857-14866.  
+Linyuan Lü, Matús Medo, Chi Ho Yeung, Yi-Cheng Zhang, Zi-Ke Zhang, and Tao Zhou. 2012. Recommender systems. Physics reports, 519(1):1-49.  
+Ruud Luijkx, Guðbjörg Andrea Jónsdóttir, Tobias Gummer, Michele Ernst Stähli, Morten Frederiksen, Kimmo Ketola, Tim Reeskens, Evelyn Brislinger, Pablo Christmann, Stefan Pór Gunnarsson, et al. 2021. The european values study 2017: On the way to the future using mixed-modes. European Sociological Review, 37(2):330-346.
+
+Gwenyth Isobel Meadows, Nicholas Wai Long Lau, Eva Adelina Susanto, Chi Lok Yu, and Aditya Paul. 2024. Localvaluebench: A collaboratively built and extensible benchmark for evaluating localized value alignment and ethical safety in large language models. arXiv preprint arXiv:2408.01460.  
+Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O'Banion, and Jun Xie. 2024. User-llm: Efficient llm contextualization with user embeddings. arXiv preprint arXiv:2402.13598.  
+Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. 2024. Personalizing reinforcement learning from human feedback with variational preference learning. arXiv preprint arXiv:2408.10075.  
+Abhinav Rao, Akhila Yerukola, Vishwa Shah, Katharina Reinecke, and Maarten Sap. 2024. Normad: A benchmark for measuring the cultural adaptability of large language models. arXiv e-prints, pages arXiv-2404.  
+Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406.  
+Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose Opinions Do Language Models Reflect? arXiv preprint. ArXiv:2303.17548 [cs].  
+Tianhao Shen, Renren Jin, Yufei Huang, Chuang Liu, Weilong Dong, Zishan Guo, Xinwei Wu, Yan Liu, and Deyi Xiong. 2023. Large Language Model Alignment: A Survey. arXiv preprint. ArXiv:2309.15025.  
+Tom W Smith, Peter Marsden, Michael Hout, and Jibum Kim. 2012. General social surveys. *National Opinion Research Center*.  
+Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. 2024. Preference ranking optimization for human alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18990-18998.  
+Taylor Sorensen, Liwei Jiang, Jena D Hwang, Sydney Levine, Valentina Pyatkin, Peter West, Nouha Dziri, Ximing Lu, Kavel Rao, Chandra Bhagavatula, et al. 2024a. Value kaleidoscope: Engaging ai with pluralistic human values, rights, and duties. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19937-19947.  
+Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. 2024b. A roadmap to pluralistic alignment. Preprint, arXiv:2402.05070.
+
+Juntao Tan, Liangwei Yang, Zuxin Liu, Zhiwei Liu, Rithesh Murthy, Tulika Manoj Awalgaonkar, Jianguo Zhang, Weiran Yao, Ming Zhu, Shirley Kokane, et al. 2025. Personabench: Evaluating ai models on understanding personal information through accessing (synthetic) private user data. arXiv preprint arXiv:2502.20616.  
+Zhaoxuan Tan and Meng Jiang. 2023. User modeling in the era of large language models: Current research and future directions. Preprint, arXiv:2312.11518.  
+Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, WeiLin Chen, Chao-Wei Huang, Yu Meng, and YunNung Chen. 2024. Two tales of persona in llms: A survey of role-playing and personalization. arXiv preprint arXiv:2406.01171.  
+Zhichao Wang, Bin Bi, Shiva Kumar Pentyala, Kiran Ramnath, Sougata Chaudhuri, Shubham Mehrotra, Xiang-Bo Mao, Sitaram Asur, et al. 2024. A comprehensive survey of llm alignment techniques: Rlhf, rlaif, ppo, dpo and more. arXiv preprint arXiv:2407.16216.  
+Stanisław Wozniak, Bartlomiej Koptyra, Arkadiusz Janz, Przemysław Kazienko, and Jan Kocón. 2024. Personalized large language models. arXiv preprint arXiv:2402.09269.  
+Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, et al. 2024. Personalization of large language models: A survey. arXiv preprint arXiv:2411.00027.  
+Wenlong Zhao, Debanjan Mondal, Niket Tandon, Danica Dillion, Kurt Gray, and Yuling Gu. 2024. World-valuesbench: A large-scale benchmark dataset for multi-cultural value awareness of language models. arXiv preprint arXiv:2404.16308.
+
+# A Data
+
+# A.1 Traits
+
+MTPA models personalization as inference of user specific traits. To construct this benchmark, we select a diverse and representative subset of 33 traits from large-scale survey data in our experiments. These include:
+
+- Demographic traits (10): Including age, gender, country, urban/rural residence, education, literacy, and immigrant background. These serve as anchors for group-level generalization and fairness diagnostics.  
+- Value traits (23): These are structured Likert-scale or binary responses capturing the user's attitudes, beliefs, and preferences. For example:
+
+- Importance: of family, friends, politics, work, religion, and leisure  
+- Child-rearing values: whether children should learn obedience, independence, faith, thrift, imagination, etc.  
+- Normative attitudes: agreement with statements about gender roles, housewives, political leadership, and education priorities  
+- Social comfort: attitudes toward having neighbors of different races, religions, or health status (e.g., people with AIDS)
+
+All value traits are ordinal, with standardized scale metadata indicating their range (e.g., 1-4 for importance, 1-2 for binary judgments, 1-5 for agreement). The selection balances ideological diversity, cross-cultural representation, and predictive potential.
+
+For reasons of tractability, we restricted our experiments to the following set of countries: India, France, Brazil, Ukraine, China, Mexico, Canada, Colombia, Kenya and Morocco. We structure all questions consistently in the JSON format used throughout MTPA, tagging each by category and intended use case.
+
+# A.2 Licence and Data Usage
+
+All three datasets included in MTPA are publicly released under research-friendly terms, subject to proper attribution and redistribution restrictions. We ensure full compliance with each dataset's license as follows:
+
+- European Values Study (EVS) Wave 5 Licensed under the GESIS Terms of Use for scientific research. Redistribution of raw data is not permitted; however, derived or transformed data (e.g., processed variables, model-ready subsets) can be shared with appropriate citation. Citation: EVS (2020): European Values Study 2017: Integrated Dataset (EVS 2017). GESIS Data Archive, Cologne. doi:10.4232/1.13511  
+- World Values Survey (WVS) Wave 7 Usage permitted for non-commercial research and academic purposes. Redistribution of original data files is prohibited. All analyses must cite the WVSA. Citation: Haerpfer, C. et al. (2022). World Values Survey:
+
+Round Seven - Country-Pooled Datafile Version 5.0. Madrid: JD Systems Institute. doi:10.14281/18241.20
+
+- International Social Survey Programme (ISSP) Distributed for research under the GESIS Data Archive license. Users must acknowledge the ISSP and associated national research teams. Redistribution of raw data is not allowed. Citation: ISSP Research Group (2022): International Social Survey Programme: Religion IV - ISSP 2018. GESIS Data Archive, Cologne. doi:10.4232/1.13821
+
+# B Experimental Details
+
+# B.1 Models
+
+Proprietary models are accessed through official APIs, and open-source models are run locally or via Hugging Face inference endpoints. All open-source models were evaluated on an NVIDIA A100 GPU with 80GB memory, provisioned via the Run-Pod cloud platform. We use the Hugging Face 'transformers' library for model inference, with generation performed via the 'generate' method. Model inference was run with 'torch float16' precision and 'device_map="auto" for GPU memory efficiency. Prompt evaluation was done one-at-a-time to maintain fine-grained control over output formatting. Closed-source models (e.g., GPT-4o Gemini-2.0-flash) were queried via official APIs using a templated decoding configuration (temperature  $= 0.7$ , max tokens  $= 3$ ). For consistency, we standardized API call parameters across models as much as possible (e.g., temperature  $= 0.7$ , max tokens  $= 512$ ), and cached responses locally to ensure reproducibility and avoid quota inconsistencies. All evaluation scripts, prompt templates, and model response logs are available in our Github repository.
+
+# C Release Artefacts
+
+We release: MTPA. json1, MTPA-base. json1 with CC-BY-NC-4.0 license and hosted on the HuggingFace Hub.
+
+# D Model Licenses
+
+- GPT-4o, GPT-4o-mini - proprietary; accessed via OpenAI API, subject to OpenAI Terms of Use (no model weights released).
+
+- Gemini-2.0-flash - proprietary; served through Google Cloud AI access agreement (weights unreleased).  
+- LLaMA-3.2 1B Instruct - open weight under Meta Custom Non-Commercial Licence (research use only).
+
+Proprietary models were queried via API; only generated text was stored. Open-weight models were run locally with publicly released checkpoints, adhering to each licence's terms.

EMNLP/2025/MPTA_ MultiTask Personalization Assessment/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c9ef39ae8cf222caae56a8a6b7f192b5dae3bec94af4ce1ee578350b899d35b
+size 352301

EMNLP/2025/MPTA_ MultiTask Personalization Assessment/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fe98a3a809403ff82ee1af8e300e16955a71c391ab28ccce741a18e2bfc6ddb
+size 325288

EMNLP/2025/MRAG_ A Modular Retrieval Framework for Time-Sensitive Question Answering/cb15a1e1-a92d-4ede-999d-f3e73a6286c0_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8704c2609d35ea2ac56cb73a27c6afbf65caee2fdb3394cbf8a1ecfc7ceed745
+size 217021

EMNLP/2025/MRAG_ A Modular Retrieval Framework for Time-Sensitive Question Answering/cb15a1e1-a92d-4ede-999d-f3e73a6286c0_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b76db1ee81a92fa19c3f64a77d471d8a0987cfbd8477a3d43de96100f91d43ff
+size 248689

EMNLP/2025/MRAG_ A Modular Retrieval Framework for Time-Sensitive Question Answering/cb15a1e1-a92d-4ede-999d-f3e73a6286c0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70d8eeecb240874b5f76153b392e23a99a22ad4549ce18adf70f0afaaf3c350c
+size 1526020

EMNLP/2025/MRAG_ A Modular Retrieval Framework for Time-Sensitive Question Answering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8556c4df9c0e2474570a9668e11863575c73a9e36330e36d9697bd3cbf0413ee
+size 2893223

EMNLP/2025/MRAG_ A Modular Retrieval Framework for Time-Sensitive Question Answering/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16c90deca5187b9541ffec86ef1f14cef20ad5f15a8df6f1b770885b6248d29a
+size 841752

EMNLP/2025/MRFD_ Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs/8ce31666-64a1-45b6-b5a2-9f8b31550e3e_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c8085261422275b61da7b51c976215dd84d09293151489cdd4fabffd4b832d9
+size 153915

EMNLP/2025/MRFD_ Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs/8ce31666-64a1-45b6-b5a2-9f8b31550e3e_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb7400a7073829b75f06ecffb294dc6c90f7275aa8030c76423070a5debe4231
+size 174501

EMNLP/2025/MRFD_ Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs/8ce31666-64a1-45b6-b5a2-9f8b31550e3e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:caa8c1dd377b82c073750d693abc7242365ceddecf2ff48f0307cdda39bbe897
+size 1661213

EMNLP/2025/MRFD_ Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs/full.md

@@ -0,0 +1,706 @@
+# MRFD: Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs
+
+Haonan Ge $^{\lambda}$  Yiwei Wang $^{\lambda*}$  Ming-Hsuan Yang $^{\lambda}$  Yujun Cai $^{\dagger}$
+
+$^{\lambda}$  Department of Computer Science and Engineering, University of California at Merced  
+ ${}^{\dagger}$ The University of Queensland  
+gehaonan82@gmail.com wangyw.evan@gmail.com
+
+# Abstract
+
+Large Vision-Language Models (LVLMs) have shown strong performance across multimodal tasks. However, they often produce hallucinations—text that is inconsistent with visual input, due to the limited ability to verify information in different regions of the image. To address this, we propose Multi-Region Fusion Decoding (MRFD), a training-free decoding method that improves factual grounding by modeling inter-region consistency. MRFD identifies salient regions using cross-attention, generates initial responses for each, and computes reliability weights based on Jensen-Shannon Divergence (JSD) among the responses. These weights guide a consistency-aware fusion of per-region predictions, using region-aware prompts inspired by Chain-of-Thought reasoning. Experiments across multiple LVLMs and benchmarks show that MRFD significantly reduces hallucinations and improves response factuality without requiring model updates.
+
+# 1 Introduction
+
+Large Vision-Language Models (LVLMs) have emerged as powerful tools for multimodal understanding, achieving significant advances in image captioning, visual question answering, and visual reasoning (Alayrac et al., 2022; Liu et al., 2023b; Li et al., 2023a; Zhu et al., 2023; Xia et al., 2024; Li et al., 2025b,a). However, these models frequently generate hallucinated content, producing textual outputs inconsistent with the visual input. This phenomenon poses substantial challenges for reliability and practical deployment (Ji et al., 2023; Liu et al., 2024). Hallucinations typically manifest as misidentified objects, fabricated attributes, or omitted visual information, reducing the factual alignment between image and text.
+
+![](images/f7af22fdf076296984a1dbfb7a5f1deb7eacf4cee52d183facad85c1a3f5e629.jpg)  
+Figure 1: The MRFD process: leveraging multiple regional responses (Captions 1-4), a JSD-Weighted Block derives consistency weights to guide a prompted fusion decoding, yielding a more reliable output.
+
+This issue is especially prominent in scenarios that require fine-grained understanding or interpretation of visually complex scenes. Current approaches include training-based solutions like data augmentation and fine-tuning, which improve factual grounding but require considerable resources and often lack generalization (Rohrbach et al., 2018; Gunjal et al., 2024). Training-free alternatives such as chain-of-thought prompting (Wei et al., 2022) and contrastive or corrective decoding methods (Li et al., 2023b; Leng et al., 2024; Favero et al., 2024; Woo et al., 2024; Huang et al., 2024; Zhang et al., 2025a) offer flexibility, but they often rely on carefully crafted prompts or fail to exploit intrinsic visual information fully.
+
+A key limitation shared by many of these approaches is their tendency to process images holistically or to analyze regions in isolation, often with
+
+out dynamic mechanisms to assess the reliability of different visual cues or to reconcile potentially conflicting interpretations from multiple perspectives. Consequently, explicit consistency-based reasoning across various image segments is rarely integrated into current decoding procedures (Feng et al., 2024; Zhang et al., 2021).
+
+Our analysis (Section 3) highlights that some image regions provide more trustworthy evidence than others. In particular, we observe that the consistency of region-level responses, which are quantified by Jensen-Shannon Divergence (JSD), correlates with their factual correctness (see Section 3.2). This strong correlation indicates that inter-region agreement is a key determinant of output quality: responses consistent with the consensus view from multiple regions are not only more factually accurate but also demonstrate higher reliability and lower hallucination rates.
+
+To this end, we propose Multi-Region Fusion Decoding (MRFD), a decoding strategy that aims to incorporate multi-perspective reasoning into the generation process. As illustrated in Figure 1, MRFD identifies several salient regions in the image based on attention mechanisms, then generates an initial response for each. By computing Jensen-Shannon Divergence (JSD) among these responses, the method derives a set of consistency-based reliability weights. These weights are used during decoding to combine predictions from all regions. Furthermore, inspired by CoT prompting, we construct region-aware prompts by combining the original question with the initial analysis of each region, preserving the localized context throughout the generation. The contributions of this work are:
+
+- We propose a training-free decoding method that integrates multiple region-level perspectives, weighted by inter-region consistency, to reduce hallucinations in LVLMs.  
+- We introduce a JSD-based scoring scheme to quantify agreement among region-wise responses and guide reliability-aware fusion.  
+- We design a region-aware prompting strategy to enhance contextual grounding during generation without modifying model parameters.
+
+# 2 Related Work
+
+LVLM Hallucinations and Grounding Deficiencies. Large Vision-Language Models (LVLMs) frequently 'hallucinate'—generating text unsupported
+
+by visual input (Ji et al., 2023; Liu et al., 2024). This often stems from flawed visual grounding due to various factors such as data biases or poor connections between vision and language components (Han et al., 2024; Jiang et al., 2024; Shu et al., 2025; Wang et al., 2023a). Although standard attention mechanisms (Anderson et al., 2018) and region-based analyzes (Li et al., 2022; Kamath et al., 2021) aim to improve grounding, LVLMs still struggle to reliably combine information from multiple, potentially conflicting, image regions or to assess their mutual consistency.
+
+Limitations in Advanced Decoding and Fusion Strategies. Advanced strategies to improve LVLM outputs also exhibit limitations. Chain-of-Thought (CoT) prompting (Wei et al., 2022; Zhang et al., 2023; Lyu et al., 2023), while enhancing reasoning, may not ensure its steps visually align with the image and can be sensitive to setup or resource-intensive. Many training-free corrective or contrastive decoding methods (e.g., (Li et al., 2023b; Leng et al., 2024; Favero et al., 2024; Woo et al., 2024; Huang et al., 2024; Zhang et al., 2025a; Wang et al., 2024)) struggle to capture finegrained local image features and risk missing valid information, resulting in poor grounding capabilities. Moreover, when attempting to fuse information from multiple sources—a concept for which consistency has proven beneficial in language modeling tasks (Wang et al., 2023b; Xiong et al., 2023)—LVLMs face distinct challenges with visual regions. Common simplistic aggregation techniques (like averaging) are often insufficient for visual data, where the trustworthiness of different regions can vary dramatically. A central challenge thus remains: developing principled methods to appropriately weigh and fuse evidence from diverse, contextually-understood visual regions based on their consistency.
+
+# 3 Motivation
+
+# 3.1 Global Decoding Misses Local Evidence
+
+To explore whether visual grounding capability is related to the hallucinations in LVLMs, we analyze their attention patterns during question answering. Figure 2 shows a case where the model is asked: "Is there a laptop in the image?" When using the full image as input, the model outputs "NO," even though a laptop is present. The attention map shows that the model distributes focus across unrelated regions.
+
+![](images/90f22129f540ca1d91f15366e04722a68c3d0c220dd037f56bbb1e56a2e9e82c.jpg)  
+Figure 2: LVLM cross-attention patterns for "Is there a laptop in the image?". (upper) Full image input results in scattered attention and potential error. (lower) Cropped image input focused on the laptop yields concentrated attention and improved accuracy.
+
+In contrast, when the image is cropped to a salient region based on high attention (the desk area), the model concentrates more narrowly on the relevant evidence and correctly answers "YES." The localized input guides the model to verify only what matters for the query, avoiding distractions from the rest of the image.
+
+This comparison reveals a key shortcoming of global decoding: attention is easily diffused across many areas, some of which may mislead the model. In cluttered or ambiguous scenes, this often results in factual errors. Cropped regions help narrow the visual focus of the model, improving the quality of the answer by strengthening the localized foundation.
+
+# 3.2 Region Consistency Reflects Reliability
+
+However, not all regions—even attention-guided ones—are equally reliable. Some may be visually ambiguous or contextually misleading. To avoid overconfident on a single (possibly misleading) region, we adopt a self-consistency approach: comparing multiple region-level responses. When different views agree on an answer, that answer is more likely to be trustworthy.
+
+To test this, we run experiments using LLaVA1.5 on 3,000 MSCOCO validation samples, an
+
+![](images/e627cca39ce37a7a4ec0a5f645225b357f505440429e13c8d81b88b2379233d8.jpg)  
+Figure 3: Density distribution of JS Divergence for correct versus hallucinated LVLM responses, indicating lower JSD correlates with higher factual accuracy.
+
+notated with hallucination labels from the POPE benchmark. For each image-question pair, we generate several responses from attention-guided image patches, along with one from the full image. For the output of each cropped region, we calculate the Jensen-Shannon divergence (JSD) between its response distribution and the average response distribution across all regions.
+
+Our results (illustrated in Figure 3) show a clear distinction: JS Divergence for hallucinated answers tends to concentrate around 0.06-0.07, whereas for correct answers, it centers around approximately 0.02. This indicates that the JSD between a specific regional response and the average across regions effectively reflects the level of hallucination of that regional response: a lower JSD correlates with fewer hallucinations. This finding resonates with the principle of self-consistency (Wang et al., 2023b), where agreement among multiple diverse outputs often signals higher quality and reliability. Motivated by this, we design a decoding strategy: A mechanism that aggregates evidence from multiple regions, and checks their agreement, can help improve the robustness and factual grounding of model predictions.
+
+# 4 Method
+
+# 4.1 Overview
+
+Current LVLM decoding methods often suffer from hallucinations due to the lack of multi-perspective consistency checking, as discussed in Section 3.
+
+To address this, we propose Multi-Region Fusion Decoding (MRFD), a training-free decoding strategy that enhances answer reliability through: 1) Attention-Guided Region Selection: selecting salient regions based on attention maps; 2) MultiRegion Analysis with JSD-Based Weighting: as
+
+![](images/65473899451427cdd3bb19db94c774e55ab7eb15a3f2f42b61d4a7faa0a9f11f.jpg)  
+Figure 4: Overall framework of Multi-Region Fusion Decoding (MRFD): Step 1 uses attention to select and crop salient regions  $(v_{k})$ , generates candidate responses  $(r_k)$  per region, and computes JSD-based consistency weights  $(w_{k})$  for each response. Step 2 forms new inputs per region with a candidate response and the original prompt. They are all processed in parallel, fusing per-region logits using the weights  $w_{k}$  during parallel decoding to select the output tokens.
+
+serving consistency across regions using Jensen-Shannon Divergence; 3) Consistency-Based Fusion Decoding: fusing predictions weighted by their mutual consistency. The full framework is illustrated in Figure 4.
+
+# 4.2 Attention-Guided Region Selection
+
+Building on the multi-perspective verification approach motivated in Section 3, the first crucial step in our MRFD's framework is to identify and isolate multiple informative regions from the input image for focused and independent analysis.
+
+Leveraging Cross-Attention for Region Identification. To identify image regions most relevant to the input query, we leverage the cross-attention mechanisms already present in LVLM architectures. Given an original image  $I_0$  and a text query  $x$ , we compute the attention matrix  $A$  from the LVLM's cross-attention layers. Let  $H_T \in \mathbb{R}^{n \times d}$  represent the sequence of hidden states from the textual input, and  $H_V \in \mathbb{R}^{m \times d}$  represent those from the visual input, where  $n$  and  $m$  denote sequence lengths and  $d$  represents the hidden dimensionality.
+
+The cross-attention weight matrix  $A \in \mathbb{R}^{n \times m}$ , capturing alignment between textual and visual modalities, is computed using scaled dot-product
+
+attention(Vaswani et al., 2017):
+
+$$
+\mathbf {A} = \operatorname {s o f t m a x} \left(\frac {\left(H _ {T} W ^ {Q}\right) \left(H _ {V} W ^ {K}\right) ^ {\top}}{\sqrt {d _ {k}}}\right) \tag {1}
+$$
+
+where  $W^{Q} \in \mathbb{R}^{d \times d_{k}}$  and  $W^{K} \in \mathbb{R}^{d \times d_{k}}$  are projection matrices for queries and keys, and  $d_{k}$  is their dimensionality.
+
+To obtain an overall map of visual focus, we aggregate these attention weights across the textual dimension:
+
+$$
+a _ {j} = \frac {1}{n} \sum_ {i = 1} ^ {n} A _ {i j} \quad \text {f o r} j = 1, \dots , m \tag {2}
+$$
+
+This produces an aggregated attention vector  $a \in \mathbb{R}^{1 \times m}$ . Assuming the visual hidden states correspond to a spatial grid of patches  $(d' \times d' = m)$ , we reshape this vector into a 2D spatial attention map:
+
+$$
+\hat {A} = \operatorname {r e s h a p e} \left(\mathbf {a}, \left(d ^ {\prime}, d ^ {\prime}\right)\right) \tag {3}
+$$
+
+This spatial attention map  $\hat{A}$  visually highlights the regions that the model finds most relevant to the input query. By utilizing the model's own attention patterns, we can directly identify which parts of the image are being primarily considered when answering the query.
+
+Selecting Salient Regions. Using the spatial attention map  $\hat{A}$ , we identify the top-K most salient
+
+regions  $R_{1}, R_{2}, \ldots, R_{K}$  within the original image. To ensure diversity and comprehensive coverage, we employ an integral image approach to efficiently search for K non-overlapping or minimally overlapping regions that maximize summed attention scores. For practical implementation, we constrain these regions to be rectangular bounding boxes with a maximum overlap (IoU) threshold between any two regions to ensure diversity in the selected perspectives.
+
+For each selected region  $R_{k}$ , we crop the original image to obtain a focused sub-image  $v_{k}$ . This process yields a set of  $\mathrm{K} + 1$  visual inputs: the original full image  $v_{0} = I_{0}$  and K cropped region images  $\{v_{1}, v_{2}, \ldots, v_{K}\}$ . These cropped regions represent the most informative parts of the image with respect to the input query, as determined by the model's own attention mechanism.
+
+# 4.3 Multi-Region Analysis with JSD-Based Weighting
+
+Having selected salient regions, we now detail their independent analysis and our JSD-based weighting mechanism. This approach is crucial for assessing inter-region consistency (Section 3.2). For each selected region  $R_{k}$ , we process the corresponding cropped sub-image  $v_{k}$  through the LVLM to generate an initial analysis:
+
+$$
+r _ {k} = \operatorname {L V L M} \left(v _ {k}, x\right) \tag {4}
+$$
+
+where  $r_k$  represents the LVLM's response when considering only the visual information from region  $k$ . We also generate a response  $r_0$  using the original full image  $v_0$ . The resulting set  $\{r_0, r_1, \dots, r_K\}$  provides multiple perspectives on the query, each grounded in different salient regions of the image.
+
+To assess the reliability of information from each region, we measure the consistency between region-specific analyses using Jensen-Shannon Divergence (JSD). The JSD provides a symmetric measure of similarity between probability distributions and is defined as:
+
+$$
+J S D (P | | Q) = \frac {1}{2} D _ {K L} (P | | M) + \frac {1}{2} D _ {K L} (Q | | M) \tag {5}
+$$
+
+where  $P$  and  $Q$  are two probability distributions,  $M = \frac{1}{2} (P + Q)$ , and  $D_{KL}$  is the Kullback-Leibler divergence.
+
+We derive a representative vocabulary distribution  $P_{k}$  for each initial analysis sequence  $r_k$  by
+
+averaging the next-token probability distributions computed during generation:
+
+$$
+P _ {k} = \operatorname {s o f t m a x} \left(\frac {1}{\left| r _ {k} \right|} \sum_ {t = 1} ^ {\left| r _ {k} \right|} \operatorname {l o g i t} _ {\theta} \left(y \mid r _ {k, <   t}, x, v _ {k}\right)\right) \tag {6}
+$$
+
+where  $|r_k|$  is the length of  $r_k$  and  $\log \mathrm{it}_{\theta}$  represents the model's output logits. This distribution  $P_{k}$  captures the overall token prediction patterns when the model considers region  $k$ . We then compute the average distribution across all regions:
+
+$$
+\bar {P} = \frac {1}{K + 1} \sum_ {i = 0} ^ {K} P _ {i} \tag {7}
+$$
+
+For each region  $k$ , we calculate its divergence from this average distribution:
+
+$$
+J _ {k} = J S D \left(P _ {k} \| \bar {P}\right) \tag {8}
+$$
+
+These divergence scores measure how much each region's predictions deviate from the consensus. A lower  $J_{k}$  indicates that region  $k$  provides information more consistent with other regions, suggesting higher reliability.
+
+Finally, we convert these divergence scores into weights with temperature parameter  $\gamma$ :
+
+$$
+w _ {k} = \frac {\exp \left(- J _ {k} / \gamma\right)}{\sum_ {i = 0} ^ {K} \exp \left(- J _ {i} / \gamma\right)} \tag {9}
+$$
+
+The temperature parameter  $\gamma$  controls the sharpness of the weight distribution—a smaller  $\gamma$  creates more contrast between weights, while a larger  $\gamma$  leads to more uniform weighting.
+
+# 4.4 JS-Weighted Integrative Fusion Decoding
+
+With the consistency-based weights  $w_{k}$  computed for each visual input  $v_{k}$  ( $k = 0\dots K$ ), we now perform the final decoding step by fusing the next-token predictions in a manner that prioritizes consistent visual evidence.
+
+Motivated by the effectiveness of Chain-of-Thought (CoT) prompting (Wei et al., 2022) in structuring reasoning, we adapt its core principle to enhance factual grounding within our multi-region fusion framework. For each region  $k$ , we construct a fixed region-aware prompt  $q_{k}$  by concatenating the original question  $x$  and the corresponding region's initial analysis  $r_{k}$ :
+
+$$
+q _ {k} = \operatorname {C o n c a t} (x, r _ {k}) \tag {10}
+$$
+
+This prompt  $q_{k}$  encapsulates both the global query and the localized preliminary analysis derived from  $v_{k}$ . It serves as a static, enriched context for region  $k$  throughout the decoding process, providing region-specific grounding information without requiring dynamic prompt updates.
+
+During autoregressive decoding, at each step  $t'$ , the LVLM produces next-token logits  $\ell_k^{(t')}$  conditioned on the current partial output sequence  $y_{< t'}$ , the visual input  $v_k$ , and the fixed region-aware prompt  $q_k$ :
+
+$$
+\ell_ {k} ^ {\left(t ^ {\prime}\right)} = \operatorname {l o g i t} _ {\theta} \left(y \mid y _ {<   t ^ {\prime}}, v _ {k}, q _ {k}\right) \tag {11}
+$$
+
+These logits represent the prediction from each region's perspective given its pre-computed analysis. We aggregate them using the pre-computed consistency weights  $w_{k}$ :
+
+$$
+\ell_ {\text {f u s e d}} ^ {(t ^ {\prime})} = \sum_ {k = 0} ^ {K} w _ {k} \cdot \ell_ {k} ^ {(t ^ {\prime})} \tag {12}
+$$
+
+The final probability distribution for the next token is obtained via softmax:
+
+$$
+P _ {\text {f u s e d}} ^ {(t ^ {\prime})} (y) = \operatorname {s o f t m a x} \left(\ell_ {\text {f u s e d}} ^ {(t ^ {\prime})}\right) \tag {13}
+$$
+
+The next token  $\hat{y}_{t'}$  is then selected (e.g., sampling) from this fused distribution, and the process repeats autoregressively until completion.
+
+# 5 Experiments
+
+# 5.1 Experimental Settings
+
+Evaluated LVLMs. We evaluate Multi-Region Fusion Decoding (MRFD) on two representative open-source LVLMs: LLaVA-1.5-7B (Liu et al., 2023a) and InstructBLIP-7B (Dai et al., 2023). These models adopt different vision-language interfaces—direct projection in LLaVA-1.5 and query-based encoding (Q-Former (Li et al., 2023a)) in InstructBLIP—enabling evaluation of MRFD's generality. MRFD is applied as a training-free, decoding-time procedure on frozen models. Architectural details are provided in Appendix A.
+
+Baselines. As a basic baseline, regular decoding samples tokens from the model's post-softmax output probabilities. These include contrastive methods such as VCD (Leng et al., 2024) and M3ID (Favero et al., 2024), robustness-enhancing methods like RITUAL (Woo et al., 2024), and approaches employing correction or feedback, such as DeGF (Zhang et al., 2025b) and Woodpecker (Yin
+
+et al., 2023). We also include comparisons with other relevant methods: HALC (Chen et al., 2024), and OPERA (Huang et al., 2024). Performance of these baselines is based on our re-implementations using publicly available code where possible. Detailed descriptions of each baseline's methodology are provided in Appendix B.
+
+Implementation Details. Across all experiments, our Multi-Region Fusion Decoding (MRFD) method selects  $K = 3$  salient regions and employs a temperature of  $\gamma = 0.02$  for the JSD-based weighting (Eq. 9). We utilize multinomial sampling for both stages of decoding. Other detailed settings are provided in Appendix C.
+
+# 5.2 Datasets and Benchmarks
+
+We evaluate MRFD on multiple benchmarks covering both hallucination detection and general vision-language understanding. Below we briefly describe the key datasets; full details and evaluation metrics are provided in Appendix D.
+
+POPE (Li et al., 2023c): A Yes/No QA benchmark for object existence hallucination, built from MSCOCO, A-OKVQA, and GQA with various negative sampling strategies.
+
+CHAIR (Rohrbach et al., 2018): Measures hallucinated object mentions in image captions via CHAIRi and CHAIRs scores on MSCOCO images. MME (Fu et al., 2023): We focus on the MME-Hallucination subset, which tests object, count, position, and attribute hallucinations.
+
+# 5.3 Results and Discussions
+
+Results on POPE. In Table 1, we compare the performance of our MRFD method against other baselines on the POPE benchmark (Li et al., 2023c) under three different negative sampling settings (Random, Popular, Adversarial), across both LLaVA-1.5 and InstructBLIP. As shown, MRFD consistently outperforms other decoding methods on both LVLMs, achieving leading F1 scores across all six configurations, with improvements of up to  $2.44\%$  in accuracy,  $6.52\%$  in precision, and  $3.05\%$  in F1 score compared to the respective second-best approaches. This suggests that MRFD's core strategy—identifying multiple salient regions via cross-attention, assessing their response consistency using JSD-weighting, and fusing their context-enriched predictions—enables LVLMs to better ground responses in relevant visual evidence, thereby effectively addressing object hallucinations.
+
+<table><tr><td rowspan="2">Setting</td><td rowspan="2">Method</td><td colspan="3">LLaVA-1.5 (Liu et al., 2023a)</td><td colspan="3">InstructBLIP (Dai et al., 2023)</td></tr><tr><td>Acc. ↑</td><td>Prec. ↑</td><td>F1 ↑</td><td>Acc. ↑</td><td>Prec. ↑</td><td>F1 ↑</td></tr><tr><td rowspan="6">Random</td><td>Regular</td><td>82.42</td><td>78.30</td><td>83.67</td><td>79.85</td><td>80.33</td><td>83.45</td></tr><tr><td>VCD (Leng et al., 2024)</td><td>84.69</td><td>80.30</td><td>85.85</td><td>84.47</td><td>83.07</td><td>84.81</td></tr><tr><td>M3ID (Favero et al., 2024)</td><td>85.46</td><td>81.54</td><td>86.42</td><td>85.32</td><td>83.51</td><td>85.58</td></tr><tr><td>RITUAL (Woo et al., 2024)</td><td>86.71</td><td>82.84</td><td>87.51</td><td>87.12</td><td>85.64</td><td>87.23</td></tr><tr><td>DeGF (Zhang et al., 2025b)</td><td>87.79</td><td>86.33</td><td>88.08</td><td>87.21</td><td>89.01</td><td>86.70</td></tr><tr><td>MRFD</td><td>88.15</td><td>88.91</td><td>88.23</td><td>88.03</td><td>88.74</td><td>88.01</td></tr><tr><td rowspan="6">Popular</td><td>Regular</td><td>76.57</td><td>71.23</td><td>79.56</td><td>75.17</td><td>70.90</td><td>77.54</td></tr><tr><td>VCD (Leng et al., 2024)</td><td>77.30</td><td>71.61</td><td>80.57</td><td>78.12</td><td>73.77</td><td>80.10</td></tr><tr><td>M3ID (Favero et al., 2024)</td><td>78.66</td><td>73.09</td><td>81.45</td><td>78.32</td><td>73.75</td><td>80.30</td></tr><tr><td>RITUAL (Woo et al., 2024)</td><td>79.75</td><td>74.55</td><td>82.31</td><td>78.40</td><td>73.63</td><td>80.55</td></tr><tr><td>DeGF (Zhang et al., 2025b)</td><td>81.94</td><td>78.33</td><td>83.31</td><td>80.37</td><td>78.84</td><td>81.24</td></tr><tr><td>MRFD</td><td>81.99</td><td>78.81</td><td>83.29</td><td>82.17</td><td>79.52</td><td>83.15</td></tr><tr><td rowspan="6">Adversarial</td><td>Regular</td><td>71.09</td><td>65.77</td><td>75.93</td><td>71.02</td><td>66.58</td><td>74.70</td></tr><tr><td>VCD (Leng et al., 2024)</td><td>71.13</td><td>65.28</td><td>76.37</td><td>73.07</td><td>68.50</td><td>76.36</td></tr><tr><td>M3ID (Favero et al., 2024)</td><td>72.10</td><td>66.27</td><td>76.87</td><td>72.97</td><td>67.92</td><td>76.55</td></tr><tr><td>RITUAL (Woo et al., 2024)</td><td>71.87</td><td>66.17</td><td>76.88</td><td>73.06</td><td>67.83</td><td>76.91</td></tr><tr><td>DeGF (Zhang et al., 2025b)</td><td>76.13</td><td>71.50</td><td>79.01</td><td>75.96</td><td>73.45</td><td>77.36</td></tr><tr><td>MRFD</td><td>77.99</td><td>76.16</td><td>79.22</td><td>77.72</td><td>74.26</td><td>79.72</td></tr></table>
+
+Table 1: Results on POPE benchmark. Higher (↑) accuracy, precision, and F1 indicate better performance. The best results are bolded, and the second-best are underlined.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">LLaVA-1.5</td><td colspan="2">InstructBLIP</td></tr><tr><td>Cs ↓</td><td>Ci ↓</td><td>Cs ↓</td><td>Ci ↓</td></tr><tr><td>Regular</td><td>26.2</td><td>9.4</td><td>31.2</td><td>11.1</td></tr><tr><td>VCD</td><td>24.4</td><td>7.9</td><td>30.0</td><td>10.1</td></tr><tr><td>M3ID</td><td>21.4</td><td>6.3</td><td>30.8</td><td>10.4</td></tr><tr><td>RITUAL</td><td>22.4</td><td>6.9</td><td>26.6</td><td>8.9</td></tr><tr><td>Woodpecker</td><td>24.9</td><td>7.5</td><td>31.2</td><td>10.8</td></tr><tr><td>HALC</td><td>21.7</td><td>7.1</td><td>24.5</td><td>8.0</td></tr><tr><td>DeGF</td><td>18.4</td><td>6.1</td><td>24.0</td><td>7.7</td></tr><tr><td>Ours (MRFD)</td><td>14.1</td><td>5.0</td><td>21.3</td><td>6.1</td></tr></table>
+
+Table 2: Results on CHAIR benchmark for caption generation. We limited the maximum number of new tokens to 64. Lower  $(\downarrow)$  CHAIRs (Cs) and CHAIRi (Ci) indicate less hallucination. Best results are bolded, second-best are underlined.
+
+Moreover, while most decoding method tend to be overcofindent in nonexistence, the consistency verification inherent in MRFD appears to promote more cautious and precise responses, which is evidenced by its strong precision, particularly in challenging adversarial settings (e.g., 76.16 on LLaVA-1.5 and 74.26 on InstructBLIP), highlighting its enhanced capability in filtering false positives and suppressing misinformation. Detailed results of POPE are attached in Appendix E.3.
+
+Results on CHAIR. We evaluate MRFD's effectiveness in mitigating object hallucination in open
+
+ended image captioning using the CHAIR benchmark (Rohrbach et al., 2018), reporting CHAIRs (Cs) and CHAIRi (Ci) scores (lower is better) for LLaVA-1.5 and InstructBLIP in Table 2. MRFD consistently achieves state-of-the-art performance on both LVLMs, significantly outperforming the strong DeGF baseline with relative CHAIR score reductions of up to  $16.4\%$  on LLaVA-1.5 and  $20.8\%$  on InstructBLIP. Given that the CHAIR task for image captioning necessitates comprehensive attention to diverse local details, MRFD's superior performance underscores its advanced capability for robust multi-region analysis and effective integration of key information from various visual segments, leading to more factually grounded image captions. Detailed results of CHAIR are appended to Appendix E.1. Some qualitative examples are shown in Appendix H.
+
+Results on MME-Hallucination. We evaluate MRFD on the MME hallucination subset (Fu et al., 2023), which assesses diverse hallucination types including object-level (Existence), attribute-level (Count, Color), and relation-level (Position) understanding in LVLMs. As illustrated in Figure 5 (presenting results for both LLaVA-1.5 and Instruct-BLIP), our MRFD method consistently demonstrates strong performance, achieving leading or state-of-the-art results across the majority of these hallucination categories for both evaluated LVLMs
+
+![](images/8f79f88f8d22f4444f18a3306afa2710752fcccd3de3b1ca7e4edbd63c7304c7.jpg)  
+Figure 5: Experimental results of MME on a hallucination subset with different decoding strategies.
+
+when compared to regular decoding and other advanced baselines. This robust performance across a spectrum of challenging hallucination types underscores that MRFD's core mechanism—multi-region analysis coupled with consistency-driven fusion—effectively enhances fine-grained visual understanding and overall factual consistency. Detialed results of MME-Hallucination are appended to Appendix E.2
+
+Efficiency Discussion. Despite employing multiple inference steps, MRFD maintains a competitive efficiency profile, particularly when compared to other advanced hallucination mitigation techniques that involve more complex iterative or feedback mechanisms. Details are in Appendix F.
+
+# 5.4 Ablation Study
+
+We conduct ablation studies to evaluate the impact of key components and hyperparameters in MRFD, using LLaVA-1.5 on the POPE-MSCOCO dataset.
+
+First, we assess the contributions of MRFD's core designs: JSD-based Consistency Weighting (CW), Fusion Prompt (FP), and attention-guided Region Selection (RS). We test our full MRFD against three main variants where these components are individually altered: (1) MRFD w/o CW, applying uniform fusion weights; (2) MRFD w/o FP, using only the original question for regional decoding; and (3) MRFD w/o RS, processing only the global image through the subsequent pipeline stages. As detailed in Table 3, full MRFD achieves an 86.21 F1 score. Removing CW degrades F1 performance by  $2.94\%$ , underscoring the importance of dynamic, consistency-based weighting. Omitting FP reduces F1 by  $3.93\%$ , highlighting the value of enriched regional context. Bypassing RS results in the largest F1 drop of  $4.07\%$ , emphasizing that robust multi-region analysis is funda
+
+<table><tr><td>Model Variants</td><td>Acc. ↑</td><td>Prec. ↑</td><td>F1 Score ↑</td></tr><tr><td>MRFD (Full)</td><td>86.50</td><td>88.11</td><td>86.21</td></tr><tr><td>w/o CW</td><td>83.76</td><td>84.27</td><td>83.74</td></tr><tr><td>w/o FP</td><td>82.87</td><td>83.95</td><td>82.70</td></tr><tr><td>w/o RS</td><td>82.77</td><td>83.94</td><td>82.58</td></tr><tr><td>Regular</td><td>80.57</td><td>78.84</td><td>81.59</td></tr></table>
+
+Table 3: Ablation study with different model variants on POPE-COCO under the average of three settings.
+
+![](images/ee109ca233f31774501c118b14e670233bfe3e1d596ee59068058952834e441e.jpg)  
+Figure 6: Sensitivity of POPE-COCO Precision (LLaVA-1.5) to JSD weighting temperature  $\gamma$  for  $K = 2,3,4$  regions, with optimal performance around  $\gamma = 0.02$ .
+
+mental. Despite these impacts, all three ablated MRFD variants still outperform Regular decoding (81.59 F1), while the complete MRFD configuration showcases the strongest synergistic benefits.
+
+Second, we analyze the sensitivity to the JSD weighting temperature  $\gamma$  (Eq. 9). As depicted in Figure 6 (POPE COCO Precision for  $K = 2,3,4$ ), performance peaks around  $\gamma = 0.02$  ( $K = 3$  at  $\approx 87.8\%$  Prec.). Higher  $\gamma$  values ( $\geq 0.08$ ) lead to more uniform weights and correspondingly reduced precision, approaching unweighted fusion. Conversely, very low  $\gamma$  values ( $\approx 0.01$ ) cause overreliance on a few regions due to extreme weighting, which can negatively impact robustness, particularly for larger  $K$  (e.g., performance degradation for  $K = 4$ ). Thus,  $\gamma = 0.02$  is adopted for providing an optimal balance. Further details are presented in Appendix G.
+
+# 6 Conclusion
+
+We presented Multi-Region Fusion Decoding (MRFD), a training-free approach that improves LVLM reliability by mimicking self-consistency multi-view verification. MRFD identifies salient regions via attention, estimates their reliability using Jensen-Shannon Divergence (JSD), and
+
+fuses region-level predictions using consistency-weighted, Chain-of-Thought-inspired prompts. Experiments show that MRFD effectively reduces hallucinations and improves factuality across various LVLMs and benchmarks, offering a simple yet robust decoding strategy.
+
+# Limitations
+
+MRFD relies on the quality and interpretability of attention maps produced by the underlying LVLM, which may vary across architectures and input conditions. The assumption that inter-region consistency indicates factual correctness holds empirically but may not generalize to all reasoning tasks, especially those requiring abstract or commonsense inference. Additionally, while MRFD improves factuality in image-grounded tasks, it has not been evaluated in broader multimodal contexts such as video or dialogue-based grounding, which may involve more complex temporal or conversational dependencies.
+
+# Ethics Statement
+
+Our work on Multi-Region Fusion Decoding (MRFD) aims to improve the reliability and factual grounding of LVLMs, thereby contributing to more trustworthy AI systems by reducing visual hallucinations. We believe this has positive implications for applications where factual accuracy is critical.
+
+# Acknowledgements
+
+The work is partially supported by the NSF of the United States Grant CRII 2451683, an NVIDIA Academic Grants Program, University of California at Merced, and a UC Merced Faculty Research Award. The views and conclusions are those of the authors and should not reflect the official policy or position of the U.S. Government.
+
+We also thank Prof. Kai-Wei Chang (Department of Computer Science, UCLA) for his insightful feedback on earlier drafts, including suggestions on clarifying terminology, adding appropriate cognitive-science references, discussing experimental settings, and presenting qualitative examples, as well as for pointing us to relevant related work. Any remaining errors are our own.
+
+# References
+
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, and 1 others. 2022. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 23716-23737.  
+Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6077-6086.  
+Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. 2024. Halc: Object hallucination reduction via adaptive focal-contrast decoding. Preprint, arXiv:2403.00425.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality. Preprint, arXiv:2303.13617.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. Instructclip: Towards general-purpose vision-language models with instruction tuning. Preprint, arXiv:2305.06500.  
+Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. 2024. Multi-modal hallucination control by visual information grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. IEEE / CVF.  
+Mingqian Feng, Yunlong Tang, Zeliang Zhang, and Chenliang Xu. 2024. Do more details always introduce more hallucinations in lvlm-based image captioning? arXiv preprint arXiv:2406.12663.  
+Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Chen, Qing Predicted, Yu Qiao, and Ping Luo. 2023. Mme: A comprehensive evaluation benchmark for multimodal large language models. Preprint, arXiv:2306.13394. MME-Hallucination evaluation is part of this benchmark.  
+Sanyam Gunjal, Ben Malin, Florian Golemo, Kiriakos N. Kutulakos, and Aaron Courville. 2024. Mitigating hallucinations in vision-language models via fine-grained vision-code data augmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3306-3316.
+
+Tianrui Han, Ruisheng Zhang, and Siyu Li. 2024. Analyzing and mitigating object hallucination in large vision-language models from perspective of training data bias. arXiv preprint arXiv:2402.16011.  
+Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 2024. OPERA: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. IEEE / CVF.  
+Drew A Hudson and Christopher D Manning. 2019. GQA: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6700-6709, Long Beach, CA, USA. IEEE / CVF.  
+Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. 2021. Openclip. Version 2.20.0.  
+Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Jenny Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38.  
+Yifan Jiang, Zhengyuan Zhu, Tianjun Zhao, and Dazchen She. 2024. Revealing and mitigating hallucinations in large vision-language models. arXiv preprint arXiv:2401.04071.  
+Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. 2021. Mdetr - modulated detection for end-to-end multi-modal understanding. Preprint, arXiv:2104.12763.  
+Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. 2024. VOLCANO: Mitigating multimodal hallucination through self-feedback guided revision. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), Mexico City, Mexico. Association for Computational Linguistics.  
+Yifan Leng, Yifan Li, Jiashi Zhang, Zidong Liu, Fang Wen, and Kaixiong Chen. 2024. Mitigating object hallucinations in large vision-language models via visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. IEEE / CVF.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023a. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning (ICML),
+
+volume 202 of Proceedings of Machine Learning Research, pages 19730-19748. PMLR.  
+Lianian Harold Li, Pengchuan Li, Zhe Zhang, Kuaida Lee, Peng Li, Yelong Zhang, Kathy Lee, Steven Hoi, Jenq-Neng Wang, Linjie Wang, Luke Zettlemoyer, and Ching Yun Liu. 2022. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10965-10975.  
+Xiang Li, Lin Li, Zihang Geng, Percy Liu, and Jian Liu. 2023b. Contrastive decoding: Open-ended text generation as optimization. In International Conference on Learning Representations (ICLR).  
+Yanshu Li, Yi Cao, Hongyang He, Qisen Cheng, Xiang Fu, Xi Xiao, Tianyang Wang, and Ruixiang Tang. 2025a.  $\mathbf{M}^{2}\mathrm{iv}$ : Towards efficient and fine-grained multimodal in-context learning via representation engineering. Preprint, arXiv:2504.04633.  
+Yanshu Li, Tian Yun, Jianjiang Yang, Pinyuan Feng, Jina Huang, and Ruixiang Tang. 2025b. Taco: Enhancing multimodal in-context learning via task mapping-guided sequence configuration. arXiv preprint arXiv:2505.17098.  
+Yifan Li, Yifan Du, Kun Wang, Yilin Zhao, Jiashi Zhang, Zidong Liu, Kaixiong Chen, and Fang Wen. 2023c. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 11221-11233, Singapore. Association for Computational Linguistics.  
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. 2014. Microsoft COCO: Common objects in context. In European conference on computer vision (ECCV), pages 740-755. Springer.  
+Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. 2024. A survey on hallucination in large vision-language models. Preprint, arXiv:2402.00253.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023a. Improved baselines with visual instruction tuning. Preprint, arXiv:2310.03744.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023b. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, pages 68178-68198.  
+Yuanhan Liu, Haochen Wang, Jingwen Li, Yanwei Fu, Chunyuan Li, and Jianfeng Gao. 2023c. MMBench: Is your multi-modal model an all-around player? Preprint, arXiv:2307.06281.  
+Qing Lyu, Shreya Chen, Johnny H GCL, Shichang Zhao, and Abhinav Shrivastava. 2023. Faithful chain-of-thought reasoning. arXiv preprint arXiv:2301.13379.
+
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. Preprint, arXiv:2103.00020.  
+Anna Rohrbach, Marcus Rohrbach, Rong Hu, Trevor Darrell, and Bernt Schiele. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4243-4248, Brussels, Belgium. Association for Computational Linguistics.  
+Dustin Schwenk, M. R. Magnusson, Borun Pan, Marco Zompero, Ferda Ofli, Michael Feldman, Russel Gabbard, Florian Metze, Adam Trischler, and Thomas Testole. 2022. A-OKVQA: A benchmark for visual question answering using world knowledge. In Proceedings of the European Conference on Computer Vision (ECCV), pages 541-558, Tel Aviv, Israel. Springer Nature Switzerland.  
+Dong Shu, Haiyan Zhao, Jingyu Hu, Weiru Liu, Ali Payani, Lu Cheng, and Mengnan Du. 2025. Large vision-language model alignment and misalignment: A survey through the lens of explainability. Preprint, arXiv:2501.01346.  
+Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. 2023. Eva-clip: Improved training techniques for clip at scale. Preprint, arXiv:2303.15389.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc.  
+P. Viola and M. Jones. 2001. Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 1, pages I-I.  
+Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. 2023a. Evaluation and analysis of hallucination in large vision-language models. Preprint, arXiv:2308.15126.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations (ICLR).  
+Zhaobin Wang, Xinyi Li, and Jing Chen. 2024. Mitigating hallucination in lvlms via disturbance instruction. arXiv preprint arXiv:2401.12876.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, and Denny
+
+Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 24824-24837. Curran Associates, Inc.  
+Sangmin Woo, Jaehyuk Jang, Donguk Kim, Yubin Choi, and Changick Kim. 2024. Ritual: Random image transformations as a universal anti-hallucination lever in large vision language models. Preprint, arXiv:2405.17821.  
+Haotian Xia, Zhengbang Yang, Yun Zhao, Yuqing Wang, Jingxi Li, Rhys Tracy, Zhuangdi Zhu, Yuan fang Wang, Hanjie Chen, and Weining Shen. 2024. Language and multimodal models in sports: A survey of datasets and applications. Preprint, arXiv:2406.12252.  
+Chen Xiong, Zhi Liu, Le Wang, and Lifan Liu. 2023. Can llms evaluate their own accuracy? a study on math problem solving. arXiv preprint arXiv:2305.07331.  
+Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Chen, Enhong Chen, Gaoang Wang, Qiao Yu, and Ping Luo. 2023. Woodpecker: Hallucination correction for multimodal large language models. Preprint, arXiv:2310.16045.  
+Ce Zhang, Zifu Wan, Zhehan Kan, Martin Q. Ma, Simon Stepputtis, Deva Ramanan, Russ Salakhutdinov, Louis-Philippe Morency, Katia Sycara, and Yaqi Xie. 2025a. Self-correcting decoding with generative feedback for mitigating hallucinations in large vision-language models. Preprint, arXiv:2502.06130.  
+Ce Zhang, Zifu Wan, Zhehan Kan, Martin Q. Ma, Simon Stepputtis, Deva Ramanan, Russ Salakhutdinov, Louis-Philippe Morency, Katia P. Sycara, and Yaqi Xie. 2025b. Self-correcting decoding with generative feedback for mitigating hallucinations in large vision-language models. In The Thirteenth International Conference on Learning Representations (ICLR).  
+Wenqiao Zhang, Haochen Shi, Siliang Tang, Jun Xiao, Qiang Yu, and Yueting Zhuang. 2021. Consensus graph representation learning for better grounded image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Zhuosheng Zhang, Jing Liu, Zheyuan Luo, Shane Storks, Zhe Li, Binhang Wang, Chen Chen, Iryna Gurevych, Honglak Lee, Graham Liu, and Aston Zhang. 2023. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923.  
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592.
+
+# A Evaluated LVLMs
+
+We evaluate our proposed Multi-Region Fusion Decoding (MRFD) framework on three representative open-source Large Vision-Language Models (VLMs): LLaVA-1.5-7B (Liu et al., 2023a), and InstructBLIP-7B (Dai et al., 2023). For the visual encoder, LLaVA-1.5 uses ViT-L-336px pre-trained from CLIP-L/14-336px (Radford et al., 2021). InstructBLIP uses ViT-g/14 pre-trained from EVA-CLIP (Sun et al., 2023). Qwen-VL uses Openclip ViT-bigG (Ilharco et al., 2021). Regarding the language model module, LLaVA-1.5 and InstructBLIP both utilize Vicuna-7B (Chiang et al., 2023).
+
+The vision-language interface varies across the evaluated models. LLaVA-1.5 directly projects visual features using MLP layers. InstructBLIP, conversely, adopts the Q-Former (Li et al., 2023a), which processes visual features via 32 learnable queries to generate a standardized input for the LLM. Our evaluation of MRFD encompasses these distinct strategies—direct MLP projection and query-based bottleneck (Q-Former)—to demonstrate its broad effectiveness in mitigating hallucinations regardless of the connection module employed.
+
+# B Baselines
+
+We compare the performance of our MRFD method with several recent training-free decoding approaches designed to mitigate hallucination. These key baselines, along with their core methodologies, are described below:
+
+- VCD (Leng et al., 2024): This method contrasts output distributions derived from an original visual input  $(v)$  and a distorted version  $(v^{\prime})$ . Given a textual query and  $v$ , the model generates two distributions. The distorted input  $v^{\prime}$  is created using pre-defined distortions (e.g., a Gaussian noise mask) to  $v$ . These two distributions are then contrasted to form the final output probability. For reproduction, we follow VCD's default setting with  $\alpha = 1$  (a parameter controlling the contrast strength) and use 500 noise steps to generate  $v^{\prime}$ . VCD aims to enhance visual grounding by reducing reliance on language priors.
+
+- M3ID (Favero et al., 2024): M3ID contrasts output distributions from original visual inputs against those from pure text inputs (lacking visual information). The final probability
+
+distribution is a combination of the distribution conditioned on both vision and text, and a contrastive term derived from the difference between vision-conditioned and text-only conditioned distributions, balanced by a hyperparameter  $\lambda$ . We adhere to their recommended  $\lambda = 0.02$ . M3ID also seeks to improve visual grounding by emphasizing visual information.
+
+- RITUAL (Woo et al., 2024): RITUAL applies common image transformations (e.g., crop, flip, color jitter) to the original visual input  $v$ , creating a transformed version  $v^{(T)}$ . It then generates the response by utilizing information from both the original  $(v)$  and transformed  $(v^{(T)})$  images. The final probability distribution combines logits from both views, with the contribution of the transformed input adjusted by a balancing hyperparameter  $\kappa$ . We follow their official implementation, setting  $\kappa = 3$ . This approach aims to improve robustness through consistency across augmented views.
+
+- DeGF (Zhang et al., 2025b): DeGF introduces a self-correction mechanism using feedback from text-to-image generative models. Specifically, it first generates an image based on the LVLM's initial textual response. This generated image then acts as an auxiliary visual reference, providing self-feedback to the LVLM to verify and correct its initial response, often through complementary or contrastive decoding techniques.
+
+- Woodpecker (Yin et al., 2023): Woodpecker is a post-hoc correction framework designed to mitigate hallucinations in the outputs of Multimodal Large Language Models (MLLMs). It operates by first prompting the MLLM itself to identify potential hallucinations (across several predefined types like object existence, attributes, etc.) in its initial response. If hallucinations are detected, Woodpecker then instructs the MLLM to revise and correct these identified errors.
+
+- HALC (Chen et al., 2024): HALC (e.g., "Mitigating Object Hallucinations in Large Vision-Language Models via Cause Analysis and Post-hoc Correction") is a post-hoc method that first analyzes the potential causes of object hallucinations to identify objects in the re
+
+sponse that are likely to be hallucinated. Subsequently, it instructs the LVLM to verify the existence of these specific, suspect objects within the image and make corrections if they are indeed confirmed as hallucinations.
+
+- OPERA (Huang et al., 2024): OPERA aims to alleviate visual relation hallucinations in LVLMs during decoding. It consists of two main components: an Over-trust Penalty (OP) term designed to penalize the model's overconfidence on unreliable visual relations during token generation, and a Retrospection-Allocation (RA) mechanism that encourages the model to retrospect previously generated tokens and re-allocate attention to relevant visual regions for verification and potential correction.
+
+We report the performance of these baselines based on our re-implementation using their released code bases where available.
+
+# C Implementation Details
+
+In all experiments using MRFD, we set the number of regions  $K = 3$  to obtain cropped images and the temperature  $\gamma = 0.02$  for JSD-based weighting (Eq. 9). For the decoding process, we employ multinomial sampling in both stages. Specifically, during the first step to generate the initial analyses  $r_k$ , we use a sampling temperature of 0.7. In the second step to generate the final output sequence  $y$ , we use a lower sampling temperature of 0.1 after fusing the logits.
+
+To obtain the aggregated spatial attention map  $\hat{A}$  (Eq. 3), we first average the attention weights across all attention heads within the relevant layer(s). For LLaVA-1.5, we utilize the attention map from the final cross-attention layer. For InstructBLIP, which uses Q-Former, we identify the query token with the maximum aggregated attention score and use its corresponding attention map. The spatial dimension  $d'$  of the map  $\hat{A}$  corresponds to the grid size of the visual patches ( $m = d' \times d'$ ), which is  $24 \times 24$  for LLaVA-1.5 (ViT-L-336px) and  $16 \times 16$  for the ViT-g/14 used by InstructBLIP.
+
+To efficiently identify the top-K salient regions  $\{R_1,\dots,R_K\}$  based on  $\hat{A}$ , we employ an integral image approach (Viola and Jones, 2001) to quickly calculate the sum of attention scores within any rectangular bounding box. We search for the K non-overlapping or minimally overlapping rectan
+
+gular regions that maximize these summed attention scores. To ensure diversity in the selected regions, we enforce a maximum Intersection over Union (IoU) of  $40\%$  between any pair of selected bounding boxes  $R_{i}$  and  $R_{j}$ $(i\neq j)$ .
+
+# D Datasets and Benchmarks
+
+We evaluate our MRFD framework on a diverse set of benchmarks targeting both hallucination detection and general vision-language capabilities.
+
+- POPE (Li et al., 2023c): POPE (Polling-based Object Probing Evaluation) is a widely used benchmark for assessing object existence hallucination in LVLMs. It presents models with Yes/No questions concerning the presence of specific objects (e.g., "Is there a {object} in the image?"). The benchmark data is structured into three main subsets derived from MSCOCO (Lin et al., 2014), A-OKVQA (Schwenk et al., 2022), and GQA (Hudson and Manning, 2019). Each of these subsets is further divided based on three negative sampling strategies for nonexistent objects: random, popular, and adversarial, which vary in difficulty. For evaluation, we report standard metrics including Accuracy, Precision, Recall, and F1 score.
+
+- MME (Fu et al., 2023): MME serves as a comprehensive benchmark for evaluating overall LVLM perception and cognition. Our evaluation specifically utilizes the MME-Hallucination subset, which is designed to assess a range of common hallucination types. These are categorized into object-level assessments like existence and count, and attribute-level assessments such as object position and color. Questions in this subset are typically Yes/No queries. We report scores based on the official benchmark protocol, which often involves combined accuracy measures reflecting both question-level and image-level correctness.
+
+- CHAIR (Rohrbach et al., 2018): The CHAIR (Caption Hallucination Assessment with Image Relevance) benchmark quantifies object hallucinations within the context of open-ended image captioning. LVLMs are prompted to generate descriptive captions for images, for which we, following prior work (Lee et al., 2024), use a random selection
+
+of 500 images from the MSCOCO (Lin et al., 2014) validation set. The generated captions are then compared against ground-truth objects within the image to calculate the CHAIRi (instance-level) and CHAIRs (category-level) scores, where lower scores indicate fewer hallucinations.
+
+$$
+\begin{array}{l} \mathrm {C H A I R} _ {S} = \frac {\text {c a p t i o n s w / h a l l u c i n a t e d o b j e c t s}}{\text {a l l c a p t i o n s}}, \\ \mathrm {C H A I R} _ {I} = \frac {\text {h a l l u c i n a t e d o b j e c t s}}{\text {a l l m e n t i o n e d o b j e c t s}}. \end{array} \tag {14}
+$$
+
+MMBench (Liu et al., 2023c): MMBench evaluates a broad spectrum of multimodal capabilities through carefully curated multiple-choice questions that span various cognitive dimensions and skills. For this benchmark, we adhere to the official evaluation protocol and report the overall accuracy score.
+
+# E Detailed Results of experiments
+
+# E.1 Detailed Results of CHAIR
+
+Detailed results of CHAIR are shown in Table 4 and Table 5, reporting CHAIRs (Cs) and CHAIRi (Ci) scores (lower is better) for LLaVA-1.5 and InstructBLIP. MRFD consistently achieves state-of-the-art performance on both LVLMs.
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">LLaVA-1.5</td><td colspan="2">InstructBLIP</td></tr><tr><td>Cs ↓</td><td>Ci ↓</td><td>Cs ↓</td><td>Ci ↓</td></tr><tr><td>Regular</td><td>55.0</td><td>16.3</td><td>57.0</td><td>17.6</td></tr><tr><td>VCD</td><td>54.4</td><td>16.6</td><td>60.4</td><td>17.8</td></tr><tr><td>M3ID</td><td>56.6</td><td>15.7</td><td>62.2</td><td>18.1</td></tr><tr><td>RITUAL</td><td>49.6</td><td>14.8</td><td>48.4</td><td>14.5</td></tr><tr><td>Woodpecker</td><td>57.6</td><td>16.7</td><td>60.8</td><td>17.6</td></tr><tr><td>HALC</td><td>51.0</td><td>14.8</td><td>53.8</td><td>15.7</td></tr><tr><td>DeGF</td><td>48.8</td><td>14.6</td><td>49.2</td><td>14.4</td></tr><tr><td>Ours (MRFD)</td><td>37.1</td><td>9.2</td><td>38.2</td><td>10.9</td></tr></table>
+
+# E.2 Detailed Results of MME
+
+In table 7 and Figure 7, we provide detailed results on the MME-Hallucination benchmark (Fu et al.,
+
+Table 4: Results on CHAIR benchmark for caption generation. We limited the maximum number of new tokens to 128. Lower  $(\downarrow)$  CHAIRs (Cs) and CHAIRi (Ci) indicate less hallucination. Best results are bolded, second-best are underlined.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">LLaVA-1.5</td><td colspan="2">InstructBLIP</td></tr><tr><td>Cs ↓</td><td>Ci ↓</td><td>Cs ↓</td><td>Ci ↓</td></tr><tr><td>Regular</td><td>58.0</td><td>17.7</td><td>61.0</td><td>18.2</td></tr><tr><td>VCD</td><td>58.2</td><td>16.7</td><td>63.0</td><td>18.6</td></tr><tr><td>M3ID</td><td>56.8</td><td>16.1</td><td>65.8</td><td>19.9</td></tr><tr><td>RITUAL</td><td>51.0</td><td>15.1</td><td>50.4</td><td>15.3</td></tr><tr><td>DeGF</td><td>49.8</td><td>14.7</td><td>49.8</td><td>15.1</td></tr><tr><td>Ours (MRFD)</td><td>39.0</td><td>11.0</td><td>38.6</td><td>11.3</td></tr></table>
+
+Table 5: Results on CHAIR benchmark for caption generation. We limited the maximum number of new tokens to 256. Lower  $(\downarrow)$  CHAIRs (Cs) and CHAIRi (Ci) indicate less hallucination. Best results are bolded, second-best are underlined.
+
+2023) for both LLaVA-1.5 and InstructBLIP. The table includes scores for object-level (existence, count) and attribute-level (position, color) tasks, and averages the score across three random seeds. The best results are bolded, and the second-best are underlined.
+
+![](images/68b9802205805cd9b32976ab2cced55e744999b8f7c5ac76b93f3a3234f6d5f9.jpg)
+
+![](images/a6c391d87a68fb4b373d39243ff19ab12205447372bb1152ce186cdfc4a9c2c4.jpg)  
+Figure 7: Experimental results of MME with different decoding strategies.
+
+# E.3 Detailed Results of POPE
+
+In table 6, we provide detailed results on the POPE benchmark (Li et al., 2023c) for both LLaVA-1.5 and InstructBLIP across three different settings (Random, Popular, Adversarial). The table includes accuracy (Acc.), precision (Prec.), recall,
+
+and F1 scores for each method. The best results are bolded, and the second-best are underlined.
+
+# F Efficiency Comparison
+
+We report MRFD's efficiency on the CHAIR benchmark with two backbones—InstructBLIP-7B and LLaVA-1.5-7B—under the same decoding setup (128 max tokens, RTX 3090 GPU). With multi-region  $(K = 3)$  analysis and fusion, MRFD increases latency by about  $\sim 2.84\times$  (from  $3.59\mathrm{s}$  to  $10.21\mathrm{s}$ ) and peak GPU memory by  $\sim 1.11\times$  (from  $15,238\mathrm{MB}$  to  $16,932\mathrm{MB}$ ) relative to Regular decoding on InstructBLIP-7B (Table 8). On LLaVA-1.5-7B, MRFD shows a  $\sim 3.30\times$  latency increase ( $3.44\mathrm{s} \rightarrow 11.34\mathrm{s}$ ) and a  $\sim 1.11\times$  memory increase ( $15,778\mathrm{MB} \rightarrow 17,458\mathrm{MB}$ ) over Regular (Table 9). Despite this overhead—and compared to lighter baselines such as VCD or the posthoc Woodpecker—MRFD remains markedly more efficient than iterative/feedback-heavy methods (DeGF, OPERA, HALC), while achieving the best CHAIRs (lower is better): 38.2 on InstructBLIP-7B and 37.1 on LLaVA-1.5-7B, representing  $\sim 33\%$  and  $\sim 32.5\%$  reductions from their respective Regular baselines. Overall, MRFD offers a compelling balance between computational cost and hallucination reduction.
+
+<table><tr><td>Method</td><td>Avg. Lat. ↓</td><td>GPU Mem. ↓</td><td>CHAIRs ↓</td></tr><tr><td>Regular</td><td>3.59 s</td><td>15,238 MB</td><td>57.0</td></tr><tr><td>VCD</td><td>6.82 s</td><td>16,078 MB</td><td>60.4</td></tr><tr><td>OPERA</td><td>25.31 s</td><td>23,575 MB</td><td>62.2</td></tr><tr><td>Woodpecker</td><td>11.72 s</td><td>22,798 MB</td><td>60.8</td></tr><tr><td>HALC</td><td>21.84 s</td><td>22,689 MB</td><td>53.8</td></tr><tr><td>DeGF</td><td>14.69 s</td><td>18,726 MB</td><td>49.2</td></tr><tr><td>Ours (MRFD)</td><td>10.21 s</td><td>16,932 MB</td><td>38.2</td></tr></table>
+
+Table 8: Efficiency comparison on InstructBLIP-7B. Lower is better for all metrics.  
+
+<table><tr><td>Method</td><td>Avg. Lat. ↓</td><td>GPU Mem. ↓</td><td>CHAIRs ↓</td></tr><tr><td>Regular</td><td>3.44 s</td><td>15,778 MB</td><td>55.0</td></tr><tr><td>VCD</td><td>6.91 s</td><td>16,634 MB</td><td>54.4</td></tr><tr><td>OPERA</td><td>24.70 s</td><td>22,706 MB</td><td>52.6</td></tr><tr><td>Woodpecker</td><td>10.68 s</td><td>22,199 MB</td><td>57.6</td></tr><tr><td>HALC</td><td>22.61 s</td><td>23,084 MB</td><td>51.0</td></tr><tr><td>DeGF</td><td>13.89 s</td><td>19,119 MB</td><td>48.8</td></tr><tr><td>Ours (MRFD)</td><td>11.34 s</td><td>17,458 MB</td><td>37.1</td></tr></table>
+
+Table 9: Efficiency comparison on LLaVA-1.5-7B. Lower is better for all metrics.
+
+# G Detailed Ablation Study
+
+# G.1 Components Ablation
+
+To assess the contributions of MRFD's core components—attention-guided Region Selection (RS), JSD-based Consistency Weighting (CW), and Fusion Prompts (FP)—we conduct ablation studies. Results on POPE-COCO (average all settings, LLaVA-1.5) are in Table 10. The evaluated model variants are:
+
+(a) MRFD (Full): Our complete method, integrating all three components.  
+(b)  $\mathbf{w} / \mathbf{o}\mathbf{CW}$  : Employs RS and FP, but uses uniform weights for fusing predictions from multiple regions, bypassing JSD-based consistency weighting.  
+(c) w/o FP : Utilizes RS and CW, but omits the enriched Fusion Prompts, using only the original question for each regional decoding pass.  
+(d) w/o RS (Global Image + FP): Bypasses attention-guided region selection, operating solely on the global image. The Fusion Prompt is constructed based on the global image's initial response. In this single-view context, the multi-region Consistency Weighting (CW) mechanism as defined is not applicable or becomes trivial (effectively a weight of 1 for the single view).  
+(e) RS only: Leverages attention-guided regions, but with uniform fusion weights and only the original question as prompt, isolating the benefit of the multi-region perspective itself.  
+(f) FP only (Global Image): Applies the Fusion Prompt (derived from the global image's initial response) directly to the global image decoding, without multi-region selection or any form of consistency weighting.  
+(g) Global Image (Token-Level Voting): Operates on the global image with the original prompt. The CW component is adapted to perform token-level voting by aggregating implicitly diversified predictions from the single global view, testing self-consistency benefits at the token level.  
+(h) Regular: Standard greedy decoding using only the global image and the original question.
+
+<table><tr><td></td><td colspan="3">Components</td><td>Acc. ↑</td><td>Prec. ↑</td><td>F1 ↑</td></tr><tr><td></td><td>CW</td><td>FP</td><td>RS</td><td></td><td></td><td></td></tr><tr><td>(a)</td><td>✓</td><td>✓</td><td>✓</td><td>86.50</td><td>88.11</td><td>86.21</td></tr><tr><td>(b)</td><td></td><td>✓</td><td>✓</td><td>83.76</td><td>84.27</td><td>83.74</td></tr><tr><td>(c)</td><td>✓</td><td></td><td>✓</td><td>82.87</td><td>83.95</td><td>82.70</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td></td><td>82.77</td><td>83.94</td><td>82.58</td></tr><tr><td>(e)</td><td></td><td></td><td>✓</td><td>82.58</td><td>82.01</td><td>82.26</td></tr><tr><td>(f)</td><td></td><td>✓</td><td></td><td>81.96</td><td>81.36</td><td>82.14</td></tr><tr><td>(g)</td><td>✓</td><td></td><td></td><td>81.52</td><td>81.05</td><td>81.68</td></tr><tr><td>(h)</td><td></td><td></td><td></td><td>80.57</td><td>78.84</td><td>81.59</td></tr></table>
+
+Table 10: Ablation study with different model variants on POPE-COCO under the average of three settings. CW: JSD-based Consistency Weighting, FP: Fusion Prompt, RS: Region Selection.
+
+# G.2 Decoding Strategies
+
+To evaluate the impact of different decoding strategies on MRFD, we conduct an ablation study using the POPE-COCO benchmark under the adversarial setting with LLaVA-1.5. The results are summarized in Table 11. We compare three decoding strategies: low temperature sampling, high temperature sampling, and high temperature sampling with top-p filtering. The best results are bolded.
+
+The low temperature strategy is the default setting in our experiments, which  $t$  is 0.1. The high temperature strategy is set to  $t = 0.5$ , which increases the randomness of the sampling process. The high temperature + top-p strategy combines high temperature sampling with top-p filtering, where we set  $p = 0.7$  to retain the top  $70\%$  of the probability mass.
+
+<table><tr><td>Decoding</td><td>Acc. ↑</td><td>Prec. ↑</td><td>F1 ↑</td></tr><tr><td>Low temp.</td><td>82.75</td><td>85.22</td><td>82.12</td></tr><tr><td>High Temp.</td><td>83.04</td><td>82.29</td><td>83.23</td></tr><tr><td>High Temp.+Top P</td><td>82.50</td><td>81.25</td><td>82.67</td></tr></table>
+
+Table 11: Ablation study on different decoding strategies.
+
+# H Qualitative Examples
+
+To provide a more intuitive understanding of MRFD's impact on mitigating hallucinations and improving caption factuality, this section presents qualitative examples from our experiments on the CHAIR benchmark using LlaVA-1.5. We compare captions generated by our Multi-Region Fusion
+
+Decoding (MRFD) method against those from the standard Regular baseline. These examples are selected to highlight differences in how the methods perceive and describe objects, attributes, and their relationships within the visual scene. Figure 8 and Figure 9 illustrate one such comparison.
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Setting</td><td rowspan="2">Method</td><td colspan="4">LLaVA-1.5</td><td colspan="4">InstructBLIP</td></tr><tr><td>Acc. ↑</td><td>Prec. ↑</td><td>Recall ↑</td><td>F1 ↑</td><td>Acc. ↑</td><td>Prec. ↑</td><td>Recall ↑</td><td>F1 ↑</td></tr><tr><td rowspan="24">MS-COCO</td><td rowspan="6">Random</td><td>Regular</td><td>83.13</td><td>81.94</td><td>85.00</td><td>83.44</td><td>83.07</td><td>83.02</td><td>83.26</td><td>83.08</td></tr><tr><td>VCD</td><td>87.00</td><td>86.13</td><td>88.18</td><td>87.15</td><td>86.23</td><td>88.14</td><td>83.73</td><td>85.88</td></tr><tr><td>M3ID</td><td>87.50</td><td>87.38</td><td>87.67</td><td>87.52</td><td>86.67</td><td>88.09</td><td>84.79</td><td>86.41</td></tr><tr><td>RITUAL</td><td>88.87</td><td>89.23</td><td>88.39</td><td>88.81</td><td>88.83</td><td>90.48</td><td>86.80</td><td>88.60</td></tr><tr><td>DeGF</td><td>89.03</td><td>91.20</td><td>86.41</td><td>88.74</td><td>88.83</td><td>93.73</td><td>82.42</td><td>87.71</td></tr><tr><td>Ours</td><td>89.50</td><td>92.55</td><td>85.94</td><td>89.12</td><td>88.52</td><td>93.83</td><td>82.61</td><td>87.86</td></tr><tr><td rowspan="6">Popular</td><td>Regular</td><td>81.17</td><td>78.28</td><td>86.26</td><td>82.08</td><td>77.00</td><td>73.82</td><td>83.68</td><td>78.44</td></tr><tr><td>VCD</td><td>83.10</td><td>79.96</td><td>88.34</td><td>83.94</td><td>80.07</td><td>77.67</td><td>84.39</td><td>80.89</td></tr><tr><td>M3ID</td><td>84.30</td><td>81.58</td><td>88.63</td><td>84.95</td><td>80.97</td><td>77.93</td><td>86.19</td><td>81.85</td></tr><tr><td>RITUAL</td><td>85.83</td><td>84.17</td><td>88.26</td><td>86.17</td><td>81.97</td><td>78.90</td><td>87.26</td><td>82.87</td></tr><tr><td>DeGF</td><td>86.63</td><td>87.75</td><td>84.86</td><td>86.28</td><td>82.73</td><td>84.02</td><td>80.27</td><td>82.10</td></tr><tr><td>Ours</td><td>87.24</td><td>86.56</td><td>88.22</td><td>87.38</td><td>83.69</td><td>85.22</td><td>81.58</td><td>83.36</td></tr><tr><td rowspan="6">Adversarial</td><td>Regular</td><td>77.43</td><td>73.31</td><td>86.26</td><td>79.26</td><td>74.60</td><td>71.26</td><td>82.46</td><td>76.45</td></tr><tr><td>VCD</td><td>77.17</td><td>72.18</td><td>88.40</td><td>79.47</td><td>77.20</td><td>74.29</td><td>83.19</td><td>78.49</td></tr><tr><td>M3ID</td><td>78.23</td><td>73.51</td><td>88.28</td><td>80.22</td><td>77.47</td><td>73.68</td><td>85.48</td><td>79.14</td></tr><tr><td>RITUAL</td><td>78.80</td><td>74.43</td><td>87.74</td><td>80.54</td><td>78.73</td><td>74.57</td><td>87.21</td><td>80.39</td></tr><tr><td>DeGF</td><td>81.63</td><td>80.59</td><td>83.33</td><td>81.94</td><td>80.30</td><td>80.90</td><td>79.33</td><td>80.11</td></tr><tr><td>Ours</td><td>82.75</td><td>85.22</td><td>79.25</td><td>82.12</td><td>82.49</td><td>83.14</td><td>81.51</td><td>82.32</td></tr><tr><td rowspan="6">Random</td><td>Regular</td><td>81.90</td><td>76.63</td><td>91.78</td><td>83.53</td><td>80.63</td><td>76.82</td><td>87.75</td><td>81.92</td></tr><tr><td>VCD</td><td>83.83</td><td>78.05</td><td>94.13</td><td>85.34</td><td>84.20</td><td>80.90</td><td>89.54</td><td>85.00</td></tr><tr><td>M3ID</td><td>84.67</td><td>79.25</td><td>93.94</td><td>85.97</td><td>85.43</td><td>81.77</td><td>91.20</td><td>86.23</td></tr><tr><td>RITUAL</td><td>85.17</td><td>79.79</td><td>94.21</td><td>86.40</td><td>87.13</td><td>83.92</td><td>91.87</td><td>87.71</td></tr><tr><td>DeGF</td><td>86.93</td><td>84.28</td><td>90.80</td><td>87.42</td><td>87.40</td><td>87.67</td><td>86.86</td><td>87.26</td></tr><tr><td>Ours</td><td>87.13</td><td>87.56</td><td>86.76</td><td>87.16</td><td>88.33</td><td>86.81</td><td>90.40</td><td>88.57</td></tr><tr><td rowspan="18">A-OKVQA</td><td rowspan="6">Popular</td><td>Regular</td><td>75.07</td><td>68.58</td><td>92.53</td><td>78.77</td><td>75.17</td><td>70.15</td><td>87.60</td><td>77.91</td></tr><tr><td>VCD</td><td>76.63</td><td>69.59</td><td>94.59</td><td>80.19</td><td>78.63</td><td>73.53</td><td>89.46</td><td>80.72</td></tr><tr><td>M3ID</td><td>77.80</td><td>70.98</td><td>94.07</td><td>80.91</td><td>78.80</td><td>73.38</td><td>90.39</td><td>81.00</td></tr><tr><td>RITUAL</td><td>78.83</td><td>71.99</td><td>94.37</td><td>81.68</td><td>78.73</td><td>72.83</td><td>91.68</td><td>81.17</td></tr><tr><td>DeGF</td><td>80.90</td><td>75.68</td><td>91.05</td><td>82.66</td><td>81.47</td><td>78.61</td><td>86.47</td><td>82.35</td></tr><tr><td>Ours</td><td>80.99</td><td>76.68</td><td>89.28</td><td>82.51</td><td>83.19</td><td>78.74</td><td>91.00</td><td>84.43</td></tr><tr><td rowspan="6">Adversarial</td><td>Regular</td><td>67.23</td><td>61.56</td><td>91.81</td><td>73.70</td><td>69.87</td><td>64.54</td><td>88.20</td><td>74.54</td></tr><tr><td>VCD</td><td>67.40</td><td>61.39</td><td>93.79</td><td>74.21</td><td>71.00</td><td>65.41</td><td>89.13</td><td>75.45</td></tr><tr><td>M3ID</td><td>68.60</td><td>62.22</td><td>94.74</td><td>75.11</td><td>70.10</td><td>64.28</td><td>90.47</td><td>75.16</td></tr><tr><td>RITUAL</td><td>68.57</td><td>62.26</td><td>94.27</td><td>74.99</td><td>70.27</td><td>64.15</td><td>91.89</td><td>75.55</td></tr><tr><td>DeGF</td><td>72.70</td><td>66.70</td><td>90.68</td><td>76.86</td><td>73.93</td><td>69.36</td><td>85.70</td><td>76.67</td></tr><tr><td>Ours</td><td>75.23</td><td>71.03</td><td>85.39</td><td>77.65</td><td>75.62</td><td>69.40</td><td>91.78</td><td>79.04</td></tr><tr><td rowspan="6">Random</td><td>Regular</td><td>82.23</td><td>76.32</td><td>93.47</td><td>84.03</td><td>79.67</td><td>76.05</td><td>86.62</td><td>80.99</td></tr><tr><td>VCD</td><td>83.23</td><td>76.73</td><td>95.38</td><td>85.05</td><td>82.83</td><td>80.16</td><td>87.26</td><td>83.56</td></tr><tr><td>M3ID</td><td>84.20</td><td>78.00</td><td>95.26</td><td>85.77</td><td>83.07</td><td>80.06</td><td>88.06</td><td>83.87</td></tr><tr><td>RITUAL</td><td>86.10</td><td>80.30</td><td>95.66</td><td>87.31</td><td>84.87</td><td>82.52</td><td>88.47</td><td>85.39</td></tr><tr><td>DeGF</td><td>87.40</td><td>83.51</td><td>93.20</td><td>88.09</td><td>85.40</td><td>85.64</td><td>84.61</td><td>85.12</td></tr><tr><td>Ours</td><td>87.81</td><td>86.62</td><td>90.32</td><td>88.41</td><td>87.24</td><td>85.57</td><td>89.69</td><td>87.58</td></tr><tr><td rowspan="12">GQA</td><td rowspan="6">Popular</td><td>Regular</td><td>73.47</td><td>66.83</td><td>93.20</td><td>77.84</td><td>73.33</td><td>68.72</td><td>85.67</td><td>76.26</td></tr><tr><td>VCD</td><td>72.37</td><td>65.27</td><td>95.60</td><td>77.58</td><td>76.13</td><td>71.10</td><td>88.07</td><td>78.68</td></tr><tr><td>M3ID</td><td>73.87</td><td>66.70</td><td>95.35</td><td>78.49</td><td>75.17</td><td>69.94</td><td>88.26</td><td>78.04</td></tr><tr><td>RITUAL</td><td>74.80</td><td>67.50</td><td>95.66</td><td>79.15</td><td>74.50</td><td>69.17</td><td>88.39</td><td>77.61</td></tr><tr><td>DeGF</td><td>78.10</td><td>71.56</td><td>93.25</td><td>80.98</td><td>76.90</td><td>73.89</td><td>83.20</td><td>78.27</td></tr><tr><td>Ours</td><td>77.72</td><td>73.17</td><td>88.24</td><td>80.00</td><td>79.62</td><td>74.61</td><td>90.15</td><td>81.65</td></tr><tr><td rowspan="6">Adversarial</td><td>Regular</td><td>68.60</td><td>62.43</td><td>93.41</td><td>74.84</td><td>68.60</td><td>63.94</td><td>85.31</td><td>73.10</td></tr><tr><td>VCD</td><td>68.83</td><td>62.26</td><td>95.67</td><td>75.43</td><td>71.00</td><td>65.75</td><td>87.66</td><td>75.14</td></tr><tr><td>M3ID</td><td>68.67</td><td>62.16</td><td>95.42</td><td>75.28</td><td>71.17</td><td>65.79</td><td>88.19</td><td>75.36</td></tr><tr><td>RITUAL</td><td>68.23</td><td>61.75</td><td>95.81</td><td>75.10</td><td>70.17</td><td>64.76</td><td>88.48</td><td>74.78</td></tr><tr><td>DeGF</td><td>74.07</td><td>67.42</td><td>93.14</td><td>78.22</td><td>73.63</td><td>70.08</td><td>80.92</td><td>75.11</td></tr><tr><td>Ours</td><td>76.00</td><td>72.22</td><td>84.50</td><td>77.88</td><td>75.06</td><td>70.25</td><td>87.18</td><td>77.80</td></tr></table>
+
+Table 6: Detailed results on POPE (Li et al., 2023d) benchmark. Higher  $(\uparrow)$  accuracy, precision, recall, and F1 indicate better performance. The best results are bolded, and the second-best are underlined.
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Method</td><td colspan="2">Object-level</td><td colspan="2">Attribute-level</td><td rowspan="2">Total Score↑</td></tr><tr><td>Existence ↑</td><td>Count ↑</td><td>Position ↑</td><td>Color ↑</td></tr><tr><td rowspan="6">LLaVA-1.5</td><td>Regular</td><td>173.75</td><td>121.67</td><td>117.92</td><td>149.17</td><td>562.50</td></tr><tr><td>VCD</td><td>186.67</td><td>125.56</td><td>128.89</td><td>139.45</td><td>580.56</td></tr><tr><td>M3ID</td><td>186.67</td><td>128.33</td><td>131.67</td><td>151.67</td><td>598.11</td></tr><tr><td>RITUAL</td><td>187.50</td><td>139.58</td><td>125.00</td><td>164.17</td><td>616.25</td></tr><tr><td>DeGF</td><td>186.22</td><td>148.33</td><td>130.50</td><td>170.33</td><td>635.38</td></tr><tr><td>Ours</td><td>190.00</td><td>148.33</td><td>128.33</td><td>173.33</td><td>640.00</td></tr><tr><td rowspan="6">InstructBLIP</td><td>Regular</td><td>155.42</td><td>72.17</td><td>72.58</td><td>115.43</td><td>415.60</td></tr><tr><td>VCD</td><td>154.49</td><td>81.67</td><td>59.11</td><td>112.67</td><td>407.94</td></tr><tr><td>M3ID</td><td>158.89</td><td>77.22</td><td>61.44</td><td>125.00</td><td>422.55</td></tr><tr><td>RITUAL</td><td>178.50</td><td>74.58</td><td>59.08</td><td>139.17</td><td>451.33</td></tr><tr><td>DeGF</td><td>184.32</td><td>82.44</td><td>56.33</td><td>150.00</td><td>473.09</td></tr><tr><td>Ours</td><td>188.67</td><td>89.58</td><td>62.33</td><td>152.33</td><td>492.91</td></tr></table>
+
+Table 7: Detailed performance on the MME-Hallucination benchmark. Scores are reported as mean. Higher scores  $\uparrow$  indicate better performance. For each model group LLaVA-1.5, InstructBLIP, results for "Ours" are bolded if they are the best in that column. Other best results in a column are also bolded. Underlined values score typically represent the second-best performing method or a notable baseline. The "Ours" rows are highlighted.
+
+![](images/21a05052314abe76dfdd7cea01e07f7ec6ec55949df75bcb201a608c7e212103.jpg)
+
+![](images/57d9addc0f27b81b47b917fb91ccf604b735c35d4545518363c643cf46a8d212.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/35e2ed97c2129b5beb5d138aef8e3a3d4da0c797296eafd948f55a84c694ddd4.jpg)
+
+Answer: A bunch of green lemons.
+
+![](images/765dff41e95ebe6265fbc95ac9cc494183c21ec34231b08b2d24ea32e2536316.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/a975f71b3c146ac393e691fddaee45ff2c382a5e013c2b7b58c4b60c8bcd4ad8.jpg)
+
+Answer: A bunch of green and yellow lemons.
+
+![](images/4440f23ded19cbaf7663ef8f98ed1ccdd276fd6f80115923ade17b95ddb4fbbe.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/172c0f74c8317528020458360ffeb47fd0812fee4d66dee8b62a4542c9846947.jpg)
+
+Answer: A bunch of green and yellow lemons.
+
+![](images/14d9e98bd79eced096673f6500ee3cba7e88d8a65fb727ccc7ec084334c0886b.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/a084610058250168f53861fd8f12fcd114ef1bf0fc2bb09057e8d96aa712ca88.jpg)
+
+Answer: A bunch of green and yellow lemons.
+
+![](images/66c2282880c02bb5c6b7b1b1c9f9192ba91c12052a30b40cad85717906af1585.jpg)
+
+Question: Please describe the image in details.
+
+![](images/f9b8b6047739d3f01b2821b10efdd7e03dfdce6bebab74e2328b531ea3f49d04.jpg)
+
+Answer: The image features a large assortment of oranges, with a prominent display prominently displayed in an accessible location. Surrounding the crate of fresh oranges, there are various other boxes containing even more lemons of different sizes and colors.
+
+![](images/6fe63f6622830136476d78e87c0acf6a8fb1b495b41fe7cdf9ecc76508250fde.jpg)
+
+![](images/e89ea7a513ebeefa053b6b3bca7ac6ba8cb0862bcd9463d741a4604cb97c18a0.jpg)
+
+Question: Please describe the image in details.
+
+![](images/5d082bcaf3734d16b93325ed85540e678e2e4af1707ce31ac88cf9991c080140.jpg)
+
+Answer: The image features a large pile of lemons, with many of them being green and yellow. The lemons are arranged in various positions, some stacked on top of each other and others sitting next to each other. The pile is quite large, occupying a significant portion of the image. The lemons are displayed in a way that showcases their vibrant colors and abundance.
+
+![](images/9d732b1204b7ab6c6780101bd863f71c42cc7716e02f12ebbc306715c528cfa7.jpg)
+
+![](images/c454b03b816598adf3eb71195a59e5437f34b000e77eaa997f3f59a0f651b56e.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/8b92f89dbc017a192e56755d91746bad0add17eb6f1f5c84c2bf8f6083c40fba.jpg)
+
+Answer: The elephant is in the water.
+
+![](images/a5821441fbc1a7705e167474de8f33db9d9bef0775f786d0ec902a2dd4ae005e.jpg)
+
+![](images/deb9e240b00405b01e3ee7aabe415dd520838d66760b55b26471cf999a2c8d93.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/c1425ba9f5f8e9d1504c3ea92036e251440c7572af32ff41b95becd6d63f35f7.jpg)
+
+Answer: The elephant is in the water.
+
+![](images/14382dfdfd7ccb9678f877149012564ac3505135c468915f4ed62504e9af5bdb.jpg)
+
+Question: The elephant is in the water.
+
+![](images/3d26710cd0d6a6cc172bc06282ea8041b4e5e0baad52b4ab37c6440e3f123a10.jpg)  
+Figure 8: Qualitative comparison of captions generated by Regular decoding (Red) and our MRFD (Green) on an image from the CHAIR evaluation set.
+
+Answer: The elephant is in the water.
+
+![](images/b5e5fadb842caf5b784a022692c8713f53d101e8ba0756fee6bcaf292303c536.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/9b3338c1f4defd6fa3c3cea8860b8d85fdcfe807bad1e45b7a3d32a57b737f91.jpg)
+
+Answer: The elephant is in the water.
+
+![](images/1621da1c44b0977b02bdf6138eef7ba48ed9ef61ca0e803d6cfaf0a0a294abfe.jpg)
+
+Question: Please describe the image in details.
+
+![](images/19ea9b3b074611001c2056acf7f4221374cdb12e674d4bd7006258719d7310dd.jpg)
+
+Answer: The image showcases two elephants splashing around and playing in a lake together. The first elephant, on the left side of the scene, is partially submerged, displaying its trunk out of the water as it interacts with the other elephant. The second elephant, located on the right side, is in full view and appears to be playfully prodding the first elephant.
+
+![](images/5a00e4989af40489d7aa33a7f07fc03082a978bd2dafe43dc8fd04b51fc60d45.jpg)
+
+Question: Please describe the image in details.
+
+![](images/7d39e957990d3102aa87b94e3a1c9b280fae61de238db33d9a6a911fb820e641.jpg)
+
+Answer: The image features two elephants in a body of water, possibly a lake or a river. One elephant is larger and appears to be an adult, while the other is smaller and likely a baby elephant. The baby elephant is resting its body on the larger elephant, creating a heartwarming scene.
+
+![](images/c7e037833216d441e52ca308fef84f5d9f6de4ea87c037afd03fba48da01865d.jpg)
+
+![](images/327bf79d04d16e9cd54728848bc1b51390adc543506c0cbd64bdaabe209c71d2.jpg)
+
+![](images/c66536c92ba608421629399d85fc788e024e3c60535122fe44a053e53fd6f2ed.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/e8cbb871eb847d904201db39c9e26b1451dab701a97fcae44760ec4d8f2e51f8.jpg)
+
+Answer: A piece of chicken in a sandwich.
+
+![](images/ab2fb58a77eaa194027b1dd36d8bd4ae9cd8193cd5d1f9bbe44533b0f3a2ffa9.jpg)
+
+![](images/d1897f01622735a8b1c427b8b088a7a2a183b17483a71bd0fd3f1372793e983a.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/320786fc2d22604465284258e26cfc3c61faae58986138eb222bdd316aba9433.jpg)
+
+Answer: A side of french fires.
+
+![](images/9e5d15e1a3ff0c76a2ed32aeb8182cb0231609922a63aeefa0ec099ce63a7eab.jpg)
+
+![](images/6c1e666e3000f1612f232f114e24f5016d4c6d5fc95bf295da9e251999689002.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/8a31c42d7599ddb7cae5a83505436e7df3bba28ef3421c2953e1a582fa580252.jpg)
+
+Answer: Two pieces of sandwich.
+
+![](images/c35579d08816a4e70beca4812ac2583af465d3c7988ea1b1e02675e000172421.jpg)
+
+![](images/653e15568c4c5d0bbaf023299ebfa2a096e985fb29aff520ecb4be7d04d1012d.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/f671a1d1baadd60dbe707724745dc75021f9d6f8fb79bc2a88df4a66cd17b6c0.jpg)
+
+Answer: Sandwich with two forks.
+
+![](images/2678f55229274fd22063508aa97c98ab3f4a7ab8cd249d633272c99276e37376.jpg)
+
+Question: Please describe the image in details.
+
+![](images/4a45a3dc9f64d70ac88083fed912fcd1ef50e5caf0bbedb677b403aba09b2ac3.jpg)
+
+Answer: The image shows a wooden dining table where a delicious combination of food and beverages is served. There are various sandwiches on a plate and a tray of chips, accompanied by cups containing drinks, possibly beers since two beer forks are present. In addition to the main dishes, the table is also set with three forks, two of which are positioned in similar directions and near the sandwiches, while three knives are placed around the plate and tray.
+
+![](images/cbea1c6f05c594899ef39b5b5b3f7dffbdc5836c602775d98b6281b327cfc946.jpg)
+
+![](images/90d88b9f2b611e809169dc3e84c5beff6f1e4b43dad318048f834c630849dc17.jpg)
+
+Question: Please describe the image in details.
+
+![](images/6ed1abdf4739401c59624dba6ab676e27a3a54bd56381ce1b411d9ceb629f7f7.jpg)
+
+Answer: The image features a delicious meal consisting of four sandwich and a side of french fries. The sandwich is cut in half, revealing the meat and vegetables inside. The fries are scattered around the sandwich, with some pieces placed near the sandwich and others further away. In addition to the main dishes, the table is also set with four forks.
+
+![](images/b98b0b6610c74d650361dde97db96e12302e8c65336809d818dedabc42a6454a.jpg)
+
+![](images/bebefba5d17aaaae630798170cb3ea310bb590a361a00efdd7b77323c672de5c.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/ef4a5b3d1b9ce054b75f508fe162ad4e53c2815337a12f6dba364d96bf8a632d.jpg)
+
+Answer: A toilet in a bathroom.
+
+![](images/108a4d95481da96b7e55335bc80d0182d5e0402270851fe101d86452a50ae78b.jpg)
+
+![](images/8e24abb999f0392b7d75af0171ea7228605be32c6bb959af903f477e188a497b.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/91d32a90ffff984daca2dd7e85feac19078268033902d13ff25a9c4641bdedad.jpg)
+
+Answer: The toilet seat is white.
+
+![](images/094077f6ad988af04d51c4bb427b70b1b9f2a24d1d7cfbbe468df0f616473b48.jpg)
+
+![](images/5c05bb67242dbab809f1e5c5d6fe676858bf13457d4fcda055e87ab0d09cbc25.jpg)
+
+Question: Provide a short description of the image.
+
+![](images/4123076169ae414efcd11ad88a64710a0e5f58529f784afc127190379d38b5e9.jpg)
+
+Answer: A silver handle on the wall.
+
+![](images/d86904919052d224df2db8b7806e6cf9c3376f5652198b784313994c4816274d.jpg)  
+Figure 9: Qualitative comparison of captions generated by Regular decoding (Red) and our MRFD (Green) on an image from the CHAIR evaluation set.
+
+Question: Provide a short description of the image.
+
+![](images/557429d1b49d5f4da1826f965cf641d2fddbfc4072ea18f9496b399b65c1b1a3.jpg)
+
+Answer: The toilet seat is white.
+
+![](images/890af807db8bb6a93bb4444163f0bdc025b0d4ca5dd16c70964198089eb61a50.jpg)
+
+Question: Please describe the image in details.
+
+![](images/2c0b71bd22461d65014cbc5cf9a0954c230f0c1e8047d2b5f6390710063daedd.jpg)
+
+Answer: The scene is a public bathroom with a toilet mounted on the wall. A white urinal is attached below a podium, featuring an electronic device likely for display purposes. There is another toilet in the room, situated on the left side with the blanket displaying a picture on it, likely giving the toilet a unique appearance.
+
+![](images/ed7aa72143a84cdb24336e2ab18d9e979bbfecc93da5f4681e44e7215b2cef88.jpg)
+
+![](images/41fda691f1f78a6594fb13d5a64c7da07eed1423910295034f846ae3116f44dd.jpg)
+
+Question: Please describe the image in details.
+
+![](images/3e9ded32b53fc8b89265301ed48fd5e06c9aeef6fe310169bafe5c54b1fae45d.jpg)
+
+Answer: The image features a white toilet in a bathroom. The toilet is situated next to a wall, and there is a silver handle on the wall. The toilet is positioned in the lower left corner of the image. The bathroom appears to be clean and well-maintained.
+
+![](images/c6435ddfad8e3a45880d476402a9e644be5eaf03db77e6ed6d3be4a8ebd96d1e.jpg)

EMNLP/2025/MRFD_ Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82f2eff19b2784239e0f846dcea4b2bc0c95d43567c4da095f455914d0343218
+size 1457725

EMNLP/2025/MRFD_ Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeabb532b0e604eb6c316225244fcad8c94ab4604e5ee856ff4567d27f2dc355
+size 747528

EMNLP/2025/MT-Mol_ Multi Agent System with Tool-based Reasoning for Molecular Optimization/58dec34e-a449-4488-b51b-35efc1c0fd7a_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:397f122a4b31293a35a61ac948ab8044d57dc620941c73f347ea03121e0996d9
+size 167949

EMNLP/2025/MT-Mol_ Multi Agent System with Tool-based Reasoning for Molecular Optimization/58dec34e-a449-4488-b51b-35efc1c0fd7a_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a66bfbfcd8edf5c5da7cfb7be99ce71869d01de316d06782caef3ff253547975
+size 199795

EMNLP/2025/MT-Mol_ Multi Agent System with Tool-based Reasoning for Molecular Optimization/58dec34e-a449-4488-b51b-35efc1c0fd7a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d63e03d8919a644d28673643099f60b4d8f3e16b15f98037c5fe221adc52c8fd
+size 6232301

EMNLP/2025/MT-Mol_ Multi Agent System with Tool-based Reasoning for Molecular Optimization/full.md

@@ -0,0 +1,834 @@
+# MT-MOL: Multi Agent System with Tool-based Reasoning for Molecular Optimization
+
+Hyomin Kim
+
+KAIST
+
+hyomin126@kaist.ac.kr
+
+Yunhui Jang
+
+KAIST
+
+yunhuijiang@kaist.ac.kr
+
+Sungsoo Ahn
+
+KAIST
+
+sungsoo.ahn@kaist.ac.kr
+
+# Abstract
+
+Large language models (LLMs) have large potential for molecular optimization, as they can gather external chemistry tools and enable collaborative interactions to iteratively refine molecular candidates. However, this potential remains underexplored, particularly in the context of structured reasoning, interpretability, and comprehensive tool-grounded molecular optimization. To address this gap, we introduce MT-MOL, a multi-agent framework for molecular optimization that leverages tool-guided reasoning and role-specialized LLM agents. Our system incorporates comprehensive RDKit tools, categorized into five distinct domains: structural descriptors, electronic and topological features, fragment-based functional groups, molecular representations, and miscellaneous chemical properties. Each category is managed by an expert analyst agent, responsible for extracting task-relevant tools and enabling interpretable, chemically grounded feedback. MT-MOL produces molecules with tool-aligned and stepwise reasoning through the interaction between the analyst agents, a molecule-generating scientist, a reasoning-output verifier, and a reviewer agent. As a result, we show that our framework shows the state-of-the-art performance of the PMO-1K benchmark on 15 out of 23 tasks and outperforms LLM baselines on ChemCoTbench benchmark.
+
+# 1 Introduction
+
+Large language models (LLMs) have demonstrated remarkable capabilities in a wide range of problems such as question answering (Dong et al., 2024; Sun et al., 2024), summarization (Kim et al., 2024; Liu et al.), translation (Alves et al., 2024; Bari et al., 2025), and code generation (Chen et al., 2021; Li et al., 2023b) using large-scale pretraining and incontext learning (Brown et al., 2020; Chowdhery et al., 2023; Zhang et al., 2022b) (Chen et al., 2021). Motivated by the success, researchers are investigating the potential of LLMs for scientific discovery
+
+in the chemical domain (Li et al., 2025; Wang et al., 2023; Luu et al., 2021; Wang et al., 2025; Nguyen and Grover, 2025; Bran et al., 2024).
+
+In particular, employing LLMs to design new molecules (e.g., drug candidates), is promising due to several advantages: (1) LLMs exhibit general understanding and reasoning capabilities obtained from large-scale pretraining, (2) they can use the off-the-shelf tools for analyzing molecules, and (3) they are capable of interact with other agents to further improve the design candidate.
+
+Recent studies have explored the application of LLMs in molecular optimization. For example, LICO (Nguyen and Grover, 2025) extends LLMs with embedding layers and in-context examples to build a surrogate modeling framework for molecular optimization. MOLLEO (Wang et al., 2025) leverages LLMs as mutation and crossover operators within an evolutionary algorithm. ChemCrow (Bran et al., 2024) integrates LLMs with chemical tools for to facilitate synthesis planning and molecular analysis. While these approaches demonstrate encouraging results, we argue that they do not fully exploit the broader capabilities of modern LLMs such as multi-agent collaboration, tool integration, and iterative reasoning, which are essential for high-quality molecular optimization.
+
+Contribution. In this work, we propose MT-MOL, a multi-agent framework for molecular optimization. Our key idea is to decompose the optimization process into four distinct roles (analyst, scientist, verifier, and reviewer) and employ specialized agent for each role. Unlike previous approaches, MT-MOL generates molecules with explicit stepwise reasoning, consistency checks, and tool-informed feedback. Furthermore, we collect a set of 154 chemistry-related functions, which serve as applicable tools for agents during molecular generation process.
+
+To be specific, we introduce four agents: (1) an-
+
+![](images/1af0cd80c5cd811af349a7ea129d21228c89b7f78f1acc0947a93466f72ccf09.jpg)
+
+# User prompt
+
+Design a drug-like molecule structurally similar to albuterol (SMILES: CC(C)(C)NCC(C1=CC(=C (C=C1)O)CO)O) ...
+
+# 1. Retrieve tools and top-k data
+
+# Tool retrieving
+
+![](images/0bc8af904076fd2ac630a028cb330e6ce3470d8906971b8f95ad18b7f4adae0b.jpg)
+
+user prompt
+
+![](images/1cd5ae33e19f1785c3d034a735a5320d02f285081d8b8538151a1e724220c64e.jpg)
+
+![](images/bf33b9ab23d90f1b3f38ddfaddcf7b20f99c3035336e7f451368c1ddb235bfe0.jpg)
+
+![](images/bfe298970cc34c70f90aeffa9d6f77fd4aea94f278d0b6ecda3f5a571c7c1a9f.jpg)
+
+![](images/c21ebb114e217b2119b3188c862066ae4dd865717e7e84a629b16da97cd9a88a.jpg)
+
+![](images/fba4727d6a65b9ef443d8a7409bc25e3530eb21cdf7143abd1fc8f5a95ae8091.jpg)
+
+5 categorized tool files
+
+![](images/707a349cfb47dd72136b7473c9119c75c2917ffa6e8e86271ba4f0a89050f3ae.jpg)
+
+![](images/bcc2de59997e58993955fa51832433896246cdcb591fb82d6d1bafb87d04cedc.jpg)
+
+![](images/e514778c73b2b7122c1081b0a08219fd864311e5b5d05304ba53c145a04ebc62.jpg)
+
+![](images/cf4c2538217c4f3215fdfaff4c5c226456b50fe4ecce7f4527ccf9b324c784dd.jpg)
+
+![](images/42007fddfad11cf7b7dd82aeaa5ca6393cffda1b849548b63626b36c15cb5f13.jpg)
+
+![](images/1326494fb7bd8c1fb5c4e743cd1add124cb5388fb92476e7793b3a3095ed5dd7.jpg)
+
+5 expert tool agents
+
+![](images/a285b05ea87a02891c5a69b064b934241240fa4933db0efc40860047cc94d3fe.jpg)
+
+![](images/128688f3330b72f657c78ce0de42609b45838e844fd8ac70eae29a3c2d2a2594.jpg)
+
+list of retrieved toolsby each agent
+
+# Top-k data retrieving
+
+![](images/dc50a9c3272175c2908e46982015bac1a638cc8b327ab370ea07a7c25f12d636.jpg)  
+dataset
+
+![](images/4108e8e679e47280313355e11d00e31f2b9ea5fdba3a80420e8c8d4074a09d16.jpg)
+
+![](images/b055854ae34ed3bf17ec58225718368e7a8c6c718b432eead6e3896fe093ef70.jpg)  
+sort by score
+
+![](images/ae5b5abdb0b00417f8c7bb9d3710b97ae529e90d6f2646a89784cbad32f05e39.jpg)
+
+![](images/585852f9ced79cca87c2b329e9bba3406c25095ea028a2272aa41753a145642c.jpg)  
+top-k data
+
+# 2. Generate SMILES with consistency
+
+![](images/66ba20b3f99da5d6902447f10361cfa2144d8554b4a9d3fcdddd514d7c4fbea3.jpg)
+
+![](images/44b3f1f7807ad7cb8ffea464f1bf528e527977b891ecfcb19498f5c5c7c8c52d.jpg)  
+top-k data
+
+![](images/2ba035ed16e163025b3cd02ea6a20d5c552498cdcfcd35c48acb2898bbbd585e.jpg)  
+tool sets
+
+![](images/38b90ece251b0c2485c77aa824cfbfa19977b3d92e004a3d0b7bf704045da51b.jpg)  
+SMILES history
+
+![](images/e9a32db02784cd21a23780f9f431be0f02659f55ad066dd3875e2e987d3022b6.jpg)  
+user prompt
+
+![](images/597b2ed8634bbd329e5994d58c24b37ad9dee0f70023d3bb720ea217e3bd04ba.jpg)  
+verifier
+
+![](images/ef5f7d8edab13de017687ed63f1f45e87048cba0890fc34e446fc79aa5d0c3a4.jpg)  
+finalized SMILES and reasoning
+
+# 3. Review final SMILES and reasoning
+
+![](images/38d233794c3a1ec6bd7952d9514f0e99f61da1ab933f18fdfc11fe9bd8665fdb.jpg)
+
+![](images/2f895652371093132d701ea8b31f037e0880d144ee80fd02997bb329229e1af0.jpg)  
+finalized SMILES and reasonings
+
+![](images/345edbe798c0922b4e18b83150e7d6d126347c19f1cb93eb937c3e0d42c09206.jpg)  
+feedback based on tool analysis of finalized SMILES  
+reviewer  
+Figure 1: Overview of our method. Given a molecular optimization task, analyst agents analyze the prompt and outputs list of relevant RDKit functions from five categories. Top-  $k$  molecules are retrieved as reference molecules for the scientist agent. Then, the scientist agent proposes a SMILES with stepwise reasoning, which the double checker validates for consistency. The reviewer finally assesses the reasoning using tool-informed descriptors and provides structured feedback. This generation and review process is repeated until the maximum number of iterations  $N$  is reached. This multi-agent pipeline enables interpretable, tool-guided molecule generation with iterative refinement toward the design objective.
+
+alyst, (2) scientist, (3) verifier, and (4) reviewer. In detail, five analyst agents proposes the task-specific relevant tools using different types of chemical functions: structure, electronic properties, functional groups, identifiers, and miscellaneous descriptors. Then a scientist agent proposes new molecules in SMILES format (Weininger, 1988) and explains each design step through structured reasoning. Next, a verifier agent evaluates whether the reasoning of the scientist is consistent with the proposed molecule. Finally, a reviewer agent assesses both the molecule and the reasoning process using the outputs from the tools and provides detailed feedback. Each agent plays a collaborative role that enables interpretable, tool-aware, and iterative molecular design. By incorporating domain-specific tools such as RDKit (Landrum, 2013), MTMOL supports chemically informed generation and transparent decision-making.
+
+In summary, we propose a multi-agent framework for molecular optimization, where each agent is assigned a specific role such as tool selection,
+
+molecule generation, consistency validation, and reasoning critique. Our system integrates 154 RDKit functions, organized into five specialized analyst agents covering structural descriptors, electronic and topological descriptors, structural descriptors, fragment-based analysis, and identifiers or representations. We achieve state-of-the-art performance on 15 out of 23 tasks from the PMO-1K benchmark, outperforming recent strong baselines including LICO and MOLLEO in terms of top-10 AUC scores. Additionally, our framework offers an interpretable reasoning pipeline in which each generated molecule is equipped with stepwise rationale, double-check verification, and tool-informed reviewer feedback.
+
+# 2 Related Work
+
+# Generative models for molecular optimization.
+
+Molecular optimization aims to design molecules that maximize desired chemical or biological properties, such as solubility, binding affinity, or synthe
+
+sizability. Generative modeling has emerged as a central approach for this task, encompassing techniques from deep learning to probabilistic search. REINVENT (Olivecrona et al., 2017) introduced reinforcement learning over SMILES strings to fine-tune molecular generation toward desired properties. Jensen (2019) showed that graph-based genetic algorithms and non-ML models combined with Monte Carlo Tree Search perform competitively in optimizing molecular properties under synthetic constraints. Augmented Memory (Guo and Schwaller, 2024) enhances sample efficiency in reinforcement learning through SMILES augmentation and experience replay. Genetic GFN (Bengio et al., 2023) enables compositional molecule generation by sampling in proportion to a reward function, offering diversity and high-reward sampling in molecular benchmarks. Srinivas et al. (2010) introduced GP BO, a Gaussian process-based optimization framework that provides sublinear regret bounds and sample-efficient exploration using information gain from kernel-based uncertainty modeling. While these models improve sample efficiency and diversity, they often lack interpretability and fail to fully utilize the available domain knowledge, such as chemical priors. Our framework complements these approaches by incorporating structured reasoning and chemical tools into the molecular generation process.
+
+LLMs for molecular optimization. LLMs have recently been applied to molecular optimization tasks. LICO (Nguyen and Grover, 2025) extends a pretrained LLM with structured embeddings to model property functions without relying on natural language prompts. MOLLEO (Wang et al., 2025) uses LLMs as evolutionary operators, enabling coherent molecule generation across single- and multi-objective settings. Prompt-MolOpt (Wu et al., 2024) introduces prompt-based editing to optimize multiple properties in low-data regimes while preserving pharmacophores. DrugAssist (Ye et al., 2025) fine-tunes an instruction-based LLM on a curated chemistry dataset to support interactive, feedback-driven molecule design. ChemCrow (Bran et al., 2024) combines general-purpose LLMs with chemistry tools and a ReAct-based reasoning loop to automate generation, retrosynthesis, and property prediction. ChemCoTbench (Li et al., 2025), a benchmark that evaluates step-by-step chemical reasoning in tasks like molecule optimization, reveals that current LLMs still have significant
+
+room for improvement in complex chemical reasoning. Despite these advances, existing approaches often lack interpretability, structured collaboration among specialized agents, and a systematic feedback loop that enhances accurate molecule design. To address these limitations, our method introduces five expert analyst agents powered by RDKit (Landrum, 2013) and a multi-agent feedback loop that ensures both accurate and interpretable molecular optimization.
+
+Multi-agent LLMs. Multi-agent LLMs have shown promise in collaborative reasoning and decomposed problem-solving. AgentVerse (Chen et al., 2023) assigns agents to roles like recruitment and evaluation, leveraging specialization for better coordination. ProAgent (Zhang et al., 2024) enables agents to infer and adapt to teammates' strategies through communication history. Self-Adaptive Multi-agent Systems (Nascimento et al., 2023) use a self-control loop to make agents responsive to dynamic environments. Theory of Mind for Multi-Agent Collaboration (Li et al., 2023a) enhances coordination by giving agents shared belief states and goal-tracking abilities. MetaGPT (Hong et al., 2024) improves communication scalability via a Shared Message Pool that standardizes agent interactions. While these frameworks contribute to multi-agent architectural design, they have overlooked domain-specific tool integration and have less focused on molecule optimization. Our framework addresses this gap by tightly coupling expert analyst agents with reasoning roles to enable targeted, tool-informed molecular design.
+
+# 3 Method
+
+In this section, we introduce our multi-agent framework for molecular optimization, coined MT-MOL. In Section 3.1, we first describe the overview of our system, which consists of four primary agent types: 1) analyst, 2) scientist, 3) verifier, and 4) reviewer agents. We describe details of the analysts in Section 3.2, and stepwise reasoning and feedback process in Section 3.3.
+
+# 3.1 Overall Framework
+
+In this section, we present a high-level overview of our multi-agent framework for molecular optimization. Given a user prompt  $T$ , analyst agents first select relevant tools, then the scientist agent proposes a molecule with structured reasoning. The verifier agent then verifies the logical consistency
+
+![](images/b80d55447cd3372003347a8f12cef7f144ce8014a14d7177a5de9ae1527c1b4f.jpg)
+
+![](images/7d09feaa0f6234892a9c26adb5a05987e317bdf94a7d32c7ce563b9dc98db1c0.jpg)
+
+![](images/b3be96d3e1b17a69172229633f94472b85ca445b2f75940feedf94a89ffa87ae.jpg)
+
+![](images/99c34202258df27be84a52af39ba5bf6b32d1a1dff27ed3a25c9d05fd916a11c.jpg)  
+Figure 2: Example of analyst agents. Example case of five analyst agents analyzing the SMILES proposed by the scientist agent for the fexofenadine_mpo task. Each analyst agent chooses task-relevant tools: electronical and topological descriptors, miscellaneous descriptors, identifiers and representations, structural descriptors, and functional groups. The molecules at the bottom visualizes how analyst agents analyze the scientist agent's proposed SMILES. We provide the description of the tools at Appendix A.
+
+![](images/df3440f115a23e92d3163c44e00acb7fcf035ac23f3f1f74a61b353e74704de3.jpg)
+
+![](images/4c12335e107358af4227e7cc252b2ad97b9515c529fe48b89111167a0a1545fd.jpg)
+
+of the proposed output. Finally, the reviewer agent provides detailed feedback grounded in chemical analysis tools. This feedback guides the generation of an improved molecule and refined reasoning, restarting the cycle, which is repeated for  $N$  iterations. We provide an overview of our method in Figure 1.
+
+Notably, our agents are informed about the details of the objective function and utilize their chemical knowledge to propose better molecules. This is in contrast to existing non-LLM works in molecular design that assume black-box objective functions. We believe that this is a strength of our approach, since in most of the tasks, we have some information about the objective function that can be described in natural language.
+
+Analysts. We design five analyst agents for different aspects of molecular analysis. Each analyst agent parses and analyzes the molecule in the task prompt  $T$  and the scientist agent's proposed SMILES. Each analyst agent wraps a curated set of RDKit or PubChem functions in one of the following categories: 1) structural descriptors, 2) electronic and topological descriptors, 3) fragment-based functional group detectors, 4) chemical identifiers and representations, and 5) miscellaneous
+
+descriptors agents. To analyze a task prompt  $T$ , the analyst agents identify the most relevant chemical features and select the tools accordingly. We illustrate the example case of how tools are used in Figure 2 and provide the details of the tool in Appendix A.
+
+Scientist. The scientist agent generates a novel molecule in SMILES format, denoted  $S$ , along with a reasoning path for proposing the molecule. To this end, the agent utilizes the tool-based analysis of the task prompt and a history of previously generated SMILES to avoid duplication. Based on the collected information, the agent proposes a molecule design strategy. It outlines this strategy in a sequence of  $k$  reasoning steps  $\{r_1, \dots, r_k\}$ , where each  $r_i$  explains how the scientist agent thinks when proposing the SMILES representation of a molecule. After the reasoning process, the agent generates a SMILES string  $S$ .
+
+Verifier. As noted by Pan et al. (2025), reasoning-action mismatch is a critical issue in multiagent frameworks. To mitigate this in our system, we introduce a verifier agent that verifies each reasoning step in  $\{r_1,\dots ,r_k\}$  ensuring that every  $r_i$  is faithfully reflected in the proposed SMILES
+
+![](images/fbe5a476091c899d2053654ecec2c6559af0d1a484664da22d2b40df54985dc2.jpg)  
+(a) Structured feedback of the verifier agent.
+
+![](images/7a54b8df85f9ef91e7d97743b64ff8fb346996f198a007a419eb26ea41f4a3b0.jpg)  
+(b) Structured feedback of the reviewer agent.  
+Figure 3: Examples of structured and stepwise response. The figures illustrate examples of structured feedback mechanisms employed by our agent system for the mestranol_similarity task. (a) The verifier flags a mismatch between reasoning and SMILES and the scientist revises both for consistency. (b) The reviewer suggests reducing rotatable bonds and the scientist reflects the design, improving the score.
+
+$S$ . In detail, it parses each step  $r_i$  and examines whether  $S$  contains the corresponding molecular feature. When discrepancies arise (e.g., when a reasoning step claims the presence of a nitro group, but  $S$  lacks it), the agent flags the inconsistency and produces stepwise feedbacks  $\{f_1^v,\dots ,f_k^v\}$ . Then, the verifier asks the scientist to re-generate the SMILES based on the feedback. This re-generation loop continues until the verifier confirms consistency between the reasoning and SMILES, or until a maximum number of iterations  $t$  reached. If there is no discrepancy detected, it passes the verified reasoning steps  $\{r_1,\ldots ,r_k\}$  and SMILES  $S$  to the reviewer agent.
+
+Reviewer. Inspired by previous works using LLMs as reviewers (Hosseini and Horbach, 2023; Zhang et al., 2022a), we introduce a chemical reviewer agent that evaluates and provides informative feedback. Specifically, the reviewer agent evaluates the verified SMILES  $S$  and reasoning steps  $\{r_1,\dots ,r_k\}$ . Using tool-based analysis of  $S$ , it provides chemically grounded, stepwise feedback  $\{f_1^r,\dots ,f_k^r\}$  aligned with the structure of the reasoning. This feedback includes confirmations of correct reasoning, identification of wrong or missing claims, and suggestions for revision. The sci
+
+entist agent then uses this feedback to refine both the reasoning and molecule  $S$  in the next iteration, enabling iterative improvement.
+
+# 3.2 Details of analyst agents
+
+We implement our multi-agent system with specialized LLM agents, with analyst agents playing a key role in analyzing molecules using domain-specific RDKit (Landrum, 2013) functions. These tools guide molecule generation by providing relevant descriptors to the scientist agent and support the reviewer with interpretable feedback. To enable comprehensive tool utilization and decomposed analysis, we categorize the analyst agents into five molecule-specialized aspects. Each agent targets a distinct aspect of molecular analysis and contributing to a chemically informed and interpretable design process. We provide a detailed list of tools that analyst agents take at Appendix A.
+
+Electronic and topological descriptors. This agent analyzes how electrons are distributed in a molecule and how its atoms are connected, helping to assess properties such as reactivity and stability. It captures patterns that are important for determining whether a molecule is likely to behave well as a drug. As shown in Figure 2, this includes features
+
+such as charge distribution.
+
+Fragment-based functional groups. This agent breaks molecules down into recognizable building blocks, such as rings or functional groups, such as acids or amines, which are commonly used in chemistry. These fragments are easy to interpret and often appear in the stepwise reasoning provided by the scientist agent. Figure 2 shows how the agent highlights specific substructures, such as aromatic rings, that are captured, which is a key component of the task.
+
+Identifiers and representations. This agent translates molecules into standardized formats such as canonicalized SMILES representation, molecular formulas, etc. Figure 2 illustrates how the functional group agent identifies chemically significant motifs such as benzene rings and carboxylic acids, which reflect specific fragment-level reasoning steps and enable chemically grounded feedback.
+
+Structural descriptors. This agent captures basic geometric and physical features of a molecule, such as the number of atoms or bonds it contains. These properties influence how a molecule might behave in real-world conditions, including how it binds to targets or dissolves. As shown in Figure 2, this agent helps evaluate aspects like bond rotatability or ring complexity.
+
+Miscellaneous descriptors. Miscellaneous descriptors agent provides additional analysis that complements the outputs of other agents. It captures properties that might be overlooked, such as molecular surface area, hybridization patterns, or structural irregularities, and helps ensure that the generated molecule is chemically reasonable. As shown in Figure 2, it offers supplementary evidence that strengthens the overall reasoning process.
+
+# 3.3 Structured and stepwise response
+
+In order to ensure that the agent responds to every desired component (e.g., stepwise reasoning, feedback, and SMILES), we guide the agent to output in JSON format using OpenAI API's function<sup>1</sup>. Specifically, the scientist agent generates stepwise reasoning and SMILES, while the verifier and reviewer agents produce stepwise feedback in a designated JSON format.
+
+Also, for a valid and interpretable response, we guide the agents to output stepwise reasoning and feedback. Specifically, the scientist output stepwise reasoning  $\{r_1,\ldots ,r_k\}$  when proposing a SMILES  $S$ . Then, the verifier agent ensures the scientist agent's stepwise reasoning  $\{r_1,\dots ,r_k\}$  is consistent with the output SMILES by providing the interpretable feedback  $\{f_1^d,\dots ,f_k^d\}$ . We visualize the example case in Figure 3a. The verifier agent identifies an inconsistency between the scientist's reasoning and the SMILES, since the butynyl group is not encoded.
+
+In addition, the reviewer critiques the reasoning of the scientist agent with stepwise feedback  $\{f_1^r,\dots ,f_k^r\}$ . As illustrated in Figure 3b, the reviewer agent highlights the issue of increased rotatable bonds. This leads the scientist to revise the design by shortening the alkyne and restoring the methoxy group, which significantly improves the structural similarity score to 1.0. This shows that our approach enables high alignment to target molecule, interpretability, and validity of the properties. We provide a detailed prompt and response example in appendix B.
+
+# 4 Experiments
+
+In this section, we evaluate the effectiveness of our multi-agent LLM system for molecular optimization in low-budget settings. We conduct experiments on the practical molecular optimization (PMO)-1K benchmark, which contains 23 chemically diverse optimization tasks, ranging from rediscovery and scaffold hopping to multi-property objectives. Our framework consists of expert analyst agents—each specialized in task decomposition, SMILES generation, verification, and tool-informed feedback—that collaborate to produce interpretable and high-quality molecular optimization. We compare our results against existing LLM-driven and evolutionary baselines, including LICO and MOLLEO, using various backbone models. We describe the dataset and baselines below, followed by the experimental setting described in Section 4.1. We then present the main benchmark results in Table 1 and provide analysis in Section 4.3.
+
+Datasets. We evaluate on the Practical Molecular Optimization benchmark (Gao et al., 2022), which comprises 23 molecular optimization tasks. Each task defines a specific molecular property or structural constraint, such as rediscovery of known drugs (e.g., celecoxib, thiothixene), similarity to target scaffolds, or maximization of molec
+
+<table><tr><td>Task</td><td>GP BO</td><td>REINVENT</td><td>LICO-L</td><td>Genetic GFN</td><td>Graph GA</td><td>Aug. Mem.</td><td>MOLLEO-B</td><td>MOLLEO-D*</td><td>MT-MOL-D*</td></tr><tr><td>albuterol_similarity</td><td>0.636</td><td>0.496</td><td>0.656</td><td>0.664</td><td>0.583</td><td>0.557</td><td>0.886</td><td>0.883</td><td>0.998</td></tr><tr><td>amlodipine_mpo</td><td>0.519</td><td>0.472</td><td>0.541</td><td>0.534</td><td>0.501</td><td>0.489</td><td>0.637</td><td>0.540</td><td>0.647</td></tr><tr><td>celecoxib_rediscovery</td><td>0.411</td><td>0.370</td><td>0.447</td><td>0.447</td><td>0.424</td><td>0.385</td><td>0.402</td><td>0.512</td><td>0.867</td></tr><tr><td>deco_hop</td><td>0.593</td><td>0.572</td><td>0.596</td><td>0.604</td><td>0.581</td><td>0.579</td><td>0.588</td><td>0.574</td><td>0.842</td></tr><tr><td>drd2</td><td>0.857</td><td>0.775</td><td>0.859</td><td>0.809</td><td>0.833</td><td>0.795</td><td>0.910</td><td>0.812</td><td>0.756</td></tr><tr><td>fexofenadine_mpo</td><td>0.707</td><td>0.650</td><td>0.700</td><td>0.682</td><td>0.666</td><td>0.679</td><td>0.674</td><td>0.680</td><td>0.883</td></tr><tr><td>gsk3b</td><td>0.611</td><td>0.589</td><td>0.617</td><td>0.637</td><td>0.523</td><td>0.539</td><td>0.397</td><td>0.496</td><td>0.308</td></tr><tr><td>isomers_c7h8n2o2</td><td>0.545</td><td>0.725</td><td>0.779</td><td>0.738</td><td>0.735</td><td>0.661</td><td>0.737</td><td>0.850</td><td>0.986</td></tr><tr><td>isomers_c9h10n2o2pf2c1</td><td>0.599</td><td>0.630</td><td>0.672</td><td>0.656</td><td>0.630</td><td>0.596</td><td>0.635</td><td>0.832</td><td>0.914</td></tr><tr><td>jnk3</td><td>0.346</td><td>0.315</td><td>0.336</td><td>0.409</td><td>0.301</td><td>0.294</td><td>0.186</td><td>0.342</td><td>0.125</td></tr><tr><td>median1</td><td>0.213</td><td>0.205</td><td>0.217</td><td>0.219</td><td>0.208</td><td>0.219</td><td>0.236</td><td>0.193</td><td>0.321</td></tr><tr><td>median2</td><td>0.203</td><td>0.188</td><td>0.193</td><td>0.204</td><td>0.181</td><td>0.184</td><td>0.191</td><td>0.197</td><td>0.322</td></tr><tr><td>mestranol_similarity</td><td>0.427</td><td>0.379</td><td>0.423</td><td>0.414</td><td>0.362</td><td>0.393</td><td>0.399</td><td>0.630</td><td>0.996</td></tr><tr><td>osimertinib_mpo</td><td>0.766</td><td>0.737</td><td>0.759</td><td>0.763</td><td>0.751</td><td>0.761</td><td>0.779</td><td>0.753</td><td>0.796</td></tr><tr><td>perindopril_mpo</td><td>0.458</td><td>0.404</td><td>0.473</td><td>0.462</td><td>0.435</td><td>0.422</td><td>0.655</td><td>0.422</td><td>0.542</td></tr><tr><td>qed</td><td>0.912</td><td>0.921</td><td>0.925</td><td>0.928</td><td>0.914</td><td>0.923</td><td>0.919</td><td>0.928</td><td>0.903</td></tr><tr><td>ranolazine_mpo</td><td>0.701</td><td>0.574</td><td>0.687</td><td>0.623</td><td>0.620</td><td>0.614</td><td>0.640</td><td>0.516</td><td>0.233</td></tr><tr><td>scaffold_hop</td><td>0.478</td><td>0.447</td><td>0.480</td><td>0.485</td><td>0.461</td><td>0.460</td><td>0.473</td><td>0.464</td><td>0.646</td></tr><tr><td>sitagliptin_mpo</td><td>0.232</td><td>0.261</td><td>0.315</td><td>0.227</td><td>0.229</td><td>0.245</td><td>0.193</td><td>0.328</td><td>0.067</td></tr><tr><td>thiothixene_rediscovery</td><td>0.351</td><td>0.311</td><td>0.343</td><td>0.377</td><td>0.322</td><td>0.336</td><td>0.416</td><td>0.478</td><td>0.719</td></tr><tr><td>troglitazone_rediscovery</td><td>0.313</td><td>0.246</td><td>0.292</td><td>0.277</td><td>0.267</td><td>0.262</td><td>0.302</td><td>0.387</td><td>0.841</td></tr><tr><td>valsartan_smarts</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td></tr><tr><td>zaleplon_mpo</td><td>0.392</td><td>0.406</td><td>0.404</td><td>0.400</td><td>0.374</td><td>0.415</td><td>0.392</td><td>0.409</td><td>0.625</td></tr><tr><td>Sum of scores (↑)</td><td>11.27</td><td>10.68</td><td>11.71</td><td>11.56</td><td>10.90</td><td>10.81</td><td>11.65</td><td>12.23</td><td>15.42</td></tr></table>
+
+Table 1: Results of PMO-1K benchmark. Tasks are assessed using AUC top-10 averaged by multiple runs. Results with (*) are evaluated from 3 independent runs while the others are assessed from 5 independent runs. We mark the best result in bold and the second-best in underline for each task.
+
+ular property scores such as quantitative estimate of drug-likeness (QED) or logP. Following Gao et al. (2022), we assess performance using the top-10 area under the curve (AUC), which measures the average property score over oracle calls. Additionally considering Nguyen and Grover (2025), the evaluation is conducted for 1K oracle calls, simulating a budget-constrained discovery setting. We use the ZINC 250K (Sterling and Irwin, 2015) dataset to retrieve the top-100 reference molecules for the scientist agent's prompt. We summarize the entire tasks and their descriptions in Appendix C. All molecules are represented in the SMILES format and evaluated using predefined black-box scoring functions consistent with the PMO benchmark protocol.
+
+Baselines. In PMO-1K benchmark, we compare MT-MOL against six baselines: GP BO (Srinivas et al., 2010), REINVENT (Olivecrona et al., 2017), LICO (Nguyen and Grover, 2025), and two variants of MOLLEO (Wang et al., 2025) (MOLLEOB, and MOLLEO-D). MOLLEO operates through LLM-guided mutation and crossover, using different base models (BioT5 (Pei et al., 2023) and DeepSeek-V3 (Liu et al., 2024). We evaluated two versions of our framework (Ours-D) using DeepSeek-V3 as a backbone for all the agent roles.
+
+In ChemCoTBenchmark, we compare our framework against 22 baselines: (1) reasoning LLMs, including Deepseek-R (Guo et al., 2025), o1-mini (OpenAI Team, 2024), o3-mini (OpenAI
+
+Team, 2025), Gemini-2.5-pro (DeepMind Team, 2025), Claude-3.7-Sonnet-thinking (Anthropic Team, 2025), Qwen-3-thinking (Qwen Team, 2025), and Llama-Nemotron-thinking (Bercovich et al., 2025), (2) general-purpose non-reasoning LLMs, such as GPT-4o (Hurst et al., 2024), Qwen-2.5/3 (Yang et al., 2025), Llama-3.3 (Grattafori et al., 2024), Gemma-2 (Team et al., 2024), Phi-4 (Abdin et al., 2024), and OLMo2 (Team OLMo, 2025) and (3) biomolecular LLMs like BioMedGPT (Luo et al., 2023), BioMistral (Labrak et al., 2024), and Text+Chem T5 (Christofidellis et al., 2023).
+
+# 4.1 PMO Benchmark
+
+Table 1 reports the performance of our framework and competing methods in all 23 PMO tasks. MT-MOL-D*, achieves the best performance in 15 of 23 tasks, significantly outperforming all baselines, including MOLLEO and LICO. In particular, MT-MOL surpasses the SOTA AUC sum of 12.23 (MOLLEO-D*) with a score of 15.42, marking a substantial improvement in the overall efficiency of optimization. The performance gap is particularly large on chemically complex tasks such as celecoxib_rediscovery and amlodipine_mpo, where MT-MOL-D* outperforms the previous best by more than 0.3 AUC points.
+
+In Figure 4, we visualize the top-10 AUC curves for every 23 PMO tasks. MT-MOL consistently achieves faster and higher AUC trajectories compared to MOLLEO-D* across tasks such
+
+![](images/66e208c1760328ced6a9647b79eb9cec5cdf6426729b63eb0b15cafa5741cc01.jpg)  
+Figure 4: Top-10 AUC curves. Top-10 average AUC curves on the PMO benchmark, averaged over three random seeds. Our method consistently surpasses MOLLEO by achieving higher and faster-rising AUC curves, highlighting the effectiveness of tool-guided reasoning and multi-agent feedback in molecular optimization.
+
+as albuterol_similarity, amlodipine_mpo, osimertinib_mpo, and troglitazon_rediscovery. These results suggest that our tool-aware reasoning, stepwise validation, and multi-agent feedback loop generate the desired molecule SMILES in the early stage while achieving high oracle value. The improvements are especially pronounced in the early stages of generation, indicating that Mol-Agent makes more efficient use of oracle calls.
+
+# 4.2 ChemCoTBenchmark
+
+To further validate our framework against diverse LLMs, we conducted an evaluation on the ChemCoTBenchmark (Li et al., 2025). In this setting, we compared the baselines with MT-MOLincreasing the feedback loop  $N = \{1,3,5\}$ . The  $N = 1$  setting is a single execution pass without iterative feedback, while other settings include multiple feedback iterations. We report performance on three challenging molecular optimization tasks, including GSK3-  $\beta$ , QED, and DRD2. Following the benchmark's experimental setting, we use the mean property improvement as the evaluation metric.
+
+Table 2 shows our framework achieves a state-of-the-art, significantly outperforming all 22 baseline models, including specialized reasoning-enhanced LLMs. Surprisingly, our method achieves competitive performance across all tasks even in the single-pass  $(N = 1)$  setting. This result highlights the importance of incorporating specialized chemical tools and a verification stage with multi-agent framework.
+
+A comparison of our framework's performance across different feedback loop iterations ( $N \in \{1,3,5\}$ ) demonstrates the clear benefit of the iterative feedback process, as performance consistently improves with more iterations on every task. When compared to other 22 large language model baselines, the performance gap is particularly large on the QED and GSK3- $\beta$  tasks. In particular, ( $N = 5$ ) setting shows substantial gains, achieving a mean improvement of 0.77 on QED and 0.18 on GSK3- $\beta$ , far exceeding the top baseline scores of 0.21 and 0.04, respectively. These results suggest that our tool-aware reasoning and multi-agent feedback loop enable a more accurate optimization process, allowing the framework to progressively refine
+
+<table><tr><td>Task</td><td>Gemini-2.5-pro
+-think</td><td>Claude3.7-sonnet
+-think</td><td>DeepSeek-R1
+-think</td><td>o3-mini
+@20250103</td><td>o1-mini
+@20240912</td><td>Qwen2-235B
+-A22B-think</td><td>Qwen3-32B
+-think</td><td>Llama-Nemo
+-49B-think</td><td>GPT-4o
+@20241120</td><td>DeepSeek
+-V3</td><td>Gemini-2.0
+-flash</td><td>Qwen2-235B
+-A22B</td><td>Qwen3
+-32B</td></tr><tr><td>GSK3-β</td><td>0.04</td><td>0.01</td><td>-0.02</td><td>-0.03</td><td>-0.08</td><td>0.01</td><td>-0.02</td><td>-0.12</td><td>-0.04</td><td>0.00</td><td>0.00</td><td>0.00</td><td>-0.02</td></tr><tr><td>QED</td><td>0.21</td><td>0.09</td><td>0.05</td><td>0.17</td><td>0.07</td><td>0.01</td><td>0.02</td><td>-0.16</td><td>0.05</td><td>0.08</td><td>0.10</td><td>0.01</td><td>0.02</td></tr><tr><td>DRD2</td><td>0.35</td><td>0.18</td><td>0.10</td><td>0.18</td><td>-0.03</td><td>0.03</td><td>0.00</td><td>-0.05</td><td>0.05</td><td>0.02</td><td>0.15</td><td>0.01</td><td>-0.01</td></tr></table>
+
+<table><tr><td>Task</td><td>Qwen2.5-72B -Instruct</td><td>Qwen2.5-32B -Instruct</td><td>Llama-3.3-70B -Instruct</td><td>Llama-Nemo -Super-49B</td><td>Gemma-2 -27b-it</td><td>Phi-4 -14B</td><td>OLMo2-32B -Instruct</td><td>BioMedGPT -7B</td><td>BioMistral -7B</td><td>MT-Mol (N=1)</td><td>MT-Mol (N=3)</td><td>MT-Mol (N=5)</td></tr><tr><td>GSK3-β</td><td>-0.01</td><td>-0.02</td><td>-0.02</td><td>-0.03</td><td>-0.01</td><td>-0.04</td><td>-0.11</td><td>-0.08</td><td>-0.01</td><td>0.12</td><td>0.16</td><td>0.18</td></tr><tr><td>QED</td><td>0.03</td><td>-0.01</td><td>0.07</td><td>0.02</td><td>0.05</td><td>0.11</td><td>-0.13</td><td>-0.29</td><td>0.00</td><td>0.59</td><td>0.73</td><td>0.77</td></tr><tr><td>DRD2</td><td>0.04</td><td>0.04</td><td>-0.02</td><td>-0.02</td><td>-0.03</td><td>-0.04</td><td>-0.11</td><td>-0.09</td><td>0.00</td><td>0.14</td><td>0.27</td><td>0.38</td></tr></table>
+
+Table 2: Results of ChemCoTBenchmark benchmark Mean property improvement on the molecular optimization task, comparing baseline LLMs against our framework using a feedback loop of  $N = \{1,3,5\}$ . We mark the best result in bold and the second-best in underline for each task.  
+
+<table><tr><td>Task</td><td>Setting</td><td>AUC-Top10</td></tr><tr><td rowspan="4">osimertinib_mpo</td><td>MT-MOL</td><td>0.796 ± 0.005</td></tr><tr><td>w/o Tool</td><td>0.694 ± 0.054</td></tr><tr><td>w/o Reviewer</td><td>0.619 ± 0.140</td></tr><tr><td>w/o Double checker</td><td>0.704 ± 0.017</td></tr><tr><td rowspan="4">albuterol_similarity</td><td>MT-MOL</td><td>0.998 ± 0.000</td></tr><tr><td>w/o Tool</td><td>0.750 ± 0.021</td></tr><tr><td>w/o Reviewer</td><td>0.991 ± 0.003</td></tr><tr><td>w/o Double checker</td><td>0.996 ± 0.003</td></tr><tr><td rowspan="4">mestranol_similarity</td><td>MT-MOL</td><td>0.996 ± 0.001</td></tr><tr><td>w/o Tool</td><td>0.831 ± 0.052</td></tr><tr><td>w/o Reviewer</td><td>0.990 ± 0.002</td></tr><tr><td>w/o Double checker</td><td>0.994 ± 0.002</td></tr></table>
+
+Table 3: Ablation study. AUC-Top10 score under different agent removals for each task.
+
+molecular candidates and achieve superior performance in complex chemical design spaces.
+
+# 4.3 Ablation studies
+
+To evaluate the contribution of each component in our multi-agent framework, we perform an ablation study on a subset of tasks from the PMO benchmark. Specifically, we assess the impact of removing (1) all five expert analyst agents, (2) the verifier agent, and (3) the reviewer agent. We report top-10 AUC scores averaged over three random seeds in Table 3.
+
+One can observe that removing the analyst agents consistently leads to a substantial drop in performance across all tasks. For instance, the AUC score on albuterol_similarity drops from 0.998 to 0.750, highlighting that the expert analyst agents provide essential domain-specific descriptors.
+
+Also removing reviwer agent causes noticeable degradation on tasks like osimertinib_mpo (from 0.796 to 0.619). Similarly, removing verifier agent shows modest performance drops when ablated, particularly on more challenging tasks.
+
+Overall, these results underscore the importance of tool-guided analysis, structured reasoning ver-
+
+ification, and feedback loops in our multi-agent system.
+
+# 5 Conclusion
+
+In this paper, we introduced MT-MOL, a multiagent framework for molecular optimization and generation that combines tool-guided reasoning with structured collaboration among specialized LLM agents. Our system integrates five expert analyst agents, each equipped with domain-specific chemistry functions, to guide and critique molecule design. Through systematic interaction among scientist, verifier, and reviewer agents, MT-MOL achieves interpretable, chemically valid, and task-aligned molecular optimization. Our experiments on the PMO-1K benchmark demonstrate that MT-MOL outperforms strong baselines, including LICO and MOLLEO, achieving state-of-the-art performance on 15 out of 23 tasks. Furthermore, evaluations on the ChemCoTbench benchmark confirm our framework's superiority, where it significantly outperforms 22 strong large language model baselines. The results highlight the effectiveness of structured reasoning, tool-based validation, and multi-agent feedback in navigating the complex chemical space. This work provides the multiagent system with comprehensive and systematic tool-augmented responses, accelerating molecular optimization and enabling transparent scientific discovery.
+
+# Broader Impact
+
+Our work may help democratize access to molecular design expertise by enabling non-expert users to interact with intelligent agents that provide chemically grounded suggestions. Furthermore, the stepwise reasoning and feedback mechanisms embedded in our framework can serve as educational tools to help students and researchers understand the rationale behind molecule design decisions.
+
+However, broader adoption of AI-assisted molecule design systems also raises potential ethical and social concerns. These include the misuse of generative tools for designing harmful substances, propagation of biases present in pretraining data, and the risk of over-reliance on AI-generated outputs without sufficient domain validation. Responsible deployment will require integrating safety checks, transparency mechanisms, and human-in-the-loop oversight.
+
+# Limitations
+
+Our framework relies heavily on rule-based cheminformatics tools (e.g., RDKit) and predefined feature sets, which may limit generalization to novel chemical spaces or underrepresented functional groups. Moreover, while the multi-agent structure enables interpretability, it introduces additional computational overhead compared to single-agent models, potentially limiting scalability in resource-constrained settings.
+
+Additionally, our experiments are conducted only in English and do not explore across other languages. This may limit usability in multilingual research environments or for integration with non-English scientific literature and databases.
+
+# Acknowledgments
+
+This work was partly supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Support Program(KAIST)), National Research Foundation of Korea(NRF) grant funded by the Ministry of Science and ICT(MSIT) (No. RS-2022-NR072184), GRDC(Global Research Development Center) Cooperative Hub Program through the National Research Foundation of Korea(NRF) grant funded by the Ministry of Science and ICT(MSIT) (No. RS-2024-00436165), and the Institute of Information & Communications Technology Planning &
+
+Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2025-02304967, AI Star Fellowship(KAIST)). This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2025-02653113, High-Performance Research AI Computing Infrastructure Support at the 2 PFLOPS Scale)
+
+# References
+
+Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, and 1 others. 2024. Phi-4 technical report. arXiv preprint arXiv:2412.08905.  
+Duarte M Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, and 1 others. 2024. Tower: An open multilingual large language model for translation-related tasks. Conference on Language Modeling.  
+Anthropic Team. 2025. Claude-3.7-sonnet: Hybrid reasoning model. Web page. Accessed: 2025-09-20.  
+M Saiful Bari, Yazeed Alnumay, Norah A Alzahrani, Nouf M Alotaibi, Hisham A Alyahya, Sultan Al-Rashed, Faisal A Mirza, Shaykhah Z Alsubaie, Hassan A Allahmed, Ghadah Alabduljabbar, and 1 others. 2025. Allam: Large language models for arabic and english. International Conference on Learning Representations.  
+Yoshua Bengio, Salem Lahlou, Tristan Deleu, Edward J Hu, Mo Tiwari, and Emmanuel Bengio. 2023. Gflownet foundations. Journal of Machine Learning Research, 24(210):1-55.  
+Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, and 1 others. 2025. Llama-nemotron: Efficient reasoning models. arXiv preprint arXiv:2505.00949.  
+Andres M Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2024. Chemcrow: Augmenting large-language models with chemistry tools. Nature Machine Intelligence.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, and 1 others. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
+
+Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, and 1 others. 2023. Agent-verse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. International Conference on Learning Representations, 2(4):6.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, and 1 others. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113.  
+Dimitrios Christofidellis, Giorgio Giannone, Jannis Born, Ole Winther, Teodoro Laino, and Matteo Manica. 2023. Unifying molecular and textual representations via multi-task language modelling. In International Conference on Machine Learning, pages 6140-6157. PMLR.  
+DeepMind Team. 2025. Gemini 2.5 pro preview: even better coding performance. Web page. Accessed: 2025-09-20.  
+Junnan Dong, Qinggang Zhang, Chuang Zhou, Hao Chen, Daochen Zha, and Xiao Huang. 2024. Cost-efficient knowledge-based question answering with large language models. Advances in Neural Information Processing Systems.  
+Wenhao Gao, Tianfan Fu, Jimeng Sun, and Connor Coley. 2022. Sample efficiency matters: a benchmark for practical molecular optimization. Advances in neural information processing systems, 35:21342-21357.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.  
+Jeff Guo and Philippe Schwaller. 2024. Augmented memory: Sample-efficient generative molecular design with reinforcement learning. *Jacs Au*, 4(6):2160-2172.  
+Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, and 1 others. 2024. Metagpt: Meta programming for multi-agent collaborative framework. International Conference on Learning Representations, 3(4):6.  
+Mohammad Hosseini and Serge PJM Horbach. 2023. Fighting reviewer fatigue or amplifying bias? considerations and recommendations for use of chatgpt and other large language models in scholarly peer review. Research integrity and peer review, 8(1):4.
+
+Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Connor W Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik. 2021. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. arXiv preprint arXiv:2102.09548.  
+Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.  
+Jan H Jensen. 2019. A graph-based genetic algorithm and generative model/monte carlo tree search for the exploration of chemical space. Chemical science, 10(12):3567-3572.  
+Jaehyung Kim, Jaehyun Nam, Sangwoo Mo, Jongjin Park, Sang-Woo Lee, Minjoon Seo, Jung-Woo Ha, and Jinwoo Shin. 2024. Sure: Summarizing retrievals using answer candidates for open-domain qa of llms. International Conference on Learning Representations.  
+Yanis Labrak, Adrien Bazoge, Emmanuel Morin, Pierre-Antoine Gourraud, Mickael Rouvier, and Richard Dufour. 2024. Biomistral: A collection of open-source pretrained large language models for medical domains. arXiv preprint arXiv:2402.10373.  
+Greg Landrum. 2013. Rdkit documentation. Release, 1(1-79):4.  
+Hao Li, He Cao, Bin Feng, Yanjun Shao, Xiangru Tang, Zhiyuan Yan, Li Yuan, Yonghong Tian, and Yu Li. 2025. Beyond chemical qa: Evaluating llm's chemical reasoning with modular chemical operations. arXiv preprint arXiv:2505.21318.  
+Huao Li, Yu Quan Chong, Simon Stepputtis, Joseph Campbell, Dana Hughes, Michael Lewis, and Katia Sycara. 2023a. Theory of mind for multi-agent collaboration via large language models. Empirical Methods in Natural Language Processing.  
+Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, and 1 others. 2023b. Starcoder: may the source be with you! Transactions on Machine Learning Research.  
+Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437.  
+Yixin Liu, AR Fabbri, P Liu, D Radev, and A Cohan. On learning to summarize with large language models as references (2024). North American Chapter of the Association for Computational Linguistics.  
+Yizhen Luo, Jiahuan Zhang, Siqi Fan, Kai Yang, Yushuai Wu, Mu Qiao, and Zaiqing Nie. 2023. Biomedgpt: Open multimodal generative pre-trained
+
+transformer for biomedicine. arXiv preprint arXiv:2308.09442.  
+Kelvin Luu, Xinyi Wu, Rik Koncel-Kedziorski, Kyle Lo, Isabel Cachola, and Noah A Smith. 2021. Explaining relationships between scientific documents. *Association for Computational Linguistics*.  
+Nathalia Nascimento, Paulo Alencar, and Donald Cowan. 2023. Self-adaptive large language model (llm)-based multiagent systems. In 2023 IEEE International Conference on Autonomic Computing and Self-Organizing Systems Companion (ACSOS-C), pages 104-109. IEEE.  
+Tung Nguyen and Aditya Grover. 2025. Lico: Large language models for in-context molecular optimization. International Conference on Learning Representations.  
+Marcus Olivecrona, Thomas Blaschke, Ola Engkvist, and Hongming Chen. 2017. Molecular de-novo design through deep reinforcement learning. Journal of cheminformatics, 9:1-14.  
+OpenAI Team. 2024. O1-mini: advancing cost-efficient reasoning. Web page. Accessed: 2025-09-20.  
+OpenAI Team. 2025. Openai o3-mini. Web page. Accessed: 2025-09-20.  
+Melissa Z Pan, Mert Cemri, Lakshya A Agrawal, Shuyi Yang, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Kannan Ramchandran, Dan Klein, and 1 others. 2025. Why do multiagent systems fail? In ICLR 2025 Workshop on Building Trust in Language Models and Applications.  
+Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. 2023. Biot5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations. Empirical Methods in Natural Language Processing.  
+Qwen Team. 2025. Qwen3: Think deeper, act faster. Web page. Accessed: 2025-09-20.  
+Niranjan Srinivas, Andreas Krause, Sham M Kakade, and Matthias Seeger. 2010. Gaussian process optimization in the bandit setting: No regret and experimental design. International Conference on Machine Learning.  
+Teague Sterling and John J Irwin. 2015. Zinc 15-ligand discovery for everyone. Journal of chemical information and modeling, 55(11):2324-2337.  
+Hongda Sun, Yuxuan Liu, Chengwei Wu, Haiyu Yan, Cheng Tai, Xin Gao, Shuo Shang, and Rui Yan. 2024. Harnessing multi-role capabilities of large language models for open-domain question answering. In Proceedings of the ACM Web Conference 2024, pages 4372-4382.
+
+Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118.  
+Team OLMo. 2025. 2 olmo 2 furious. Web page. Accessed: 2025-09-20.  
+Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, and 1 others. 2023. Scientific discovery in the age of artificial intelligence. Nature, 620(7972):47-60.  
+Haorui Wang, Marta Skreta, Cher-Tian Ser, Wenhao Gao, Lingkai Kong, Felix Strieth-Kalthoff, Chenru Duan, Yuchen Zhuang, Yue Yu, Yanqiao Zhu, and 1 others. 2025. Efficient evolutionary search over chemical space with large language models. International Conference on Learning Representations.  
+David Weininger. 1988. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of chemical information and computer sciences, 28(1):31-36.  
+Zhenxing Wu, Odin Zhang, Xiaorui Wang, Li Fu, Huifeng Zhao, Jike Wang, Hongyan Du, Dejun Jiang, Yafeng Deng, Dongsheng Cao, and 1 others. 2024. Leveraging language model for advanced multiproducty molecular optimization via prompt engineering. Nature Machine Intelligence, pages 1-11.  
+An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.  
+Geyan Ye, Xibao Cai, Houtim Lai, Xing Wang, Junhong Huang, Longyue Wang, Wei Liu, and Xiangxiang Zeng. 2025. Drugassist: A large language model for molecule optimization. Briefings in Bioinformatics, 26(1):bbae693.  
+Ceyao Zhang, Kaijie Yang, Siyi Hu, Zihao Wang, Guanghe Li, Yihang Sun, Cheng Zhang, Zhaowei Zhang, Anji Liu, Song-Chun Zhu, and 1 others. 2024. Proagent: building proactive cooperative agents with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17591-17599.  
+Jiayao Zhang, Hongming Zhang, Zhun Deng, and Dan Roth. 2022a. Investigating fairness disparities in peer review: A language model enhanced approach. arXiv preprint arXiv:2211.06398.  
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, and 1 others. 2022b. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068.
+
+# A List of tools
+
+List of tools are provided by categories. Tools of electronic and topological descriptors are provided at Table 4, fragment based functional groups at Table 5, Table 6, and Table 7, identifiers and representations at Table 8, structural descriptors at Table 10, Table 11, and miscellaneous descriptors at Table 9.
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>bcut2d</td><td>Implements BCUT descriptors From J. Chem. Inf. Comput. Sci., Vol. 39, No. 1, 1999Diagonal elements are (currently) atomic mass, gasteiger charge,crippen logP and crippen MR- Returns the 2D BCUT2D descriptors vector as described in re-turns [mass eigen value high, mass eigen value low, gasteiger charge eigenvalue high, gasteiger charge low,</td></tr><tr><td>calcautocorr2d</td><td>Returns 2D Autocorrelation descriptor vector using a speci-fied atom property.</td></tr><tr><td>calcchi0n</td><td>Calculates the Chi0n index, a valence-based topological de-scriptor.</td></tr><tr><td>calcchi0v</td><td>Calculates the Chi0v index, a non-valence-based topological descriptor.</td></tr><tr><td>calcchi1n</td><td>Calculates the Chi1n index using atom connectivity and op-tionally forces calculation.</td></tr><tr><td>calcchi1v</td><td>Calculates the Chi1v index, a valence-corrected form of Chi1n.</td></tr><tr><td>calcchi2n</td><td>Calculates the Chi2n index, a higher-order topological de-scriptor (non-valence- based).</td></tr><tr><td>calcchi2v</td><td>Calculates the Chi2n index, a higher-order topological de-scriptor (non-valence- based).</td></tr><tr><td>calcchi3n</td><td>Calculates the Chi3n index for extended connectivity (non-valence-based).</td></tr><tr><td>calcchi3v</td><td>Calculates the Chi3v index with valence correction for deeper molecular topology.</td></tr><tr><td>calcchi4n</td><td>Calculates the Chi4n index, further extending non-valence connectivity descriptors.</td></tr><tr><td>calcchi4v</td><td>Calculates the Chi4v index, a valence-aware descriptor at a 4th topological level</td></tr></table>
+
+Table 4: List of Electronic Topological Descriptors tools
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>get_min_ring_freqency</td><td>Return the least frequent known ring system in the molecule with its frequency.</td></tr><tr><td>remove_stereo_from_smiles</td><td>Removes stereochemistry from SMILES and returns canonical SMILES and InChI Key.</td></tr><tr><td>get_spiro_atoms</td><td>Returns atom indices that are shared between two rings (spiro atoms).</td></tr><tr><td>max_ring_size</td><td>Returns the size of the largest ring in the molecule.</td></tr><tr><td>ring.stats</td><td>Returns the number of rings and the size of the largest ring in the molecule.</td></tr><tr><td>countisable</td><td>Returns the number of molecular fragments present in the SMILES.</td></tr><tr><td>get_largest Fragment</td><td>Returns the SMILES of the largest fragment by atom count in a molecule.</td></tr><tr><td>fr_phos_acid</td><td>Number of phosphoric acid groups</td></tr><tr><td>fr_Al_COO</td><td>Number of aliphatic carboxylic acids</td></tr><tr><td>fr_Al_OH</td><td>Number of aliphatic hydroxyl groups</td></tr><tr><td>fr_Al_OH_noTert</td><td>Number of aliphatic hydroxyl groups excluding tert-OH</td></tr><tr><td>fr_ArN</td><td>Number of N functional groups attached to aromatics</td></tr><tr><td>fr_Ar_COO</td><td>Number of Aromatic carboxylic acid</td></tr><tr><td>fr_Ar_N</td><td>Number of aromatic nitrogens</td></tr><tr><td>fr_Ar_NH</td><td>Number of aromatic amines</td></tr><tr><td>fr_Ar_OH</td><td>Number of aromatic hydroxyl groups</td></tr><tr><td>fr_COO</td><td>Number of carboxylic acids</td></tr><tr><td>fr_CO02</td><td>Number of carboxylic acids</td></tr><tr><td>fr_C_O</td><td>Number of carbonyl O</td></tr><tr><td>fr_C_O_noCOO</td><td>Number of carbonyl O, excluding COOH</td></tr><tr><td>fr_C_S</td><td>Number of thiocarbonyl</td></tr><tr><td>fr_HOCCN</td><td>Number of C(OH)CCN-Ctert-alkyl or C(OH)CCNcyclic</td></tr><tr><td>fr_Imine</td><td>Number of Imines</td></tr><tr><td>fr_NH0</td><td>Number of Tertiary amines</td></tr><tr><td>fr_NH1</td><td>Number of Secondary amines</td></tr><tr><td>fr_NH2</td><td>Number of Primary amines</td></tr><tr><td>fr_N_0</td><td>Number of hydroxylamine groups</td></tr><tr><td>fr_Nhpyrrole</td><td>Number of H-pyrrole nitrogens</td></tr><tr><td>fr_SH</td><td>Number of thiol groups</td></tr></table>
+
+Table 5: List of Fragment Based Functional Groups tools (1/3)
+
+<table><tr><td>fr_aldehyde</td><td>Number of aldehydes</td></tr><tr><td>fr_alkyl_carbamate</td><td>Number of alkyl carbamates (subject to hydrolysis)</td></tr><tr><td>fr_alkyl_halide</td><td>Number of alkyl halides</td></tr><tr><td>fr_allylic_oxid</td><td>Number of allylic oxidation sites excluding steroid dienone</td></tr><tr><td>fr_amide</td><td>Number of amides</td></tr><tr><td>fr_amidine</td><td>Number of amidine groups</td></tr><tr><td>fr_aniline</td><td>Number of anilines</td></tr><tr><td>Function</td><td>Description</td></tr><tr><td>fr_aryl_methyl</td><td>Number of aryl methyl sites for hydroxylation</td></tr><tr><td>fr_azide</td><td>Number of azide groups</td></tr><tr><td>fr_azo</td><td>Number of azo groups</td></tr><tr><td>fr_barbitur</td><td>Number of barbiturate groups</td></tr><tr><td>fr_benzene</td><td>Number of benzene rings</td></tr><tr><td>fr_benzodiazepine</td><td>Number of benzodiazepines with no additional fused rings</td></tr><tr><td>fr_bicyclic</td><td>Bicyclic</td></tr><tr><td>fr_diazo</td><td>Number of diazo groups</td></tr><tr><td>fr_dihydropyridine</td><td>Number of dihydropyridines</td></tr><tr><td>fr_epoxide</td><td>Number of epoxide rings</td></tr><tr><td>fr_ester</td><td>Number of esters</td></tr><tr><td>fr_ether</td><td>Number of ether oxygens (including phenoxy)</td></tr><tr><td>fr_furan</td><td>Number of furan rings</td></tr><tr><td>fr_guanido</td><td>Number of guanidine groups</td></tr><tr><td>fr HALogen</td><td>Number of halogens</td></tr><tr><td>fr_hdrzine</td><td>Number of hydrazine groups</td></tr><tr><td>fr_hdrzone</td><td>Number of hydrazone groups</td></tr><tr><td>fr_imidazole</td><td>Number of imidazole rings</td></tr><tr><td>fr_imide</td><td>Number of imide groups</td></tr><tr><td>fr_isocyan</td><td>Number of isocyanates</td></tr><tr><td>fr_isothiocyan</td><td>Number of isothiocyanates</td></tr><tr><td>fr_ketone</td><td>Number of ketones</td></tr><tr><td>fr_ketone_TOpliss</td><td>Number of ketones excluding diaryl, a,b-unsat. dienones, heteroatom on Calpha</td></tr></table>
+
+Table 6: List of Fragment Based Functional Groups tools (2/3)
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>fr_lactam</td><td>Number of beta lactams</td></tr><tr><td>fr_lactone</td><td>Number of cyclic esters (lactones)</td></tr><tr><td>fr_methoxy</td><td>Number of methoxy groups -OCH3</td></tr><tr><td>fr_morpholine</td><td>Number of morpholine rings</td></tr><tr><td>fr_nitrile</td><td>Number of nitriles</td></tr><tr><td>fr_nitro</td><td>Number of nitro groups</td></tr><tr><td>fr_nitro_aram</td><td>Number of nitro benzene ring substituents</td></tr><tr><td>fr_nitro_aram_nonortho</td><td>Number of non-ortho nitro benzene ring substituents</td></tr><tr><td>fr_nitroso</td><td>Number of nitroso groups, excluding NO2</td></tr><tr><td>fr_oxazole</td><td>Number of oxazole rings</td></tr><tr><td>fr_oxime</td><td>Number of oxime groups</td></tr><tr><td>fr_para_hydroxylation</td><td>Number of para-hydroxylation sites</td></tr><tr><td>fr_phenol</td><td>Number of phenols</td></tr><tr><td>fr_phenol_noOrthoHbond</td><td>Number of phenolic OH excluding ortho intramolecular Hbond substituents</td></tr><tr><td>fr_phos_ester</td><td>Number of phosphoric ester groups</td></tr><tr><td>fr_piperdine</td><td>Number of piperidine rings</td></tr><tr><td>fr_piperzine</td><td>Number of piperzine rings</td></tr><tr><td>fr_priamide</td><td>Number of primary amides</td></tr><tr><td>fr_prisulfonamd</td><td>Number of primary sulfonamides</td></tr><tr><td>fr_pyridine</td><td>Number of pyridine rings</td></tr><tr><td>fr_quatN</td><td>Number of quaternary nitrogens</td></tr><tr><td>fr_sulfide</td><td>Number of thioether</td></tr><tr><td>fr_sulfonamd</td><td>Number of sulfonamides</td></tr><tr><td>fr_sulfone</td><td>Number of sulfone groups</td></tr><tr><td>fr_term_acetylene</td><td>Number of terminal acetylenes</td></tr><tr><td>fr_tetrazole</td><td>Number of tetrazole rings</td></tr><tr><td>fr_thiazole</td><td>Number of thiazole rings</td></tr><tr><td>fr_thiocyan</td><td>Number of thiocyanates</td></tr><tr><td>fr_thiophene</td><td>Number of thiophene rings</td></tr><tr><td>fr_unbrch_alkane</td><td>Number of unbranched alkanes of at least 4 members (excludes halogenated alkanes)</td></tr><tr><td>fr_urea</td><td>Number of urea groups</td></tr></table>
+
+Table 7: List of Fragment Based Functional Groups tools (3/3)  
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>get_rdkit_complexity</td><td>Returns the Bertz molecular complexity index of the molecule.</td></tr><tr><td>get_rdkit_number_of_atoms</td><td>Returns the number of atoms in the molecule.</td></tr><tr><td>get_rdkit_number_of_bonds</td><td>Returns the number of bonds in the molecule.</td></tr><tr><td>get_rdkit_rotatable_bond_count</td><td>Returns the number of rotatable bonds in the molecule.</td></tr><tr><td>get_rdkit_h_bond_donor_count</td><td>Returns the number of hydrogen bond donors in the molecule.</td></tr><tr><td>get_rdkit_h_bond.acceptor_count</td><td>Returns the number of hydrogen bond acceptors in the molecule.</td></tr><tr><td>get_rdkit_molecular_formula</td><td>Returns the molecular formula of the molecule.</td></tr><tr><td>get_rdkit_canonical_smiles</td><td>Returns the canonical SMILES of the molecule.</td></tr><tr><td>get_rdkit_inchi</td><td>Returns the InChI string of the molecule.</td></tr></table>
+
+Table 8: List of Identifiers and Representations tools
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>smi2mol_with Errors</td><td>Attempts to parse SMILES and returns validation status with error/warning messages.</td></tr><tr><td>calcmolformula</td><td>Returns the molecule 2019s formula</td></tr><tr><td>calccrippendescriptors</td><td>Returns a 2-tuple with the Wildman-Crippen logp,mr values</td></tr><tr><td>calcfractioncsp3</td><td>Returns the fraction of C atoms that are SP3 hybridized</td></tr><tr><td>calckappa1</td><td>Calculates the first Kier shape index, reflecting molecular linearity based on atom and bond counts.</td></tr><tr><td>calckappa2</td><td>Computes the second Kier shape index, indicating molecular cyclicity and branching.</td></tr><tr><td>calckappa3</td><td>Computes the third Kier shape index, sensitive to molecular flexibility and complex ring structures.</td></tr><tr><td>calclabuteasa</td><td>Returns the Labute ASA value for a molecule</td></tr><tr><td>calcpbf</td><td>Returns the PBF (plane of best fit) descriptor</td></tr><tr><td>calcphi</td><td>Estimates the molecular flexibility index based on the number of rotatable bonds and ring structures.</td></tr><tr><td>getconnectivityinvariants</td><td>Returns connectivity invariants (ECFP-like) for a molecule.</td></tr><tr><td>getfeatureinvariants</td><td>Returns feature invariants (FCFP-like) for a molecule.</td></tr><tr><td>mqns_</td><td>Computes Molecular Quantum Numbers, a 42-dimensional vector of counts for various atom types, bonds, and topological features.</td></tr><tr><td>peoe_vsa_</td><td>Computes descriptors combining partial charges (Gasteiger) with van der Waals surface areas in defined bins.</td></tr><tr><td>smr_vsa_</td><td>Calculates descriptors combining molar refractivity contributions with surface areas in predefined ranges.</td></tr><tr><td>slogp_vsa_</td><td>Computes descriptors by combining atomic logP contributions (Wildman-Crippen) with van der Waals surface areas.</td></tr></table>
+
+Table 9: List of Other Descriptors tools
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>get_center</td><td>Computes the geometric center of a conformer generated from the input SMILES.</td></tr><tr><td>get_shape Moments</td><td>Calculates NPR1 and NPR2 shape descriptors from a generated conformer.</td></tr><tr><td>refine_conformers</td><td>Refines 3D conformers based on energy and RMSD thresholds.</td></tr><tr><td>get_conformer_energies</td><td>Returns the energies of multiple conformers generated from the input molecule.</td></tr><tr><td>calcnumaliphaticcarbocycles</td><td>Returns the number of aliphatic (containing at least one non-aromatic bond) carbocycles for a molecule</td></tr><tr><td>calcnumaliphaticheterocycles</td><td>Returns the number of aliphatic (containing at least one non-aromatic bond) heterocycles for a molecule</td></tr><tr><td>calcnumaliphaticrings</td><td>Returns the number of aliphatic (containing at least one non-aromatic bond) rings for a molecule</td></tr><tr><td>calcnumamidebonds</td><td>Returns the number of amide bonds in a molecule</td></tr><tr><td>calcnumaromaticcarbocycles</td><td>Returns the number of aromatic carbo-cycles for a molecule</td></tr><tr><td>calcnumaromaticheterocycles</td><td>Returns the number of aromatic hetero-cycles for a molecule</td></tr><tr><td>calcnumaromaticrings</td><td>Returns the number of aromatic rings for a molecule</td></tr><tr><td>calcnumatomstereocenters</td><td>Returns the total number of atomic stereocenters (specified and unspecified)</td></tr><tr><td>calcnumatoms</td><td>Returns the total number of atoms for a molecule</td></tr><tr><td>calcnumhba</td><td>Returns the number of H-bond acceptors for a molecule</td></tr><tr><td>calcnumhbd</td><td>Returns the number of H-bond donors for a molecule</td></tr></table>
+
+Table 10: List of Structural Descriptors tools
+
+<table><tr><td>Function</td><td>Description</td></tr><tr><td>calcnumheavyatoms</td><td>Returns the number of heavy atoms for a molecule</td></tr><tr><td>calcnumheteroatoms</td><td>Returns the number of heteroatoms for a molecule</td></tr><tr><td>calcnumheterocycles</td><td>Returns the number of heterocycles for a molecule</td></tr><tr><td>calcnumlipinskihba</td><td>Returns the number of Lipinski H-bond acceptors for a molecule</td></tr><tr><td>calcnumlipinskihbd</td><td>Returns the number of Lipinski H-bond donors for a molecule</td></tr><tr><td>calcnumrings</td><td>Returns the number of rings for a molecule</td></tr><tr><td>calcnumrotatablebonds</td><td>Returns the number of rotatable bonds for a molecule. strict = NumRotatable-BondsOptions.NonStrict - Simple rotatable bond definition.</td></tr><tr><td>calcnumsaturatedcarbocycles</td><td>Returns the number of saturated carbocycles for a molecule</td></tr><tr><td>calcnumsaturatedheterocycles</td><td>Returns the number of saturated heterocycles for a molecule</td></tr><tr><td>calcnumsaturatedrings</td><td>Returns the number of saturated rings for a molecule</td></tr><tr><td>calcnumunspecifiedatomstereocenters</td><td>Returns the number of unspecified atomic stereocenters</td></tr><tr><td>calcooxidationnumbers</td><td>Adds the oxidation number/state to the atoms of a molecule as property OxidationNumber on each atom. Use Pauling electronegativities. This is experimental code, still under development.</td></tr></table>
+
+Table 11: List of Structural Descriptors tools
+
+# B Prompts
+
+# B.1 Prompt for analyst
+
+You are a professional AI chemistry assistant specialized in resolving [category name] using RDKit tools.
+
+Your job is to identify how to retrieve standardized molecular information such as CIDs, InChI, and canonical SMILES for downstream processing.
+
+Follow this structured reasoning process step-by-step:
+
+Step 1. Analyze the molecule design condition which is the goal of the task.  
+Step 2. Parse list of all valid SMILES strings mentioned anywhere in the user prompt and output them in the provided JSON format.  
+Step 3. Based on your chemical knowledge, explain why standardizing identifiers and resolving canonical formats might be important for this task.
+
+- E.g., checking uniqueness, linking to external data, verifying molecular identity.  
+Step 4. Choose as many tools as necessary from the identifier toolset that help you access consistent molecular representations or external references.  
+Step 5. Output your final answer in the provided JSON format.
+
+This is a molecule design condition of the [task name] task: [task description]
+
+Now output the tools to use by using the following JSON format. Take a deep breath and think carefully before writing your answer. "json {}
+
+```txt
+"parseed_smiles": [  
+{  
+"smiles": "Parsing SMILES string",  
+} ],  
+...  
+],  
+"tools_to_use": [  
+{  
+"tool_name": "fr_Ar_OH",  
+"purpose": "Detect aromatic hydroxyl groups, similar to those in albuterol."  
+} ],  
+...  
+]  
+}
+```
+
+# B.2 Prompt for scientist
+
+You are a skilled chemist.
+
+Your task is to design a SMILES string for a molecule that satisfies the following condition: [task description]
+
+Functional groups and molecule tool analysis results of task related molecules: [result of tool analysis]
+
+You are provided with:
+
+- Top-100 example molecules with high relevance to the task, listed below. You may use these as inspiration, but YOU MUST NOT COPY THEM EXACTLY.  
+- A list of previously generated SMILES, which YOU MUST NOT REPEAT.
+
+Top-100 Relevant SMILES Examples (SMILES, score)
+
+YOU MUST FAITHFULLY REFER TO THESE EXAMPLES WHEN DESIGNING YOUR MOLECULE. BUT DO NOT COPY THEM EXACTLY:
+
+[top100 SMLIES]
+
+You must return your response in the following json format. The text inside each key explains what kind of answer is expected — it is a guideline, not the answer.
+
+DO NOT repeat the example text or instructions. Instead, write your own scientifically reasoned content based on the task.
+
+Use the following format. Take a deep breath and think carefully before writing your answer.
+
+```
+```
+```
+```
+```
+``
+
+"step1": "List of the target's critical structural/property features (e.g., 'Target: phenyl ring,  $\beta$ -hydroxyamine, catechol-like substitution'). If property-based, specify requirements (e.g., "logP  $>3$ : add hydrophobic groups").",
+
+"step2": "Propose modifications or scaffolds to meet the condition (e.g., 'Replace catechol with 3-hydroxy-4-pyridone').
+
+Justify each change chemically (e.g., "Maintains H-bonding but improves metabolic stability").", "step3": "Describe the full structure of your designed molecule in natural language before writing the SMILES. (e.g., "A tert-butyl group attached to the amine  $(-\mathrm{NH}-\mathrm{C}(\mathrm{CH}_3)_3)$  to mimic target's bulky substituent.")".
+
+"smiles": "Your valid SMILES string here"  
+}
+
+# B.3 Prompts for scientist with feedback
+
+YOU MUST NOT REPEAT ANY OF THE PREVIOUSLY GENERATED SMILES: [smiles_history] Task: Take [verification/reviewer]'s feedback actively and design a SMILES string for a molecule that satisfies the condition:
+
+Condition for molecule design: [task description]
+
+Functional groups and molecule tool analysis results of task related molecules: [target functional groups]
+
+Top-100 Relevant SMILES Examples (SMILES, score)  
+YOU MUST FAITHFULLY REFER TO THESE EXAMPLES WHEN DESIGNING YOUR MOLECULE. BUT DO NOT COPY THEM EXACTLY: [topk smiles]
+
+You will be provided with:
+
+1. Previous SMILES string  
+2. Task score (0-1)  
+3. Detected functional groups in your previous molecule
+
+MOLECULE SMILES TO IMPROVE
+
+MOLECULE SMILES: [previous smiles]
+
+- Task score: [score] (0-1)  
+- Functional groups detected:
+
+[functional groups]
+
+— YOUR PREVIOUS THOUGHT AND REVIEWER'S FEEDBACK —
+
+Step1: List Key Features
+
+Your previous thought process:
+
+[scientist step1 reasoning]
+
+Accordingly, reviewer's feedback is:
+
+[verifier/reviewer step1 feedback]
+
+Step2: Design Strategy:
+
+Your previous thought process:
+
+[scientist step2 think]
+
+Accordingly, reviewer's feedback is:
+
+[verifier/reviewer step2 feedback]
+
+Step 3: Construct the Molecule: Your previous thought process:
+
+[verifier/scientist step3 think]
+
+Accordingly, reviewer's feedback is:
+
+[verifier/reviewer step3 feedback]
+
+Now based on your previous thoughts and the reviewer's feedback, you need to improve your design. You must return your response in the following json format.
+
+The text inside each key explains what kind of answer is expected — it is a guideline, not the answer.
+
+DO NOT repeat the example text or instructions.
+
+Instead, write your own scientifically reasoned content based on the task.
+
+Use the following format.
+
+Take a deep breath and think carefully before writing your answer.
+
+```snap
+```
+```
+```
+```
+``
+```
+
+"step1": "List of the target's critical structural/property features (e.g., 'Target: phenyl ring,  $\beta$ -hydroxyamine, catechol-like substitution'). If property-based, specify requirements (e.g., "logP  $>3$ : add hydrophobic groups").",
+
+"step2": "Propose modifications or scaffolds to meet the condition (e.g., 'Replace catechol with 3-hydroxy-4-pyridone'). Justify each change chemically (e.g., "Maintains H-bonding but improves metabolic stability").",
+
+"step3": "Describe the full structure of your designed molecule in natural language before writing the SMILES. (e.g., "A tert-butyl group attached to the amine  $(-NH - C(CH_3)_3)$  to mimic target's bulky substituent.")".
+
+"smiles": "Your valid SMILES string here"  
+}  
+...
+
+# B.4 Prompts for verifier
+
+You are a meticulous double-checker LLM. Your task is to verify whether each step of the scientist's reasoning is chemically valid and faithfully and logically reflected in the final SMILES string. You will be given:
+
+- A user prompt describing the target objective,  
+- The scientist's reasoning broken into Step1 through Step3,  
+- The SMILES string proposed by the scientist.
+
+Evaluate each step independently, comparing the described logic to the molecular structure in the SMILES.
+
+Provide a reasoning assessment for each step.  $== = = =$  SCIENTIST'S TASK  $= = = =$
+
+If any step is inconsistent, mark "Consistency" as "Inconsistent" and provide specific suggestions for improvement.
+
+[task description]
+
+Functional groups and molecule tool analysis results of task related molecules: [target functional groups]
+
+```javascript
+$= = =$  SCIENTIST'S THINKING  $= = =$  Step1:[thinking['step1']] Step2:[thinking['step2']] Step3:[thinking['step3']]
+```
+
+```javascript
+=== SCIENTIST'S SMILES===
+```
+
+- SMILES: [smiles]  
+- Detected functional groups and molecule tool analysis results:
+
+[functional groups]
+
+You must return your response in the following json format.
+
+The text inside each key explains what kind of answer is expected — it is a guideline, not the answer.
+
+DO NOT repeat the example text or instructions.
+
+Instead, write your own scientifically reasoned content based on the task.
+
+Use the following format.
+
+Take a deep breath and think carefully before writing your answer.
+
+```snap
+```
+```
+```
+```
+``
+```
+
+"step1": "Your analysis of whether scientist's Step1 thinking is chemically valid and reflected in the SMILES.",
+
+"step2": "Your analysis of whether scientist's Step2 thinking is chemically valid and reflected in the SMILES.",
+
+"step3": "Your analysis of whether scientist's Step3 thinking is chemically valid and reflected in the SMILES.",
+
+```txt
+"consistency": "Consistent" or "Inconsistent",  
+}  
+...
+```
+
+# B.5 Prompts for reviewer
+
+You are a rigorous chemistry reviewer.
+
+Evaluate the Scientist LLM's reasoning steps and final SMILES molecule for:
+
+- Validity  
+- Chemical soundness  
+- Adherence to the design condition:
+
+Scientist LLM's task:
+
+[task description]
+
+Be constructive: Provide fixes for issues (e.g., "Replace  $\mathrm{C} = \mathrm{O} = \mathrm{C}$  with  $\mathrm{O} = \mathrm{C} = \mathrm{O}$  for carbon dioxide").
+
+You are provided with:
+
+- Scientist's thinking  
+- Scientist-generated SMILES  
+- Task score  
+- Detected functional groups in the generated molecule
+
+SCIENTIST'S STEP-WISE THINKING
+
+Step 1: [scientist step1 reasoning]
+
+Step 2: [scientist step2 reasoning]
+
+Step 3: [scientist step3 reasoning]
+
+SCIENTIST-MOLECULE SMILES
+
+SMILES: [scientist proposed SMILES]
+
+- Task score: [score] (range: 0 to 1)  
+- Detected functional groups and molecule tool analysis results:
+
+[functional groups]
+
+You must return your response in the following json format.
+
+The text inside each key explains what kind of answer is expected — it is a guideline, not the answer.
+
+DO NOT repeat the example text or instructions.
+
+Instead, write your own scientifically reasoned content based on the task.
+
+Use the following format.
+
+Take a deep breath and think carefully before writing your answer.
+
+"json
+
+1
+
+"step1": "List accurate features and functional groups identified. Mention any critical features and functional groups that were missed or misinterpreted.",
+
+"step2": "Evaluate if the proposed design strategy aligns with the structural and functional similarity goal.
+
+Comment on whether the design aligns with the initial objectives. Suggest improvements or alternatives if needed."
+
+6
+
+"step3": "Review the structural construction and positional assignments. Check for missing elements or mismatches in reasoning. (e.g., "Claimed 'para hydroxyl' but SMILES places it meta)", }}
+
+# C Task description
+
+In this section, we describe the 23 tasks of practical molecular benchmark (Gao et al., 2022). For more details about the task and oracle, refer to Therapeutics Data Commons (Huang et al., 2021, TDC) document: https://tdc.readthedocs.io/en/main/_modules/tdc/chem_utils/oracle/oracle.html.
+
+# 1. albuterol_similarity
+
+Design a molecule similar to albuterol while preserving key functional groups.
+
+# 2.amlodipine_mpo
+
+Generate molecules similar to amlodipine with good drug-like properties (e.g., 3-ring topology).
+
+# 3. celecoxib_rediscovery
+
+Recreate the anti-inflammatory drug celecoxib.
+
+# 4.deco_hop
+
+Modify the decorations of a molecule while preserving a fixed scaffold. Avoid forbidden substructures and stay below similarity cap.
+
+# 5. drd2
+
+Generate molecules predicted to strongly bind to the dopamine D2 receptor using a predictive model.
+
+# 6. fexofenadine_mpo
+
+Create molecules structurally similar to fexofenadine with TPSA  $\approx 90$  and  $\log \mathrm{P}\approx 4$
+
+# 7. gsk3b
+
+Design molecules predicted to have high binding affinity for the  $\mathrm{GSK3}\beta$  protein.
+
+# 8.isomers_c7h8n2o2
+
+Generate any molecule that is an exact isomer of  $\mathrm{C_7H_8N_2O_2}$ . Must match the molecular formula exactly.
+
+# 9.isomers_c9h10n2o2pf2cl
+
+Generate an exact isomer of  $\mathrm{C_9H_{10}N_2O_2PF_2Cl}$
+
+# 10.jnk3
+
+Design molecules with high predicted inhibitory activity against the JNK3 protein.
+
+# 11. median1
+
+Find a molecule similar to both camphor and menthol.
+
+# 12. median2
+
+Design a molecule similar to both tadalafil and sildenafil.
+
+# 13. mestranol_similarity
+
+Generate molecules similar to the hormone mestranol, preserving the core scaffold.
+
+# 14. osimertinib_mpo
+
+Create osimertinib-like molecules with low logP  $(\approx 1)$  and TPSA  $\approx 100$
+
+# 15. perindopril_mpo
+
+Design perindopril-like molecules.
+
+# 16. qed
+
+Maximize a quantitative estimate of drug-likeness (QED) score.
+
+# 17. ranolazine_mpo
+
+Create ranolazine-like molecules with TPSA  $\approx$  95 and  $\log \mathrm{P}\approx 7$
+
+# 18. scaffold_hop
+
+Replace the molecular scaffold while keeping key functional groups unchanged.
+
+# 19. sitagliptin_mpo
+
+Design sitagliptin-like molecules matching formula  $\mathrm{C_{16}H_{15}F_6N_5O}$
+
+# 20. thiothixene_rediscovery
+
+Reproduce the structure of thiothixene.
+
+# 21. troglitazone_rediscovery
+
+Reconstruct the diabetes drug troglitazone.
+
+# 22. valsartan_smarts
+
+Generate molecules containing the substructure SMARTS with  $\log \mathrm{P}\approx 2.0$  and TPSA  $\approx 95$
+
+# 23. zaleplon_mpo
+
+Design zaleplon-like molecules with formula  $\mathrm{C_{19}H_{17}N_3O_2}$
+
+# D PMO-1K experiment result
+
+We provide the full PMO-1K experiment result in Table 12.
+
+# E ZINC 250K statistics
+
+We provide the data statistics of ZINC250K (Sterling and Irwin, 2015) that we used in our setting at Table 13.
+
+# F Usage of AI assistants
+
+We used AI writing assistants (e.g., ChatGPT) to improve the clarity, grammar, and style of the manuscript during the writing process. These tools were employed strictly for language refinement and did not contribute to the development of ideas, methods, or analysis. All scientific contributions and experimental results are the original work of the authors.
+
+# G Scientific Artifacts
+
+The License for artifacts. We used dataset and tools accordingly with their respective licenses. In detail, We use open-source ZINC250K dataset (Sterling and Irwin, 2015) and the publicly available RDKIt tools (Landrum, 2013). We provide our source code at https://anonymous.4open.science/r/mt_mol-0448 for reproducibility with an appropriate open-source license.
+
+Artifact use consistency with intended use. We used dataset and tools in line of their intended use. Specifically, ZINC250K (Sterling and Irwin, 2015) incorporates molecule with property scores for molecular optimization task which aligns with goal of our study. Also, RDKit tools are used to analyze the chemical properties of the given molecule which is used in our study.
+
+<table><tr><td>Task</td><td>GP BO</td><td>REINVENT</td><td>LICO-L</td><td>Genetic GFN</td><td>Graph GA</td><td>Aug. Mem.</td><td>MOLLEO-B</td><td>MOLLEO-D*</td><td>Ours-D*</td></tr><tr><td>albuterol_similarity</td><td>0.636 ± 0.106</td><td>0.496 ± 0.020</td><td>0.656 ± 0.125</td><td>0.664 ± 0.054</td><td>0.583 ± 0.065</td><td>0.557 ± 0.048</td><td>0.886 ± 0.023</td><td>0.883 ± 0.001</td><td>0.998 ± 0.000</td></tr><tr><td>amlodipine_mpo</td><td>0.519 ± 0.014</td><td>0.472 ± 0.008</td><td>0.541 ± 0.026</td><td>0.534 ± 0.019</td><td>0.501 ± 0.016</td><td>0.489 ± 0.009</td><td>0.637 ± 0.023</td><td>0.540 ± 0.072</td><td>0.647 ± 0.010</td></tr><tr><td>celecoxib_rediscovery</td><td>0.411 ± 0.046</td><td>0.370 ± 0.029</td><td>0.447 ± 0.073</td><td>0.447 ± 0.028</td><td>0.424 ± 0.049</td><td>0.385 ± 0.027</td><td>0.402 ± 0.003</td><td>0.512 ± 0.119</td><td>0.867 ± 0.007</td></tr><tr><td>deco_hop</td><td>0.593 ± 0.018</td><td>0.572 ± 0.006</td><td>0.596 ± 0.010</td><td>0.604 ± 0.017</td><td>0.581 ± 0.006</td><td>0.579 ± 0.010</td><td>0.588 ± 0.007</td><td>0.574 ± 0.001</td><td>0.842 ± 0.077</td></tr><tr><td>drd2</td><td>0.857 ± 0.080</td><td>0.775 ± 0.086</td><td>0.859 ± 0.066</td><td>0.809 ± 0.045</td><td>0.833 ± 0.065</td><td>0.795 ± 0.024</td><td>0.910 ± 0.017</td><td>0.812 ± 0.027</td><td>0.756 ± 0.410</td></tr><tr><td>fexofenadine_mpo</td><td>0.707 ± 0.021</td><td>0.650 ± 0.007</td><td>0.700 ± 0.023</td><td>0.682 ± 0.021</td><td>0.666 ± 0.009</td><td>0.679 ± 0.021</td><td>0.674 ± 0.002</td><td>0.680 ± 0.007</td><td>0.883 ± 0.02</td></tr><tr><td>gsk3b</td><td>0.611 ± 0.059</td><td>0.589 ± 0.063</td><td>0.617 ± 0.063</td><td>0.637 ± 0.018</td><td>0.523 ± 0.047</td><td>0.539 ± 0.097</td><td>0.397 ± 0.013</td><td>0.496 ± 0.073</td><td>0.308 ± 0.009</td></tr><tr><td>isomers_c7h8n2o2</td><td>0.545 ± 0.158</td><td>0.725 ± 0.064</td><td>0.779 ± 0.099</td><td>0.738 ± 0.039</td><td>0.735 ± 0.112</td><td>0.661 ± 0.039</td><td>0.737 ± 0.043</td><td>0.850 ± 0.009</td><td>0.986 ± 0.015</td></tr><tr><td>isomers_c9h10n2o2pf2c1</td><td>0.599 ± 0.059</td><td>0.630 ± 0.032</td><td>0.672 ± 0.075</td><td>0.656 ± 0.075</td><td>0.630 ± 0.086</td><td>0.596 ± 0.066</td><td>0.635 ± 0.017</td><td>0.832 ± 0.007</td><td>0.914 ± 0.031</td></tr><tr><td>jnk3</td><td>0.346 ± 0.067</td><td>0.315 ± 0.042</td><td>0.336 ± 0.051</td><td>0.409 ± 0.165</td><td>0.301 ± 0.071</td><td>0.294 ± 0.110</td><td>0.186 ± 0.076</td><td>0.342 ± 0.044</td><td>0.125 ± 0.020</td></tr><tr><td>median1</td><td>0.213 ± 0.020</td><td>0.205 ± 0.014</td><td>0.217 ± 0.019</td><td>0.219 ± 0.008</td><td>0.208 ± 0.015</td><td>0.219 ± 0.014</td><td>0.236 ± 0.021</td><td>0.193 ± 0.005</td><td>0.321 ± 0.029</td></tr><tr><td>median2</td><td>0.203 ± 0.009</td><td>0.188 ± 0.010</td><td>0.193 ± 0.009</td><td>0.204 ± 0.011</td><td>0.181 ± 0.009</td><td>0.184 ± 0.010</td><td>0.191 ± 0.009</td><td>0.197 ± 0.023</td><td>0.322 ± 0.024</td></tr><tr><td>mestranol_similarity</td><td>0.427 ± 0.025</td><td>0.379 ± 0.026</td><td>0.423 ± 0.016</td><td>0.414 ± 0.022</td><td>0.362 ± 0.017</td><td>0.393 ± 0.021</td><td>0.399 ± 0.020</td><td>0.630 ± 0.171</td><td>0.996 ± 0.001</td></tr><tr><td>osimertinib_mpo</td><td>0.766 ± 0.006</td><td>0.737 ± 0.007</td><td>0.759 ± 0.008</td><td>0.763 ± 0.008</td><td>0.751 ± 0.005</td><td>0.761 ± 0.006</td><td>0.779 ± 0.006</td><td>0.753 ± 0.018</td><td>0.796 ± 0.005</td></tr><tr><td>perindopril_mpo</td><td>0.458 ± 0.019</td><td>0.404 ± 0.008</td><td>0.473 ± 0.009</td><td>0.462 ± 0.033</td><td>0.435 ± 0.016</td><td>0.422 ± 0.013</td><td>0.655 ± 0.054</td><td>0.422 ± 0.006</td><td>0.542 ± 0.027</td></tr><tr><td>qed</td><td>0.912 ± 0.010</td><td>0.921 ± 0.002</td><td>0.925 ± 0.005</td><td>0.928 ± 0.002</td><td>0.914 ± 0.007</td><td>0.923 ± 0.002</td><td>0.919 ± 0.006</td><td>0.928 ± 0.006</td><td>0.903 ± 0.003</td></tr><tr><td>ranolazine_mpo</td><td>0.701 ± 0.023</td><td>0.574 ± 0.044</td><td>0.687 ± 0.029</td><td>0.623 ± 0.022</td><td>0.620 ± 0.014</td><td>0.614 ± 0.033</td><td>0.640 ± 0.000</td><td>0.516 ± 0.024</td><td>0.233 ± 0.018</td></tr><tr><td>scaffold_hop</td><td>0.478 ± 0.009</td><td>0.447 ± 0.010</td><td>0.480 ± 0.008</td><td>0.485 ± 0.015</td><td>0.461 ± 0.008</td><td>0.460 ± 0.010</td><td>0.473 ± 0.000</td><td>0.464 ± 0.002</td><td>0.646 ± 0.055</td></tr><tr><td>siglatigpin_mpo</td><td>0.232 ± 0.083</td><td>0.261 ± 0.026</td><td>0.315 ± 0.097</td><td>0.227 ± 0.041</td><td>0.229 ± 0.053</td><td>0.245 ± 0.030</td><td>0.193 ± 0.073</td><td>0.328 ± 0.091</td><td>0.067 ± 0.006</td></tr><tr><td>thiothixene_rediscovery</td><td>0.351 ± 0.033</td><td>0.311 ± 0.021</td><td>0.343 ± 0.035</td><td>0.377 ± 0.015</td><td>0.322 ± 0.023</td><td>0.336 ± 0.073</td><td>0.416 ± 0.075</td><td>0.478 ± 0.028</td><td>0.719 ± 0.001</td></tr><tr><td>troglitazone_rediscovery</td><td>0.313 ± 0.018</td><td>0.246 ± 0.009</td><td>0.292 ± 0.028</td><td>0.277 ± 0.015</td><td>0.267 ± 0.015</td><td>0.262 ± 0.012</td><td>0.302 ± 0.022</td><td>0.387 ± 0.013</td><td>0.841 ± 0.042</td></tr><tr><td>valsartanSMARTs</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 0.000</td><td>0.000 ± 1.000</td></tr><tr><td>zaleplon_mpo</td><td>0.392 ± 0.034</td><td>0.406 ± 0.017</td><td>0.404 ± 0.022</td><td>0.400 ± 0.014</td><td>0.374 ± 0.024</td><td>0.415 ± 0.013</td><td>0.392 ± 0.003</td><td>0.409 ± 0.005</td><td>0.625 ± 0.046</td></tr><tr><td>Sum of scores (↑)</td><td>11.27</td><td>10.68</td><td>11.71</td><td>11.56</td><td>10.90</td><td>10.81</td><td>11.65</td><td>12.23</td><td>15.42</td></tr></table>
+
+Table 12: Detailed results of PMO-1K benchmark. Tasks are assessed using AUC top-10 with mean ± standard deviation. Results with (*) are evaluated from 3 independent runs while the others are assessed from 5 independent runs. We mark the best result in bold and the second-best are underlined for each task.  
+
+<table><tr><td>Oracle</td><td>Min</td><td>Max</td><td>Mean</td><td>Std</td></tr><tr><td>albuterol_similarity</td><td>0.053</td><td>0.667</td><td>0.251</td><td>0.062</td></tr><tr><td>amlodipine_mpo</td><td>0.000</td><td>0.686</td><td>0.214</td><td>0.144</td></tr><tr><td>celecoxib_rediscovery</td><td>0.000</td><td>0.447</td><td>0.142</td><td>0.060</td></tr><tr><td>deco_hop</td><td>0.291</td><td>0.878</td><td>0.768</td><td>0.048</td></tr><tr><td>drd2</td><td>0.000</td><td>0.987</td><td>0.009</td><td>0.038</td></tr><tr><td>fexofenadine_mpo</td><td>0.000</td><td>0.756</td><td>0.232</td><td>0.206</td></tr><tr><td>gsk3b</td><td>0.000</td><td>0.990</td><td>0.030</td><td>0.045</td></tr><tr><td>isomers_c7h8n2o2</td><td>0.000</td><td>1.000</td><td>0.004</td><td>0.037</td></tr><tr><td>isomers_c9h10n2o2pf2c1</td><td>0.000</td><td>0.869</td><td>0.018</td><td>0.071</td></tr><tr><td>jnk3</td><td>0.000</td><td>0.680</td><td>0.016</td><td>0.026</td></tr><tr><td>median1</td><td>0.000</td><td>0.324</td><td>0.066</td><td>0.037</td></tr><tr><td>median2</td><td>0.000</td><td>0.291</td><td>0.108</td><td>0.027</td></tr><tr><td>mestranol_similarity</td><td>0.004</td><td>0.886</td><td>0.170</td><td>0.059</td></tr><tr><td>osimertinib_mpo</td><td>0.000</td><td>0.829</td><td>0.179</td><td>0.209</td></tr><tr><td>perindopril_mpo</td><td>0.000</td><td>0.560</td><td>0.176</td><td>0.113</td></tr><tr><td>qed</td><td>0.117</td><td>0.948</td><td>0.732</td><td>0.139</td></tr><tr><td>ranolazine_mpo</td><td>0.000</td><td>0.586</td><td>0.059</td><td>0.069</td></tr><tr><td>scaffold_hop</td><td>0.176</td><td>0.526</td><td>0.373</td><td>0.026</td></tr><tr><td>sitagliptin_mpo</td><td>0.000</td><td>0.479</td><td>0.012</td><td>0.035</td></tr><tr><td>thiothixene_rediscovery</td><td>0.000</td><td>0.408</td><td>0.162</td><td>0.047</td></tr><tr><td>troglitazon_rediscovery</td><td>0.000</td><td>0.391</td><td>0.135</td><td>0.035</td></tr><tr><td>valsartan_smarts</td><td>0.000</td><td>0.320</td><td>0.000</td><td>0.001</td></tr><tr><td>zaleplon_mpo</td><td>0.000</td><td>0.545</td><td>0.072</td><td>0.100</td></tr></table>
+
+Table 13: Data statistics of ZINC 250k that we retrieved for each oracle.

EMNLP/2025/MT-Mol_ Multi Agent System with Tool-based Reasoning for Molecular Optimization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0712716ba27eac5e13851eb45eb8a3b30c91c3bd74b1046bab5c54a06ced639
+size 2449949

EMNLP/2025/MT-Mol_ Multi Agent System with Tool-based Reasoning for Molecular Optimization/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c2eacf050babe10946c3a571fbb1aac2988cc2e43245d4f597da8bfb5ddf596
+size 751847

EMNLP/2025/MT-R1-Zero_ Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning/18e1861f-b64d-4027-ae8f-74cad0b42af2_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e2ef9675193f655687c1b66826d58dbefc026f14be8fd75827a4d947315d0aa
+size 143363

EMNLP/2025/MT-R1-Zero_ Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning/18e1861f-b64d-4027-ae8f-74cad0b42af2_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:644c0a1c482ae39cedb26d180ebcb388566dd3078c4eabebf3b174e76709cd2b
+size 164121

EMNLP/2025/MT-R1-Zero_ Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning/18e1861f-b64d-4027-ae8f-74cad0b42af2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b5c4f62ddad4fe23250fc8da221e7095ba703a734a28e109dcc42cde87b3204
+size 2235833

EMNLP/2025/MT-R1-Zero_ Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning/full.md

@@ -0,0 +1,617 @@
+# MT-R1-Zero: Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning
+
+Zhaopeng Feng $^{1}$  Shaosheng Cao $^{2\dagger}$  Jiahan Ren $^{1}$  Jiayuan Su $^{1}$
+
+Ruizhe Chen $^{1}$  Yan Zhang $^{1}$  Jian Wu $^{1}$  Zuozhu Liu $^{1\dagger}$
+
+$^{1}$ Zhejiang University  $^{2}$ Xiaohongshu Inc.
+
+{zhaopeng.23, zuozhuliu}@intl.zju.edu.cn
+
+{caoshaosheng}@xiaohongshu.com
+
+# Abstract
+
+Large-scale reinforcement learning (RL) methods have proven highly effective in enhancing the reasoning abilities of large language models (LLMs), particularly for tasks with verifiable solutions such as mathematics and coding. However, applying this idea to machine translation (MT), where outputs are flexibly formatted and difficult to automatically evaluate with explicit rules, remains underexplored. In this work, we introduce MT-R1-Zero, the first open-source adaptation of the R1-Zero RL framework for MT without supervised fine-tuning or cold-start. We propose a rule-metric mixed reward mechanism to guide LLMs towards improved translation quality via emergent reasoning. On the WMT 24 English-Chinese benchmark, our MT-R1-Zero3B-Mix achieves competitive performance, surpassing TowerInstruct-7B-v0.2 by an average of 1.26 points. Meanwhile, our MT-R1-Zero7B-Mix attains a high average score of 62.25 across all metrics, placing it on par with advanced proprietary models such as GPT-4o and Claude-3.5-Sonnet, while the MT-R1-Zero7B-Sem variant achieves state-of-the-art scores on semantic metrics. Moreover, our work exhibits strong generalization capabilities on out-of-distribution MT tasks, robustly supporting multilingual and low-resource settings. Extensive analysis of model behavior across different initializations and reward metrics offers pioneering insight into the critical role of reward design, LLM adaptability, training dynamics, and emergent reasoning patterns within the R1-Zero paradigm for MT. Our code is available at https://github.com/fzp0424/MT-R1-Zero.
+
+# 1 Introduction
+
+Large-scale Reinforcement Learning (RL) has empowered Large Language Models (LLMs) with strong reasoning capabilities (OpenAI, 2024; Team,
+
+![](images/581cc4ff2092b3f56ea8df32589c89e309da03a5c9dd07f5516d26707033b708.jpg)  
+Figure 1: Performance comparison of contemporary LLM-based translation systems on the WMT 24 EN-ZH test set, plotted by average score across BLEU, COMETKiwi, and XCOMET versus model size.
+
+2025a,b), demonstrating significant success in tasks such as mathematical reasoning or coding in which answers can be clearly verified. In particular, DeepSeek-R1-Zero (DeepSeek-AI et al., 2025) introduced a pure rule-based RL approach that directly fosters emergent reasoning ability without requirements on structured Chain-of-Thought (CoT) data (Huang et al., 2025; Cui et al., 2025) or sophisticated techniques such as Monte Carlo Tree Search (MCTS) (Silver et al., 2016; Zhao et al., 2024; Luo et al., 2024; Guan et al., 2025). However, the applicability of these methods to machine translation (MT) remains challenging and underexplored, as MT outputs are flexibly generated and hard to evaluate automatically with explicit rules.
+
+Recent work has launched attempts to empower LLMs for MT with reasoning capabilities (Chen et al., 2025; Liu et al., 2025). Early studies investigate explicit reasoning methods for improved translation, such as finetuning with CoT (Wang et al., 2024a) or MCTS (Zhao et al., 2024), where advanced multi-step pipelines with self-correction or long-thought agentic mechanisms are further ex
+
+plored (Feng et al., 2024b; Wang et al., 2024b,a). Another line of work leverages RL to empower LLMs for MT through process reward models or supervised finetuning (SFT) with manually annotated CoT data (Feng et al., 2025; He et al., 2025). However, these methods often depend on manually designed or synthetically generated structured CoT data, rely on complex search algorithms, or require explicit multi-stage prompting, leaving the potential of pure RL-based approaches largely unexplored. Furthermore, the performance reported in these studies often lags behind state-of-the-art (SoTA) open-source or proprietary models.
+
+Developing pure RL methods to directly enhance the reasoning ability of LLMs for better translation requires answering three key questions: 1) Feasibility: How to design R1-Zero-like RL pipelines with effective reward signals to directly solve MT tasks without binary rule-based rewards; 2) Reasoning capability: Could pure RL training cultivate emergent reasoning abilities and induce models to generate explicit thinking patterns for MT, such as multi-step CoT or verification/reflection; 3) Generalizability: Could the training paradigm generalize across different models (e.g., pre-trained base models, instruction-tuned models, or models pretrained on translation data) or diverse downstream settings (e.g., out-of-distribution, multilingual or low-resource scenarios).
+
+In this work, we introduce MT-R1-Zero, the first open-source implementation that extends the R1-Zero-like RL training paradigm to MT. We propose a rule-metric mixed reward mechanism that adapts the original rule-based reward concept to effectively guide training in MT scenarios. We explore different rewards optimizing over lexical (Lex), semantic (Sem), and Lex-Sem mixed (Mix) objectives to guide LLMs towards improved translation quality via emergent reasoning. Our experiments demonstrate the efficacy of this approach: as RL training progresses, our MT-R1-Zero-3B-Mix achieves competitive performance, surpassing TowerInstruct-7B-v0.2 by an average of 1.26 points across all metrics (BLEU, COMETKiwi, XCOMET) on the WMT 24 English-Chinese (EN-ZH) benchmark. Meanwhile, our MT-R1-Zero-7B-Mix surpasses LLaMA-3.1-70B by an average of 1.24 points and Qwen2.5-72B by 0.48 points, even on par with top proprietary models such as GPT4o and Claude-3.5-Sonnet. The MT-R1-Zero further demonstrates promising generalizability across multilingual and low-resource settings.
+
+Extensive experiments further provide key findings and insight into the adaptation of R1-Zero paradigm to MT. First, we empirically demonstrate that the choice of metric reward plays a pivotal role in steering RL optimization and translation style (semantic or lexical) (Finding 1). Further analysis reveals that MT-R1-Zero induces diverse emergent reasoning patterns, including dynamic language-of-thought transition during translation (Finding 2). We also identify distinct RL adaptability of different LLMs (Finding 3). Ablation studies suggest that pure RL process alone can lead to substantial translation improvements, independent of thinking morbidity. Our core contributions are as follows:
+
+- We present the first open-source implementation of the DeepSeek-R1-Zero paradigm for MT, achieving superior performance across in-domain and out-of-distribution MT tasks.  
+- Our analysis reveals key findings and recipes for effective R1-Zero adaptation to MT, including reward metric selection, emergent reasoning patterns, training dynamics and LLM adaptability.  
+- Extensive experiments and ablations show that pure RL serves as the primary driver of MT improvements, with minimal dependence on forced reasoning or output length, highlighting the significant potential of RL for diverse translation applications and broader language tasks.
+
+# 2 Related Work
+
+LLM Reasoning with Post-training. Recent research indicates that scaling test-time computation can significantly enhance the ability of LLMs to tackle complex reasoning tasks (OpenAI, 2024; Zeng et al., 2024; Xiang et al., 2025). Many approaches rely on sophisticated techniques such as step-level process reward models (PRMs) that provide granular feedback (Lightman et al., 2024; Yuan et al., 2024; Snell et al., 2024) or MCTS to explore potential reasoning paths (Feng et al., 2023; Qi et al., 2024; Guan et al., 2025). A recent alternative, DeepSeek-R1-Zero (DeepSeek-AI et al., 2025), demonstrated that large-scale pure RL, guided only by formatting rules and correctness of final predictions (rule-based reward), can motivate LLMs to develop self-emergent reasoning processes for complex reasoning tasks. This paradigm has been successfully replicated and extended to mathematical, logical, and visual reasoning (Hu et al., 2025; Face, 2025; Xie et al., 2025; Huang
+
+et al., 2025). Despite its potential, the application of the R1-Zero RL paradigm to challenging generation tasks like MT, in which the accuracy/quality of outputs is not rule-based and difficult to validate automatically, remains an open question.
+
+LLM Reasoning for MT. Leveraging reasoning to improve MT has garnered increasing attention, as explored by Chen et al. (2025); Liu et al. (2025). Previous work have designed multi-step processes for MT, e.g., Feng et al. (2024b) introduced an API-based self-correcting framework, and Wang et al. (2024b) employed multi-task training followed by a multistage inference phase. Wang et al. (2024a) integrated a similar procedure into inference-time CoT, using a multi-agent mechanism to synthesize long CoT prompts for English-Chinese literary translation. Efforts have also focused on reward modeling for MT reasoning. Feng et al. (2025) constructed implicit process reward models for translation and explored their effectiveness when combined with test-time search. Recent study further evaluated explicit reasoning for MT using CoT fine-tuning and MCTS to expand test-time computation (Zhao et al., 2024). He et al. (2025) demonstrated that models can acquire reasoning-based translation capabilities through multi-stage training with manually constructed CoT templates.
+
+# 3 Method
+
+In this section, we present our method that trains a translation model with pure RL using a hybrid reward system. Unlike tasks with fixed correct answers, translation allows for multiple valid outputs, making the evaluation more complicated. In this work, we introduce a rule-metric mixed reward that integrates reasoning format checking with multiple translation quality assessment metrics, which is used within the Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm to ensure stable and efficient RL training.
+
+# 3.1 Rule-Metric Mixed Reward
+
+The reward signal  $r$  is crucial in RL. DeepSeek-R1-Zero (DeepSeek-AI et al., 2025) employs simple rule-based rewards that check whether the answer is correct and whether the response follows a specific format. This works well for tasks with fixed format correct answers such as math or coding. However, there is often no single "correct" output for MT, impeding the design of rule-based rewards. Fortunately, the MT community has developed
+
+many evaluation metrics to measure translation quality. Recent advancements in automated MT evaluation metrics have shown promise in aligning automated assessments with human translation quality judgments (Freitag et al., 2023). Thus, we design a rule-metric mixed reward, which consists of two parts: a Format Reward that checks output structure, and a Metric Reward that evaluates translation quality. We use a structured prompt template similar to that in DeepSeek-R1-Zero:
+
+# Template for MT-R1-Zero
+
+A conversation between User and Assistant. The User asks for a translation from {src_language} to {tgt_language}, and the Assistant solves it. The Assistant first thinks about the reasoning process in the mind and then provides the user with the final translation. The reasoning process and final translation are enclosed within <think> </think> and <translate> </translate> tags, respectively, i.e., <think> reasoning process here </think><translate> final translation here </translate>.
+
+User:{src_text}
+
+Assistant:
+
+Here, src_language and tgt_language indicate the source and target languages, and src_text denotes the source text requiring translation.
+
+Format Reward: We use regular expression extraction to enforce a structured response format. The model is required to place reasoning process within <think></think> tags and provide the translation inside <translate></translate>. The format reward score  $(S_{\text{format}})$  is computed as:
+
+$$
+S _ {f o r m a t} = \left\{ \begin{array}{l l} 1, & \text {i f f o r m a t i s c o r r e c t} \\ - 1, & \text {i f f o r m a t i s i n c o r r e c t} \end{array} \right.
+$$
+
+Metric Reward: In the NMT era, works often used string-level metrics like BLEU directly as a reward signal to fine-tune encoder-decoder models (Wu et al., 2018; Choshen et al., 2020; Kiegeland and Kreutzer, 2021). We use automatic evaluation metrics to calculate a translation quality score  $S_{\text{metric}}$ . We explore three approaches:
+
+1. N-gram Lexical Matching Reward (RewardLex): Metrics such as BLEU (Papineni et al., 2002) orchrF (Popovic, 2015) evaluate translation quality by measuring the difference (primarily lexical overlap) between the translation
+
+and the human-written reference. In our experiments, we calculate BLEU via sacrebleu<sup>1</sup>.
+
+2. Semantic and Contextual Reward (Reward-Sem): Learning-based metrics like COMET (Rei et al., 2020) and COMETKiwi (Rei et al., 2022) are trained on human judgments (e.g., MQM quality assessments (Freitag et al., 2021)). These metrics can recognize good translations even if the wording differs from the reference, as long as the meaning is preserved. We use the COMETKiwi-23-XL, which was used in the WMT 24 (Kocmi et al., 2024) and only needs the source sentence and the model's translation.  
+3. Lexical and Semantic Mixed Reward (Reward-Mix): To capture both lexical fidelity and semantic adequacy, we use a hybrid reward (Reward-Mix) that adds together Lexical Matching Reward (Reward-Lex) and Semantic and Contextual Reward (Reward-Sem). Beyond simple mixing, we also explore in-group score normalization strategies before addition in Appendix B.
+
+Accordingly, the computation of  $S_{\text{metric}}$  depends on the selected reward configuration:
+
+$$
+S _ {m e t r i c} = \left\{ \begin{array}{l l} \mathrm {B} (\text {t r a n s}, \text {r e f}), & \text {i f R e w a r d - L e x} \\ \mathrm {C K} (\text {s r c}, \text {t r a n s}), & \text {i f R e w a r d - S e m} \\ \mathrm {B} (\text {t r a n s}, \text {r e f}) + \mathrm {C K} (\text {s r c}, \text {t r a n s}), & \text {i f R e w a r d - M i x} \end{array} \right.
+$$
+
+where B denotes normalized BLEU score, CK denotes the COMETKiwi score, trans is the generated translation, ref is the reference translation, and src is the source text.
+
+Rule-Metric Mixed Reward: The final reward  $r$  combines both the format reward ( $S_{\text{format}}$ ) and the metric reward ( $S_{\text{metric}}$ ). Formally, it is calculated using the following rule:
+
+$$
+r = \left\{ \begin{array}{l l} S _ {f o r m a t} - 2, & \text {i f} S _ {f o r m a t} = - 1 \\ S _ {f o r m a t} + S _ {m e t r i c}, & \text {i f} S _ {f o r m a t} = 1 \end{array} \right.
+$$
+
+where  $S_{\text{metric}}$  is calculated only if the response format is correct.  $S_{\text{format}} = 1$ . Then the final reward becomes  $r = 1 + S_{\text{metric}}$ . Unlike traditional rule-based rewards that give a fixed score for correct outputs, our approach uses a continuous metric score. This means the reward can vary within the [1, 2] or [1, 3] range, depending on translation quality. As a result, the model receives more detailed feedback and can learn to improve even small differences in translation quality.
+
+# 3.2 RL Algorithm
+
+We use the GRPO algorithm (Shao et al., 2024) to train the translation model with our rule-metric mixed reward. In each training step, for a given query  $q$ , we sample a group of candidate outputs  $\{o_1, o_2, \dots, o_G\}$  from the policy model  $\pi_{\theta_{old}}$ .  $A_i = \frac{r_i - \mathrm{mean}(\{r_1, r_2, \dots, r_G\})}{\mathrm{std}(\{r_1, r_2, \dots, r_G\})}$  is the computed advantage using the group rule-metric mixed rewards  $\{r_1, r_2, \dots, r_G\}$ . GRPO then maximizes the following objective function to optimize  $\pi_\theta$ :
+
+$$
+\begin{array}{l} J _ {\mathrm {G R P O}} (\theta) = \mathbb {E} _ {q \sim P (Q), \{o _ {i} \} _ {i = 1} ^ {G} \sim \pi_ {\theta_ {\mathrm {o l d}}} (O | q)} \\ \left[ \frac {1}{G} \sum_ {i = 1} ^ {G} \min  \left(\frac {\pi_ {\theta} (o _ {i} \mid q)}{\pi_ {\theta_ {\mathrm {o l d}}} (o _ {i} \mid q)} A _ {i}, \right. \right. \\ \operatorname {c l i p} \left(\frac {\pi_ {\theta} (o _ {i} \mid q)}{\pi_ {\theta_ {\mathrm {o l d}}} (o _ {i} \mid q)}, 1 - \varepsilon , 1 + \varepsilon\right) A _ {i}\left. \right) \\ \left. - \beta D _ {\mathrm {K L}} \left(\pi_ {\theta} \| \pi_ {\text {r e f}}\right) \right], \tag {1} \\ \end{array}
+$$
+
+where  $\varepsilon$  and  $\beta$  are hyperparameters controlling the PPO clipping threshold and the weight of the Kullback-Leibler (KL) divergence penalty (Schulman et al., 2017; Shao et al., 2024), respectively.
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+Dataset and Benchmarks. Our experiments primarily focus on English (EN) and Chinese (ZH). Following Xu et al. (2023) and Feng et al. (2024a), we sourced EN=ZH parallel examples from WMT 2017-2020, totaling 13,130 pairs. For in-domain evaluation, we use WMT 24 (EN-ZH) and WMT 23 (ZH-EN). Out-of-distribution (OOD) generalization is assessed on benchmarks covering: (1) unseen language pairs from WMT: English-Japanese (EN-JA, WMT 2024) and German-English (DE-EN, WMT 2023 Document-level); and (2) unseen language pair from a distinct dataset: German-Chinese (DE-ZH, Flores-200 (Costa-jussà et al., 2022)). Detailed statistics are in Appendix F.
+
+Baselines. Our primary baselines encompass leading proprietary models, namely Claude-3.5-Sonnet (Anthropic, 2024), GPT-4o (OpenAI, 2023), and Gemini-1.5-Pro (Team et al., 2024), alongside advanced open-source models such as the Qwen2.5 series (Yang et al., 2024), LLaMA-3.1 series (Grattafori et al., 2024), advanced multilingual models Aya series (Aryabumi et al., 2024), and translation-specific Tower family (Alves et al., 2024). More evaluation details are in Appendix E.
+
+<table><tr><td rowspan="2">MODEL</td><td colspan="4">ZH-EN</td><td colspan="4">EN-ZH</td></tr><tr><td>BLEU</td><td>COMETKiwi</td><td>XCOMET</td><td>Avg.</td><td>BLEU</td><td>COMETKiwi</td><td>XCOMET</td><td>Avg.</td></tr><tr><td colspan="9">Closed</td></tr><tr><td>Claude-3.5-Sonnet (2024/10)</td><td>22.55</td><td>71.69</td><td>87.32</td><td>60.52</td><td>38.63</td><td>70.39</td><td>78.24</td><td>62.42</td></tr><tr><td>GPT-4o (2024/08)</td><td>22.57</td><td>71.63</td><td>87.22</td><td>60.47</td><td>41.13</td><td>69.01</td><td>75.43</td><td>61.86</td></tr><tr><td>Gemini-1.5-Pro (2025/03)</td><td>18.34</td><td>69.23</td><td>85.55</td><td>57.71</td><td>39.82</td><td>67.47</td><td>76.26</td><td>61.18</td></tr><tr><td colspan="9">Open</td></tr><tr><td colspan="9">General Purpose LLMs</td></tr><tr><td>LLaMA-3.1-70B-Instruct</td><td>25.19</td><td>70.43</td><td>86.21</td><td>60.61</td><td>39.82</td><td>68.05</td><td>75.17</td><td>61.01</td></tr><tr><td>Qwen2.5-72B-Instruct</td><td>21.96</td><td>70.95</td><td>87.07</td><td>59.99</td><td>39.29</td><td>69.04</td><td>76.97</td><td>61.77</td></tr><tr><td>Qwen2.5-32B-Instruct</td><td>20.54</td><td>69.35</td><td>85.47</td><td>58.45</td><td>36.36</td><td>68.43</td><td>74.90</td><td>59.90</td></tr><tr><td colspan="9">Multilingual and Translation-Specific LLMs</td></tr><tr><td>TowerInstruct-13B-v0.1</td><td>24.72</td><td>70.17</td><td>85.69</td><td>60.19</td><td>37.06</td><td>66.22</td><td>73.13</td><td>58.80</td></tr><tr><td>TowerInstruct-7B-v0.2</td><td>23.32</td><td>69.99</td><td>84.93</td><td>59.41</td><td>34.93</td><td>64.04</td><td>70.67</td><td>56.55</td></tr><tr><td>Aya-23-35B</td><td>21.99</td><td>68.68</td><td>84.32</td><td>58.33</td><td>36.33</td><td>64.40</td><td>72.10</td><td>57.61</td></tr><tr><td>Aya-23-8B</td><td>19.13</td><td>66.74</td><td>82.89</td><td>56.25</td><td>33.28</td><td>63.11</td><td>70.78</td><td>55.72</td></tr><tr><td colspan="9">Ours</td></tr><tr><td>Qwen2.5-3B-Base</td><td>14.26</td><td>64.86</td><td>76.76</td><td>51.96</td><td>15.90</td><td>52.05</td><td>67.13</td><td>45.03</td></tr><tr><td>MT-R1-Zero-3B-Lex</td><td>21.53</td><td>66.33</td><td>81.69</td><td>56.52</td><td>33.70</td><td>60.58</td><td>65.67</td><td>53.32</td></tr><tr><td>MT-R1-Zero-3B-Sem</td><td>18.41</td><td>70.33</td><td>85.98</td><td>58.24</td><td>24.32</td><td>69.75</td><td>76.92</td><td>57.00</td></tr><tr><td>MT-R1-Zero-3B-Mix</td><td>22.54</td><td>68.84</td><td>84.08</td><td>58.49</td><td>36.27</td><td>65.05</td><td>72.10</td><td>57.81</td></tr><tr><td>Qwen2.5-7B-Base</td><td>18.23</td><td>68.27</td><td>84.99</td><td>57.16</td><td>31.14</td><td>63.38</td><td>69.83</td><td>54.78</td></tr><tr><td>MT-R1-Zero-7B-Lex</td><td>23.56</td><td>65.35</td><td>82.12</td><td>57.01</td><td>40.11</td><td>64.57</td><td>70.21</td><td>58.30</td></tr><tr><td>MT-R1-Zero-7B-Sem</td><td>16.62</td><td>71.66</td><td>86.07</td><td>58.12</td><td>23.07</td><td>72.07</td><td>79.37</td><td>58.17</td></tr><tr><td>MT-R1-Zero-7B-Mix</td><td>23.98</td><td>70.81</td><td>86.17</td><td>60.32</td><td>40.97</td><td>69.43</td><td>76.36</td><td>62.25</td></tr></table>
+
+Evaluation Metrics. We assess translation quality using a suite of complementary metrics, including the lexical metric BLEU (Post, 2018), the reference-free learning-based metric COMETKiwi-23-XL (Rei et al., 2022), and a SoTA, reference-based learning metric XCOMET-XL (Guerreiro et al., 2024) that was not directly used as an optimization target. To provide an even more holistic view, we also report scores from MetricX-23-XL (Juraska et al., 2023), another high-performing reference-based metric from a distinct framework, and  $\mathrm{chrF} + +$  (Popovic, 2015), which integrates character-level and word-level matching.
+
+Training Details. Our implementation is based on verl². We selected Qwen2.5-base series (3B and 7B parameter variants) as starting models for MT-R1-Zero training. More details are in Appendix G.
+
+# 4.2 Main Results
+
+In-Domain Performance. As detailed in Table 1 and Table 7, MT-R1-Zero models demonstrate substantial gains over their base versions and achieve competitive performance against existing SoTA benchmarks. For EN-ZH, MT-R1-Zero-7B-Mix surpasses advanced models like GPT-4o and Qwen2.5-72B on average scores. The MT-R1-Zero-7B-Sem variant particularly ex
+
+cels in semantic-level evaluations (COMETKiwi, XCOMET), outperforming strong proprietary models. This strength is further supported by a strong MetricX-23 score of 2.42, notably better than GPT-4o (3.29) and Qwen2.5-72B (3.08). On ZHEN, MT-R1-Zero-7B-Mix remains highly competitive. MT-R1-Zero-7B-Sem achieves COMETKiwi scores comparable to leading closed models and surpasses strong open-source counterparts. Furthermore, the MT-R1-Zero-3B-Sem delivers impressive performance for its scale. It scores 69.75 COMETKiwi on EN-ZH, which is approximately 1.7 points higher than the much larger LLaMA-3.1-70B and over 0.7 points above Qwen2.5-72B.
+
+Table 1: Performance comparison on in-domain translation directions (EN-ZH, ZH-EN) using BLEU, COMETKiwi, and XCOMET metrics, with average metric scores (Avg.). MT-R1-Zero variants (-Lex, -Sem, -Mix) are compared against closed and open baselines, which are further categorized by accessibility and specialization. The -Mix variant often achieves the best balance, while -Sem reaches peak semantic scores. Bold and underline denote the best and the second best score for each column respectively.  
+
+<table><tr><td rowspan="2">MODEL</td><td colspan="4">OUT-OF-DISTRIBUTION</td></tr><tr><td>EN-JA</td><td>DE-EN (DOC)</td><td>DE-ZH</td><td>Avg.</td></tr><tr><td colspan="5">Strong Baseline</td></tr><tr><td>Qwen2.5-72B-Instruct</td><td>76.86</td><td>89.51</td><td>88.42</td><td>84.93</td></tr><tr><td>LLaMA3.1-70B-Instruct</td><td>75.67</td><td>88.72</td><td>87.42</td><td>83.94</td></tr><tr><td colspan="5">Same-size Baseline</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>63.74</td><td>87.45</td><td>84.43</td><td>78.54</td></tr><tr><td>LLaMA-3.1-8B-Instruct</td><td>64.50</td><td>86.84</td><td>82.23</td><td>77.86</td></tr><tr><td>TowerInstruct-7B-v0.2</td><td>56.73</td><td>89.47</td><td>84.28</td><td>76.83</td></tr><tr><td>Aya-23-8B</td><td>72.23</td><td>87.97</td><td>83.84</td><td>81.35</td></tr><tr><td>MT-R1-Zero-7B-Lex</td><td>60.65</td><td>85.25</td><td>83.86</td><td>76.59</td></tr><tr><td>MT-R1-Zero-7B-Sem</td><td>71.95</td><td>87.68</td><td>87.66</td><td>82.43</td></tr><tr><td>MT-R1-Zero-7B-Mix</td><td>68.49</td><td>88.69</td><td>88.69</td><td>81.96</td></tr></table>
+
+Table 2: Out-of-distribution performance comparison evaluated by the XCOMET metric.
+
+Out-of-Distribution Performance. Our models also demonstrate robust generalization across different OOD scenarios, which encompass unseen
+
+![](images/2811ea8253d97e9328f7ae90aed1b4a8b8e2a38760abf3a2a3452014998cd086.jpg)
+
+![](images/66ce8b1724fddf268cddd9f249a55f5c3dbeafcc4943976031e4bf9a37474adc.jpg)
+
+![](images/5aa850c5bcc2864317b1069f0decd30a758e93e68cf78162b98ee4bd73d18aee.jpg)
+
+![](images/d3a9615006087f42568ade95a43178f48bb2e6d65f959546f2e29cb00a863cec.jpg)  
+Figure 2: Progression of translation quality metrics using Reward-Lex, Reward-Sem, and Reward-Mix.  
+Figure 3: Qualitative examples illustrates the effect of different reward functions (Reward-Lex, Reward-Sem, Reward-Mix) on EN-ZH translation, where the stylistic differences are driven by reward optimization (Finding 1).
+
+WMT language pairs (EN-JA, DE-EN) and the distinct Flores-200 DE-ZH dataset. XCOMET scores presented in Table 2 highlight this: MT-R1-Zero7B-Sem excels, achieving an average XCOMET score of 82.43 across these tasks, which is 3.89 points higher than Qwen2.5-7B-Instruct. While our 7B variants do not consistently surpass much larger strong baselines, they significantly outperform other evaluated same-size baselines by a considerable margin, an outperformance that extends to defeating same-size multilingual and translation-specific baselines (such as TowerInstruct-7B-v0.2 and Aya-23-8B). These OOD results suggest that the quality improvements in MT-R1-Zero can effectively transfer to unseen language pairs and benchmarks. More results are provided in Appendix I.
+
+# 5 Key Findings and Insight
+
+# 5.1 Impact of Reward Metric Selection
+
+As detailed in Section 3.1, we explore three metric rewards: Reward-Lex, Reward-Sem, and Reward-
+
+Mix. Our results demonstrate that the choice among these significantly affects the learning target and
+
+Finding 1: Reward metric selection critically shapes optimization targets and translation style.
+
+final model outputs, as stated in Finding 1.
+
+Figure 2 presents the training dynamics with different rewards. Training with Reward-Lex maximizes BLEU scores, often at the expense of semantic scores, while Reward-Sem maximizes COMETKiwi, leading to a decline in BLEU. Training with Reward-Mix improves both metrics. Independent evaluation with XCOMET further supports this finding, showing consistent improvements for Sem and Mix variants while fluctuating for Lex. This finding aligns with the insight from Chen et al. (2025), suggesting that lexical and semantic assessments are complementary, particularly for reasoning-oriented LLMs, and combining them can offer a more comprehensive evaluation signal.
+
+![](images/cc8a9e7d8ff68e9e574c26b2e79dffd19bc96084cfc8bf8b400b54ed606abe6c.jpg)  
+Step 0
+
+![](images/aadd8a879e63938d0337c56b3e8290db7e534392c38b97c86bb355852cd96519.jpg)  
+Figure 4: Examples illustrating language-of-thought phenomenon, i.e., transition of the internal reasoning language in MT-R1-Zero models. The reasoning language transits from English at Step 0 to target language at Step 1600, indicated by bold text across various OOD test pairs (Finding 2).
+
+![](images/3a387b2d378583894b7234f84c162dcf61cdc22834b3aaeb413da403372f910f.jpg)  
+Figure 5: Comparison of training dynamics for different model families (Qwen2.5, LLaMA-3.1, Tower) undergoing MT-R1-Zero RL training, highlighting differences in adaptability (Finding 3).
+
+![](images/068540205047d5d52d44d5f419507c3f4bb08830eaf3f8d52be98aac65eaef73.jpg)
+
+Qualitatively (Figure 3), this optimization alignment manifests as distinct translation styles. Reward-Lex encourages literal focused translations, potentially sacrificing nuance. Reward-Sem fosters translations that prioritize semantic faithfulness, even if lexically divergent from references. In contrast, the mixed reward yields balanced translations. This demonstrates that the metric reward fundamentally dictates the nature of the translation quality learned (e.g., semantic v.s. lexical). Therefore, careful metric selection and deliberate fusion are essential for tailoring RL-based MT refinement towards specific and desired translations.
+
+Moreover, a closer inspection of Figure 2 reveals that the BLEU progression for Reward-Mix closely mirrors that of Reward-Lex, suggesting the unnormalized reward signal may be disproportionately influenced, or even dominated, by the more volatile BLEU metric score. This observation motivated
+
+our deeper investigation into reward normalization strategies (see Appendix B) to ensure a more balanced contribution from each component.
+
+# 5.2 Emergence and Evolution of Translation Thinking Patterns
+
+As R1-Zero-like training lacks a cold-start phase with predefined reasoning patterns, the observed thinking processes should be emergent and shaped
+
+Finding 2: Diverse reasoning patterns emerge, varying in style and complexity, and moreover, the internal reasoning language could dynamically transit to target languages even for OOD settings.
+
+by the RL objective.
+
+Our framework incentivizes a variety of reasoning styles within the <think></think> tags (Fig-
+
+<table><tr><td rowspan="3">Model</td><td colspan="4">In-domain</td><td colspan="6">Out-of-distribution</td></tr><tr><td colspan="2">ZH-EN</td><td colspan="2">EN-ZH</td><td colspan="2">EN-JA</td><td colspan="2">DE-ZH</td><td colspan="2">DE-EN (Doc)</td></tr><tr><td>COMETKiwi</td><td>XCOMET</td><td>COMETKiwi</td><td>XCOMET</td><td>COMETKiwi</td><td>XCOMET</td><td>COMETKiwi</td><td>XCOMET</td><td>COMETKiwi</td><td>XCOMET</td></tr><tr><td>Qwen2.5-7B (SFT)</td><td>69.29</td><td>84.80</td><td>67.25</td><td>74.29</td><td>67.77</td><td>65.39</td><td>67.01</td><td>86.17</td><td>67.44</td><td>86.74</td></tr><tr><td>Qwen2.5-7B (RL w/o thinking)</td><td>70.78</td><td>86.26</td><td>69.62</td><td>76.03</td><td>68.68</td><td>68.77</td><td>67.84</td><td>86.67</td><td>68.31</td><td>88.30</td></tr><tr><td>Qwen2.5-7B (RL w/ thinking)</td><td>70.81</td><td>86.17</td><td>69.43</td><td>76.36</td><td>69.27</td><td>68.49</td><td>70.25</td><td>89.25</td><td>68.74</td><td>88.69</td></tr></table>
+
+Table 3: Performance comparison of different training paradigms: Supervised Fine-Tuning (SFT) vs. RL with explicit thinking (RL w/ thinking) vs. RL without explicit thinking (RL w/o thinking). Results shown for in-domain and out-of-distribution tasks support the finding that the RL process itself is the primary driver of gains.
+
+ure 11). While some instances include explicit "review/refine" steps, these generally appear as pre-planned components rather than the conversational, iterative self-correction characteristic of the "Aha moment" reported in mathematical reasoning tasks (DeepSeek-AI et al., 2025; Hu et al., 2025). This suggests that while MT-R1-Zero successfully encourages thinking, the complexity and specific nature of emergent reasoning are task-dependent.
+
+Furthermore, we observe a striking and interesting "language-of-thought" (transition in the language used for internal reasoning) phenomenon during OOD testing (Figure 4). While base models often use English as default thinking language based on template, MT-R1-Zero models progressively transit to utilize the target language of the translation task for their reasoning process within the <think></think> block during training (see bold Japanese or Chinese text in step 1600). This dynamic adaptation of the internal "language of thought", conditioned on the task, emerges even without direct supervision on reasoning language.
+
+# 5.3 Training Dynamics of Different LLMs
+
+The effectiveness and training behavior of MT-R1-Zero are significantly influenced by the base LLM architecture and its initial state (pre-trained vs. instruction-tuned). We compare three distinct model families: general (Qwen2.5 and LLaMA-3.1 series) and translation-specific (Tower family).
+
+Finding 3: LLM architectures exhibit distinct adaptability and effectiveness under MT-R1-Zero, with Qwen showing the highest compatibility in format learning and reasoning generation, while LLaMA and Tower face more challenges and tend towards "format hacking".
+
+As shown in Figure 5, both the translation-specific (Tower) and LLaMA-3.1 models exhibit significantly slower adaptation to the required format compared to Qwen models, as evidenced by their delayed format error reduction. Furthermore,
+
+qualitative analysis (Figure 10) reveals that these models often circumvent meaningful reasoning by generating minimal or templated placeholder content in the <think></think> tags, potentially "hacking" the format reward. In contrast, Qwen2.5 models demonstrate stronger adaptability, consistently producing coherent reasoning text within the structured framework. This suggests that architectures like Qwen may possess inherent advantages for integrating structured reasoning via RL, a finding that aligns with prior work on cognitive behaviors (Gandhi et al., 2025). However, even Qwen2.5 models occasionally regress to simplistic one-sentence outputs during reasoning tasks, underscoring the instability of exploration in R1-Zero-like training paradigms.
+
+# 5.4 Disentangling RL and Explicit Thinking
+
+To determine whether the explicit <think> step or the underlying RL optimization is the primary driver of performance gains, we conducted an ablation study. We compared three paradigms: Supervised Fine-Tuning (SFT), our standard MT-R1-Zero-Mix (RL w/ thinking), and an RL variant without the explicit <think> step (RL w/o thinking).
+
+The results presented in Table 3 reveal a key finding: both RL configurations achieve comparable performance while substantially outperforming the SFT baseline across both in-domain and OOD settings. This demonstrates that the major performance improvements in MT-R1-Zero are primarily driven by the RL framework itself, rather than the mere presence of an explicit reasoning step. This core conclusion is further corroborated by similar findings on the DRT literature translation benchmark (see Appendix J for the full analysis).
+
+# 6 Conclusion
+
+In this work, we introduced MT-R1-Zero, the first successful adaptation of R1-Zero RL framework to MT using a novel rule-metric mixed reward mechanism that combines format enforcement with quality metrics. Our MT-R1-Zero significantly im
+
+proves translation quality, achieving leading results on multiple benchmarks, i.e., our 3B models compete with much larger open-source models, while our 7B models are on par with advanced proprietary models. The MT-R1-Zero also demonstrates strong OOD generalization and multilingual applicability. Through extensive experiments and analysis, we highlight the significant impact of reward metric choice for optimization, showcase distinct adaptability across different LLMs, and reveal that performance gains are principally from the RL process itself rather than reasoning steps or vocabulary, establishing R1-Zero as a viable and potent paradigm for advancing MT. More broadly, our work highlights the great potential of RL for diverse language processing tasks beyond translation.
+
+# Limitations
+
+While MT-R1-Zero represents a significant advance, certain limitations remain. The emergent reasoning observed, though diverse, did not achieve the sophisticated iterative self-correction capabilities demonstrated in mathematical reasoning tasks using similar RL or R1-like methods. This discrepancy may reflect fundamental differences in task structure or indicate the need for specialized design in translation tasks. One promising direction would be developing task-specific cold-start datasets for SFT before RL optimization, though this would deviate from the pure RL paradigm we investigated here. Future work could focus on inducing deeper reasoning structures specifically beneficial for the MT task, investigating architectural adaptability across a broader range of LLMs, and developing more appropriate reward mechanisms. Exploring applications to specialized domains (e.g., law and healthcare) and general language processing tasks presents promising opportunities to extend this work.
+
+# Acknowledgments
+
+This work is supported by the National Key R&D Program of China (Grant No. 2024YFC3308304), the "Pioneer" and "Leading Goose" R&D Program of Zhejiang (Grant no. 2025C01128), the National Natural Science Foundation of China (Grant No. 62476241), the Natural Science Foundation of Zhejiang Province, China (Grant No. LZ23F020008), the Zhejiang Zhejiang Key Laboratory of Medical Imaging Artificial Intelligence.
+
+# References
+
+Duarte M Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peeters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. 2024. Tower: An open multilingual large language model for translation-related tasks. arXiv preprint arXiv:2402.17733.  
+Anthropic. 2024. Claude 3.5 sonnet.  
+Viraat Aryabumi, John Dang, Dwarak Talupuru, Saurabh Dash, David Cairuz, Hangyu Lin, Bharat Venkitesh, Madeline Smith, Jon Ander Campos, Yi Chern Tan, Kelly Marchisio, Max Bartolo, Sebastian Ruder, Acyr Locatelli, Julia Kreutzer, Nick Frosst, Aidan Gomez, Phil Blunsom, Marzieh Fadaee, Ahmet Üstün, and Sara Hooker. 2024. Aya 23: Open weight releases to further multilingual progress. Preprint, arXiv:2405.15032.  
+Andong Chen, Yuchen Song, Wenxin Zhu, Kehai Chen, Muyun Yang, Tiejun Zhao, et al. 2025. Evaluating o1-like llms: Unlocking reasoning for translation through comprehensive analysis. arXiv preprint arXiv:2502.11544.  
+Leshem Choshen, Lior Fox, Zohar Aizenbud, and Omri Abend. 2020. On the weaknesses of reinforcement learning for neural machine translation. In International Conference on Learning Representations.  
+Marta R Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.  
+Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, et al. 2025. Process reinforcement through implicit rewards. arXiv preprint arXiv:2502.01456.  
+DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen
+
+Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang and Zhen Zhang. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint arXiv:2501.12948.
+
+Hugging Face. 2025. Open r1: A fully open reproduction of deepseek-r1.
+
+Xidong Feng, Ziyu Wan, Muning Wen, Stephen Marcus McAleer, Ying Wen, Weinan Zhang, and Jun Wang. 2023. Alphazero-like tree-search can guide large language model decoding and training. arXiv preprint arXiv:2309.17179.
+
+Zhaopeng Feng, Ruizhe Chen, Yan Zhang, Zijie Meng, and Zuozhu Liu. 2024a. Ladder: A model-agnostic framework boosting LLM-based machine translation to the next level. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15377-15393, Miami, Florida, USA. Association for Computational Linguistics.
+
+Zhaopeng Feng, Jiahan Ren, Jiayuan Su, Jiamei Zheng, Zhihang Tang, Hongwei Wang, and Zuozhu Liu. 2025. Mt-rewardtree: A comprehensive framework for advancing llm-based machine translation via reward modeling. arXiv preprint arXiv:2503.12123.
+
+Zhaopeng Feng, Yan Zhang, Hao Li, Wenqiang Liu, Jun Lang, Yang Feng, Jian Wu, and Zuozhu Liu. 2024b. Improving llm-based machine translation with systematic self-correction. arXiv preprint arXiv:2402.16379.
+
+Markus Freitag, George Foster, David Grangier, Viresh Ratnakar, Qijun Tan, and Wolfgang Macherey. 2021. Experts, errors, and context: A large-scale study of
+
+human evaluation for machine translation. Transactions of the Association for Computational Linguistics, 9:1460-1474.
+
+Markus Freitag, Nitika Mathur, Chi-kiu Lo, Eleftherios Avramidis, Ricardo Rei, Brian Thompson, Tom Kocmi, Frederic Blain, Daniel Deutsch, Craig Stewart, Chrysoula Zerva, Sheila Castilho, Alon Lavie, and George Foster. 2023. Results of WMT23 metrics shared task: Metrics might be guilty but references are not innocent. In Proceedings of the Eighth Conference on Machine Translation, pages 578-628, Singapore. Association for Computational Linguistics.
+
+Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. 2025. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. arXiv preprint arXiv:2503.01307.
+
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
+
+Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. rstar-math: Small llms can master math reasoning with self-evolved deep thinking. arXiv preprint arXiv:2501.04519.
+
+Nuno M Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and Andre FT Martins. 2024. xcomet: Transparent machine translation evaluation through fine-grained error detection. Transactions of the Association for Computational Linguistics, 12:979-995.
+
+Minggui He, Yilun Liu, Shimin Tao, Yuanchang Luo, Hongyong Zeng, Chang Su, Li Zhang, Hongxia Ma, Daimeng Wei, Weibin Meng, et al. 2025. R1-t1: Fully incentivizing translation capability in llms via reasoning learning. arXiv preprint arXiv:2502.19735.
+
+Jingcheng Hu, Yinmin Zhang, Qi Han, Daxin Jiang, and Heung-Yeung Shum Xiangyu Zhang. 2025. Open-reasoner-zero: An open source approach to scaling reinforcement learning on the base model. https://github.com/Open-Reasoner-Zero/Open-Reasoner-Zero.
+
+Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Yao Hu, and Shaohui Lin. 2025. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749.
+
+Juraj Juraska, Mara Finkelstein, Daniel Deutsch, Aditya Siddhant, Mehdi Mirzazadeh, and Markus Freitag. 2023. MetricX-23: The Google submission to the WMT 2023 metrics shared task. In Proceedings of the Eighth Conference on Machine Translation, pages 756-767.
+
+Samuel Kiegeland and Julia Kreutzer. 2021. Revisiting the weaknesses of reinforcement learning for neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1673-1681, Online. Association for Computational Linguistics.  
+Tom Kocmi, Eleftherios Avramidis, Rachel Bawden, Ondrej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, et al. 2024. Preliminary wmt24 ranking of general mt systems and llms. arXiv preprint arXiv:2407.19884.  
+Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let's verify step by step. In *The Twelfth International Conference on Learning Representations*.  
+Sinuo Liu, Chenyang Lyu, Minghao Wu, Longyue Wang, Weihua Luo, and Kaifu Zhang. 2025. New trends for modern machine translation with large reasoning models. arXiv preprint arXiv:2503.10351.  
+Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, et al. 2024. Improve mathematical reasoning in language models by automated process supervision. arXiv preprint arXiv:2406.06592.  
+OpenAI. 2023. GPT-4: technical work.  
+OpenAI. 2024. Introducing openai o1. https://openai.com/o1/. Accessed: 2024-10-02.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Maja Popovic. 2015. *chrF: character n-gram F-score* for automatic MT evaluation. In *Proceedings of the Tenth Workshop on Statistical Machine Translation*, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics.  
+Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Brussels, Belgium. Association for Computational Linguistics.  
+Zhenting Qi, Mingyuan Ma, Jiahang Xu, Li Lyna Zhang, Fan Yang, and Mao Yang. 2024. Mutual reasoning makes smaller llms stronger problem-solvers. arXiv preprint arXiv:2408.06195.  
+Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. Comet: A neural framework for mt evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702.
+
+Ricardo Rei, Marcos Treviso, Nuno M Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José GC De Souza, Taisiya Glushkova, Duarte Alves, Luísa Coheur, et al. 2022. Cometkiwi: Ist-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 634-645.  
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.  
+Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.  
+David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, L. Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Vedavyas Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy P. Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. 2016. Mastering the game of go with deep neural networks and tree search. Nature, 529:484-489.  
+Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314.  
+Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530.  
+Kimi Team. 2025a. Kimi k1.5: Scaling reinforcement learning with llms.  
+Qwen Team. 2025b. Qwq-32b: Embracing the power of reinforcement learning.  
+Jiaan Wang, Fandong Meng, Yunlong Liang, and Jie Zhou. 2024a. Drt-o1: Optimized deep reasoning translation via long chain-of-thought. arXiv preprint arXiv:2412.17498.  
+Yutong Wang, Jiali Zeng, Xuebo Liu, Fandong Meng, Jie Zhou, and Min Zhang. 2024b. Taste: Teaching large language models to translate through self-reflection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6144-6158.  
+Lijun Wu, Fei Tian, Tao Qin, Jianhuang Lai, and TieYan Liu. 2018. A study of reinforcement learning for neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3612-3621, Brussels, Belgium. Association for Computational Linguistics.
+
+Violet Xiang, Charlie Snell, Kanishk Gandhi, Alon Balak, Anikait Singh, Chase Blagden, Duy Phung, Rafael Rafailov, Nathan Lile, Dakota Mahan, et al. 2025. Towards system 2 reasoning in llms: Learning how to think with meta chain-of-though. arXiv preprint arXiv:2501.04682.
+
+Tian Xie, Zitian Gao, Qingnan Ren, Haoming Luo, Yuqian Hong, Bryan Dai, Joey Zhou, Kai Qiu, Zhirong Wu, and Chong Luo. 2025. Logic-rl: Unleashing llm reasoning with rule-based reinforcement learning. arXiv preprint arXiv:2502.14768.
+
+Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models. arXiv preprint arXiv:2309.11674.
+
+Haoran Xu, Kenton Murray, Philipp Koehn, Hieu Hoang, Akiko Eriguchi, and Huda Khayrallah. 2024. X-alma: Plug & play modules and adaptive rejection for quality translation at scale. arXiv preprint arXiv:2410.03115.
+
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.
+
+Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. 2025. Demystifying long chain-of-thought reasoning in llms. arXiv preprint arXiv:2502.03373.
+
+Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. 2025. Dapo: An opensource llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476.
+
+Lifan Yuan, Wendi Li, Huayu Chen, Ganqu Cui, Ning Ding, Kaiyan Zhang, Bowen Zhou, Zhiyuan Liu, and Hao Peng. 2024. Free process rewards without process labels. arXiv preprint arXiv:2412.01981.
+
+Zhiyuan Zeng, Qinyuan Cheng, Zhangyue Yin, Bo Wang, Shimin Li, Yunhua Zhou, Qipeng Guo, Xuanjing Huang, and Xipeng Qiu. 2024. Scaling of search and learning: A roadmap to reproduce o1 from reinforcement learning perspective. arXiv preprint arXiv:2412.14135.
+
+Yu Zhao, Huifeng Yin, Bo Zeng, Hao Wang, Tianqi Shi, Chenyang Lyu, Longyue Wang, Weihua Luo, and Kaifu Zhang. 2024. Marco-o1: Towards open reasoning models for open-ended solutions. Preprint, arXiv:2411.14405.
+
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient finetuning of  $100+$  language models. arXiv preprint arXiv:2403.13372.
+
+# A Progression of Response Length and Performance
+
+By observing the training process, we provide several insights into model adaptation and the emergence of reasoning.
+
+Finding 2 (Continued): Response length initially declines rapidly and then gradually increases as training progresses.
+
+Figure 7 (Right) depicts the pattern in Finding 2 (Continued) alongside consistent COMETKiwi improvements (Left). Qualitative analysis (Figure 8) reveals that this length trajectory reflects evolving reasoning strategies. The initial decline corresponds to the model mastering the required format while transitioning from naive decomposition (Step 0) to more efficient, direct translations. The subsequent increase aligns with the development of richer semantic analysis and deeper contextual reasoning within the <think></think> tags (Step 1600).
+
+# B Analysis of Reward Normalization Strategies
+
+As noted in Section 5.1, naively mixing metrics like BLEU and COMETKiwi can be problematic due to their different scales and statistical properties. To investigate this, we experimented with in-group normalization strategies. As shown in Table 4, Z-score standardization yields the best overall performance, improving semantic scores while maintaining lexical fidelity.
+
+<table><tr><td>Strategy</td><td>BLEU</td><td>XCOMET</td><td>COMETKiwi</td></tr><tr><td>w/o Norm</td><td>20.75</td><td>86.51</td><td>70.83</td></tr><tr><td>Min-Max Norm</td><td>20.30</td><td>86.79</td><td>71.10</td></tr><tr><td>Z-score Norm</td><td>20.67</td><td>87.12</td><td>71.46</td></tr></table>
+
+Table 4: Performance on ZH-EN at Step 1600 with different reward normalization strategies.
+
+To understand the underlying mechanism, we analyzed the correlation between the raw metric scores and the final reward signal across the training process (Table 5). Without normalization, the reward's ranking is almost entirely dominated by BLEU's ranking (Spearman correlation of 0.93). This is likely due to BLEU's high variance, where small n-gram changes cause large score jumps. By applying normalization, particularly Z-score, we successfully decouple the final reward from being
+
+dominated by BLEU's volatility. This creates a more robust and balanced learning signal based on relative in-group performance, directly addressing the concerns of naive metric mixing.
+
+<table><tr><td>Correlation</td><td>Metric vs. Reward</td><td>w/o Norm</td><td>Min-Max</td><td>Z-score</td></tr><tr><td rowspan="2">Spearman</td><td>BLEU-Reward</td><td>0.9295</td><td>0.1263</td><td>0.0886</td></tr><tr><td>COMETK-Reward</td><td>0.7097</td><td>0.0661</td><td>0.1165</td></tr><tr><td rowspan="2">Pearson</td><td>BLEU-Reward</td><td>0.6440</td><td>0.2509</td><td>0.2657</td></tr><tr><td>COMETK-Reward</td><td>0.8357</td><td>0.5487</td><td>0.6849</td></tr></table>
+
+Table 5: Correlation analysis of raw metric scores vs. the final reward signal over the training process.
+
+# C KL Penalty Constrains Response Length but Not Quality Gains
+
+We investigate the effectiveness of the KL term in the GRPO objective (Equation 1) on response length and translation quality, as it would regularize the policy by discouraging large deviations from the initial reference model. We conducted experiments without the KL penalty (setting  $\beta = 0$ , Figure 6), and found that the average response length, after an initial drop, began to fluctuate and trend upward during training. This pattern is consistent with R1-Zero-like results in mathematical tasks (Yu et al., 2025; Yeo et al., 2025). Additional ablation of the KL penalty with COMETKiwi reveals that the improvement of translation quality appears to be largely independent of the thinking vocabulary. Significant quality gains were achieved in early-stage training (e.g., before Steps 400) before a substantial increase in response length, even in experiments conducted without the KL penalty. This suggests that performance improvements in the MT-R1-Zero setup could not be attributed solely or primarily to increasing reasoning vocabulary.
+
+# D Multilingual and Low-Resource Support
+
+To evaluate the broader applicability of our framework, we examine its effectiveness in multilingual training scenarios and its potential benefits for low-resource languages. We train multilingual MT-R1-Zero models using the Germanic language data split in the X-ALMA (Xu et al., 2024), augmented with Chinese (see Table 9 for detailed data statistics). We set the batch size to 16 and used COMET $22^{3}$  as the metric reward (Reward-Sem), consistent with the evaluation protocols in X-ALMA. All models are trained for 1 epoch on 16 NVIDIA H800
+
+![](images/02d4459d567a59959d88d68cb82db798eca459b4d2d95c749ba5fb7d78199642.jpg)
+
+![](images/3affe6599c57e77c8c6b0ba9069e3c096a75075b9a2ac09eb7f34bba292c47d0.jpg)  
+Figure 6: Effect of the KL divergence penalty on EN-ZH COMETKiwi score and response length progression for models trained with (w/ KL,  $\beta = 0.01$ ) and without (w/o KL,  $\beta = 0$ ) the penalty. Experiments are conducted three times with MT-R1-Zero-7B-Sem.
+
+80G GPUs for about 12 hours. All other hyperparameters follow the configuration described in Section 4.1. The training progress, measured by COMET-22 for English-to-target directions, is depicted in Figure 9.
+
+The learning curves demonstrate consistent improvement in translation quality across languages spanning diverse resource levels, including those typically considered low-resource (e.g., Icelandic (IS) and Norwegian (NO)). The steady performance improvement observed throughout training confirms that the MT-R1-Zero framework remains effective when applied in multilingual settings.
+
+# E Evaluation Details
+
+When evaluating model performance on the test set, we deployed open-source models locally using frameworks like vLLM or HuggingFace implementations. Proprietary models were accessed via their APIs. We use the sampling decoding strategy with a temperature of 0.2, and top_p set to 0.95. The maximum generation length was capped
+
+![](images/b960ab9b3489f6068bc5e89c25a51158dc6d7c56ea019e09ee8cf7fa4461c721.jpg)  
+Figure 7: Training dynamics of MT-R1-Zero models (using Reward-Sem). Left: COMETKiwi score progression for 3B and 7B models on EN-ZH and ZH-EN test sets. Right: Average response length changes over training steps, exhibiting the classic decrease-then-increase pattern (Finding 2 Continued).
+
+![](images/290482d11b85d1e741f288393a85868d178d55df66ab5fa7b70eca51374fea96.jpg)
+
+![](images/916c826f68668ff57dba4f4fa1b27508b78784ed88d932f0a5a27a4dc5e3729a.jpg)  
+Figure 8: Evolution of an MT-R1-Zero model's reasoning process and translation output for the Chinese source text "其影响可能类似于2008年的经济危机" at different training steps (0, 400, 1600), showcasing the shift from decomposition to more semantic analysis (Finding 2 Continued).
+
+# Inference Prompt
+
+Translate the following text from {src_language} into {tgt_language}.  
+{src_language}:{src_text}  
+{tgt_language}:
+
+Table 6: Prompt used for translation generation. {tgt_language}: target language; {src_language}: source language; {src_text}: the source test sentence.
+
+at 1024 tokens. We adopt the prompt showcasing in Table 6 to sample the translation (applying specific chat template when needed). For the multilingual and translation-specific models, we utilize the prompts in their official model cards.
+
+# F Data Statistics
+
+This section provides further details on the datasets used in our experiments. Table 8 outlines the statistics for the data employed in our main EN  $\rightleftharpoons$  ZH experiments. The training set for these experiments consists of  $13,130\mathrm{EN}\rightleftharpoons \mathrm{ZH}$  parallel sentence pairs, obtained after filtering out sentences with fewer than 30 characters. For model training, these examples were evenly divided between the two translation directions (6,565 pairs each for EN  $\rightarrow$  ZH and  $\mathrm{ZH}\rightarrow \mathrm{EN}$ ). Table 9 presents the statistics for the multilingual experiments discussed in Section D.
+
+![](images/adffe0d1800cd1b95ce3144b508cf9102be983771582a96dec9ffa8cb63e7600.jpg)
+
+![](images/f604c13b91ddff18b0a260594ba2454dd2a88f2da319bc713890d8d70197fce0.jpg)
+
+![](images/0784cd49f24e6bc61d2d02ba84c522f76ea788d61c414cc2beaf2a64c3d3c50c.jpg)
+
+![](images/c193ed946f14e97eb7339e0a72913d67cf8e7441974f899f6826bc944ce21682.jpg)
+
+![](images/568c8ec7b0ac71c7ca1877e0a6913f8002342687b7ec9e7b21e4740f61fef89a.jpg)  
+Figure 9: Training progression (COMET-22) for multilingual MT-R1-Zero models based on LLaMA-3.1-8B and Qwen2.5-7B across multiple EN-XX test sets, demonstrating applicability in multilingual settings (Section D).
+
+![](images/af4f4a0f77d6aabea1eedb56387ef420d2f30ca9777d046cf5942e7f207a14e5.jpg)
+
+![](images/5dd6f75176505fff4ad416ff2185079d262db822d236328f0321511e3c7d60be.jpg)
+
+![](images/e6eaf86090fd8cc85f578b4e24f7a62ea47273f1784112115b9b34537880393f.jpg)
+
+# G RL Training Details
+
+During training, we configure a batch size of 8 and utilize 8 rollouts per prompt within the GRPO algorithm. We employ a constant learning rate of 5e-7 and set the sampling temperature to 1.0. The maximum generation length for responses is capped at 1024 tokens. We set the KL penalty coefficient  $\beta$  to 0, thereby removing the KL constraint against the reference policy. This decision stems from our empirical observation that the KL penalty tends to restrict the model's exploration of diverse response lengths, which we will discuss further in Section C. The PPO clipping range  $\epsilon$  is set to 0.2. All models are trained for 1 epoch on 4 NVIDIA H800 80G GPUs for about 13 hours.
+
+# H Complementary Metric Scores for Main Results
+
+In addition to the primary metrics discussed in Section 4.2 for in-domain tasks, Table 7 reports complementary scores using MetricX-23-XL and  $\mathrm{chrF} + +$ . These metrics provide further perspectives on semantic quality (MetricX-23-XL) and lexical fidelity ( $\mathrm{chrF} + +$ ) for the EN=ZH evaluations, generally aligning with the strong performance trends of our MT-R1-Zero models.
+
+# I OOD COMETKiwi and BLEU Scores
+
+To supplement the OOD XCOMET results presented in Section 4.2, we provide COMETKiwi and BLEU scores for the same OOD tasks here.
+
+<table><tr><td rowspan="2">MODEL</td><td colspan="2">ZH-EN</td><td colspan="2">EN-ZH</td></tr><tr><td>ChrF++ (↑)</td><td>MetricX-23 (↓)</td><td>ChrF++ (↑)</td><td>MetricX-23 (↓)</td></tr><tr><td colspan="5">Closed</td></tr><tr><td>Claude-3.5-Sonnet (2024/10)</td><td>50.37</td><td>2.01</td><td>34.51</td><td>2.91</td></tr><tr><td>GPT-4o (2024/08)</td><td>51.06</td><td>2.08</td><td>32.81</td><td>3.29</td></tr><tr><td>Gemini-1.5-Pro (2025/03)</td><td>45.58</td><td>2.21</td><td>35.04</td><td>3.26</td></tr><tr><td colspan="5">General Purpose LLMs</td></tr><tr><td>LLaMA-3.1-70B-Instruct</td><td>51.43</td><td>2.32</td><td>35.85</td><td>3.40</td></tr><tr><td>Qwen2.5-72B-Instruct</td><td>52.17</td><td>2.21</td><td>39.27</td><td>3.08</td></tr><tr><td>Qwen2.5-32B-Instruct</td><td>50.15</td><td>2.46</td><td>37.47</td><td>3.31</td></tr><tr><td colspan="5">Multilingual and Translation-Specific LLMs</td></tr><tr><td>TowerInstruct-13B-v0.1</td><td>51.15</td><td>2.49</td><td>34.17</td><td>3.58</td></tr><tr><td>TowerInstruct-7B-v0.2</td><td>50.17</td><td>2.60</td><td>32.45</td><td>3.96</td></tr><tr><td>Aya-23-35B</td><td>47.98</td><td>2.83</td><td>33.33</td><td>3.86</td></tr><tr><td>Aya-23-8B</td><td>44.69</td><td>2.86</td><td>31.23</td><td>3.89</td></tr><tr><td colspan="5">Ours</td></tr><tr><td>MT-R1-Zero-3B-Lex</td><td>47.88</td><td>3.33</td><td>31.74</td><td>4.94</td></tr><tr><td>MT-R1-Zero-3B-Sem</td><td>48.47</td><td>2.21</td><td>30.27</td><td>2.78</td></tr><tr><td>MT-R1-Zero-3B-Mix</td><td>49.43</td><td>2.78</td><td>33.12</td><td>3.83</td></tr><tr><td>MT-R1-Zero-7B-Lex</td><td>49.25</td><td>3.17</td><td>34.27</td><td>4.13</td></tr><tr><td>MT-R1-Zero-7B-Sem</td><td>48.37</td><td>2.12</td><td>30.57</td><td>2.42</td></tr><tr><td>MT-R1-Zero-7B-Mix</td><td>51.55</td><td>2.44</td><td>38.03</td><td>3.11</td></tr></table>
+
+Table 7:  $\mathrm{ChrF}++$  and MetricX-23 scores for all models and translation directions (ZH-EN, EN-ZH). For each column, bold marks the best and underline marks the second best performance.  $(\uparrow)$ : higher is better,  $(\downarrow)$ : lower is better. MT-R1-Zero rows are aligned with the corresponding variants in Table 1.
+
+Table 10 details the COMETKiwi results. Table 11 contains the BLEU scores. These additional scores offer further perspectives on the generalization capabilities of our MT-R1-Zero models across the evaluated OOD scenarios.
+
+# J Details in Disentangling RL and Explicit Thinking
+
+The ablation study discussed in Section 5.4 aimed to determine whether the explicit <think> step or the underlying RL optimization primarily drives performance gains. We compared three training paradigms: 1) Supervised Fine-Tuning (SFT),
+
+<table><tr><td rowspan="2"></td><td colspan="2">Train</td><td colspan="5">Test</td></tr><tr><td>EN-ZH</td><td>ZH-EN</td><td>EN-ZH</td><td>ZH-EN</td><td>EN-JA</td><td>DE-EN</td><td>DE-ZH</td></tr><tr><td># of cases</td><td>6565</td><td>6565</td><td>997</td><td>1976</td><td>997</td><td>549</td><td>1012</td></tr><tr><td>Source</td><td colspan="2">WMT 17-20</td><td>WMT 24</td><td>WMT 23</td><td>WMT 24</td><td>WMT 23</td><td>Flores</td></tr></table>
+
+Table 8: Data statistics for the training and test sets used in the main experiments (EN=ZH).  
+
+<table><tr><td rowspan="2"></td><td colspan="5">Parallel Data</td></tr><tr><td>Train (from EN)</td><td>Train (to EN)</td><td>Test (from EN)</td><td>Test (to EN)</td><td>Resource</td></tr><tr><td>Afrikaans (AF)</td><td>2994</td><td>341</td><td>1012</td><td>1012</td><td>Mid</td></tr><tr><td>Danish (DA)</td><td>2994</td><td>355</td><td>1012</td><td>1012</td><td>Mid</td></tr><tr><td>Dutch (NL)</td><td>2994</td><td>403</td><td>1012</td><td>1012</td><td>High</td></tr><tr><td>German (DE)</td><td>7015</td><td>885</td><td>1012</td><td>1012</td><td>High</td></tr><tr><td>Icelandic (IS)</td><td>4994</td><td>678</td><td>1012</td><td>1012</td><td>Low</td></tr><tr><td>Norwegian (NO)</td><td>2994</td><td>360</td><td>1012</td><td>1012</td><td>Low</td></tr><tr><td>Swedish (SV)</td><td>2994</td><td>339</td><td>1012</td><td>1012</td><td>High</td></tr><tr><td>Chinese (ZH)</td><td>6906</td><td>874</td><td>1012</td><td>1012</td><td>High</td></tr><tr><td>English (EN)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr></table>
+
+Table 9: Parallel data statistics for languages used in multilingual experiments (Section D), detailing training/test pairs and resource level classification.  
+
+<table><tr><td rowspan="2">MODEL</td><td colspan="4">OUT-OF-DISTRIBUTION</td></tr><tr><td>EN-JA</td><td>DE-EN (DOC)</td><td>DE-ZH</td><td>Avg.</td></tr><tr><td colspan="5">Strong Baseline</td></tr><tr><td>Qwen2.5-72B-Instruct</td><td>73.25</td><td>69.13</td><td>69.89</td><td>70.76</td></tr><tr><td>LLaMA3.1-70B-Instruct</td><td>71.84</td><td>69.28</td><td>68.67</td><td>69.93</td></tr><tr><td colspan="5">Same-size Baseline</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>64.79</td><td>67.20</td><td>67.82</td><td>66.60</td></tr><tr><td>LLaMA-3.1-8B-Instruct</td><td>62.42</td><td>66.77</td><td>64.28</td><td>64.49</td></tr><tr><td>TowerInstruct-7B-v0.2</td><td>58.33</td><td>69.03</td><td>65.45</td><td>64.27</td></tr><tr><td>Aya-23-8B</td><td>66.44</td><td>67.21</td><td>63.40</td><td>65.68</td></tr><tr><td>MT-R1-Zero-7B-Lex</td><td>63.33</td><td>66.17</td><td>64.32</td><td>64.61</td></tr><tr><td>MT-R1-Zero-7B-Sem</td><td>72.00</td><td>68.41</td><td>71.51</td><td>70.64</td></tr><tr><td>MT-R1-Zero-7B-Mix</td><td>69.27</td><td>68.74</td><td>70.25</td><td>69.42</td></tr></table>
+
+Table 10: Out-of-distribution performance comparison using the COMETKiwi metric on EN-JA, DE-EN (Doc), and DE-ZH. (Complements Table 2).  
+
+<table><tr><td rowspan="2">MODEL</td><td colspan="4">OUT-OF-DISTRIBUTION</td></tr><tr><td>EN-JA</td><td>DE-EN (Doc)</td><td>DE-ZH</td><td>Avg.</td></tr><tr><td colspan="5">Strong Baseline</td></tr><tr><td>Qwen2.5-72B-Instruct</td><td>25.02</td><td>45.54</td><td>40.83</td><td>37.13</td></tr><tr><td>LLaMA3.1-70B-Instruct</td><td>24.64</td><td>45.98</td><td>37.85</td><td>36.16</td></tr><tr><td colspan="5">Same-size Baseline</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>18.91</td><td>41.17</td><td>35.25</td><td>31.78</td></tr><tr><td>LLaMA-3.1-8B-Instruct</td><td>16.22</td><td>40.28</td><td>31.08</td><td>29.19</td></tr><tr><td>TowerInstruct-7B-v0.2</td><td>10.52</td><td>43.40</td><td>34.74</td><td>29.55</td></tr><tr><td>Aya-23-8B</td><td>20.20</td><td>37.28</td><td>31.86</td><td>29.78</td></tr><tr><td>MT-R1-Zero-7B-Lex</td><td>14.94</td><td>40.01</td><td>37.00</td><td>30.65</td></tr><tr><td>MT-R1-Zero-7B-Sem</td><td>14.12</td><td>33.19</td><td>22.83</td><td>23.38</td></tr><tr><td>MT-R1-Zero-7B-Mix</td><td>20.27</td><td>43.17</td><td>32.31</td><td>31.92</td></tr></table>
+
+serving as a non-RL baseline (further SFT training details can be found in Appendix D); 2) our MT-R1-Zero-Mix (RL w/ thinking), which incorporates explicit `<think>`/ `<translate>` prompting and is trained with the Reward-Mix signal; and 3) an RL w/o thinking variant, trained with the same Reward-Mix RL optimization but without prompts enforcing the `<think>` step.
+
+Table 11: Out-of-distribution performance comparison using the BLEU metric on EN-JA, DE-EN (Doc), and DE-ZH. (Complements Table 2).  
+
+<table><tr><td rowspan="2">MODEL</td><td colspan="4">DRT TEST SET</td></tr><tr><td>BLEU</td><td>COMETKIWI-22</td><td>XCOMET</td><td>Avg.</td></tr><tr><td>Qwen2.5-7B-Instruct</td><td>24.17</td><td>69.66</td><td>61.84</td><td>51.89</td></tr><tr><td>TowerInstruct-13B</td><td>22.71</td><td>70.55</td><td>62.77</td><td>52.01</td></tr><tr><td>DRT-7B</td><td>35.51</td><td>71.77</td><td>68.40</td><td>58.56</td></tr><tr><td>DRT-14B</td><td>36.37</td><td>72.15</td><td>69.64</td><td>59.39</td></tr><tr><td>Qwen2.5-7B (SFT)</td><td>21.61</td><td>69.91</td><td>63.20</td><td>51.57</td></tr><tr><td>Qwen2.5-7B (RL w/o thinking)</td><td>28.44</td><td>72.92</td><td>66.17</td><td>55.84</td></tr><tr><td>Qwen2.5-7B (RL w/ thinking)</td><td>28.42</td><td>73.20</td><td>66.64</td><td>56.09</td></tr></table>
+
+Table 12: Performance comparison on the DRT literature translation dataset (Wang et al., 2024a) using BLEU, COMETKiwi-22, and XCOMET metrics.
+
+The main results from this ablation (Table 3) indicate that both RL configurations ("RL w/ thinking" and "RL w/o thinking") achieve comparable performance and substantially outperform SFT, suggesting the RL framework itself is the main driver of improvements. This core conclusion is further corroborated by evaluations on the DRT literature translation benchmark (Wang et al., 2024a). Table 12 consistently shows that both RL variants perform similarly to each other and are significantly
+
+Reference (EN): It's encouraging to see, although there is some disappointment
+
+# Qwen2.5-7B-Base
+
+User:好看，但有点点小遗憾
+
+Assistant: <think> I translated the Chinese text "User:好看，但有点点小遗憾" to English as "It looks good, but there is a tiny bit of disappointment." The user is saying that the object or situation being discussed appears nice, but there is a small aspect that they are not completely satisfied with. </think>
+
+<translate> "It looks good, but there is a tiny bit of disappointment.</translate>
+
+# LLaMA-3.1-8B-Base
+
+User:好看，但有点点小遗憾
+
+Assistant: <think> reasoning to generate translation </think><translate>"It's beautiful, but there are a few minor drawbacks." </translate>
+
+# Qwen2.5-7B-Instruct
+
+<im_start>user 好看，但有点点小遗憾 <im_end>
+<im_start>assistant <think> The user is asking to translate the Chinese phrase "好看，但有点点小遗憾" to English. Let's break down the phrase:
+
+- "好看" means "it looks good" or "it is attractive."
+- "但" is a conjunction that means "but."  
+- "有点点小遗憾" means "there is a slight bit of regret" or "there is a small sense of disappointment."
+
+Putting it all together, the phrase suggests that something looks good, but there is a small element of disappointment or regret. A possible translation could be: "It looks good, but there's a tiny bit of disappointment."
+
+To ensure the translation captures the nuance, we can refine it to: "It looks good, but there's a slight sense of regret." </think><translate> "It looks good, but there's a slight sense of regret." </translate>
+
+# LLaMA-3.1-8B-Instruct
+
+<start_header_id>user</end_header_id>  
+好看，但有点点小遗憾
+
+<eot_id> <start_header_id> assistant <end_header_id>
+
+<think> "It's nice, but there's a little bit of a disappointment." </think> <translate> "It's nice, but there's a little bit of a disappointment." </translate>
+
+# TowerBase-7B
+
+User:好看，但有点点小遗憾
+
+Assistant: <think> Reasoning
+
+</think><translate>  
+Translation
+
+</translate>
+
+# TowerInstruct-7B
+
+<im_start>user好看，但有点点小遗憾<im_end>
+
+<im_start> assistant <think>好看，但有点点小遗憾
+
+</think><translate> It was beautiful, but there was a small disappointment:</translate>
+
+Figure 10: Qualitative comparison of final outputs from different starting models trained with MT-R1-Zero for the Chinese input "好看,但有点点小遗憾", illustrating varying degrees of format adherence and reasoning generation, including format hacking by some models (Finding 3).
+
+better than SFT on this distinct literary domain. This reinforces our main finding across different benchmarks, underscoring that the RL optimization process, rather than the mere structural prompting for explicit thinking steps, is the key factor behind the enhanced performance of MT-R1-Zero.
+
+# K SFT Training Details
+
+For the Supervised Fine-Tuning (SFT) baseline compared in the ablation study (Section 5.4), we utilized LLaMA-Factory (Zheng et al., 2024). The SFT process started from the same base model architecture as the corresponding RL experiments (e.g., Qwen2.5-7B) and was performed on the identical parallel translation dataset (13,130 examples from WMT 2017-2020 after filtering, detailed in Section 4.1). The model was fine-tuned on 8 NVIDIA H800 80G GPUs for 2 epochs using a learning rate of 5e-6 and a batch size of 64, totaling approximately 400 training steps.
+
+User:"It's not that bad, right, Kayel?"
+
+# Assistant: <think>
+
+When approaching the task of translating the given English text into Japanese, my primary goal is to create a faithful and engaging rendition that maintains the structure, tone, and level of detail of the original while ensuring it reads smoothly and naturally in Japanese. I will pay particular attention to preserving the conversational nature of the exchange and the use of question tags ("right, Kayel?"), as well as the balance between the two characters' statements.
+
+In this case, I will consider the following aspects:
+
+1. Maintaining the conversational tone: This involves using appropriate phrasing, word order, and intonation to convey the casual nature of the conversation.
+
+2. Translating question tags: Japanese question tags are typically formed differently from English, so I will find an equivalent structure that maintains the flow of the conversation.
+
+3. Preserving the balance between the two characters' statements: I
+
+will ensure that the translation reflects the equal weight given to both Kayel's response and the speaker's initial statement in the original text.
+
+4. Adapting to Japanese sentence structure: Japanese often places the verb at the end of a sentence, so I will adjust the word order accordingly while maintaining the overall meaning and rhythm of the conversation.
+
+</think>
+
+<translate>
+
+“大丈夫！也实在，久工兀？”
+
+</translate>
+
+# User:好看，但有点点小遗憾
+
+# Assistant: <think>
+
+The task at hand involves translating a sentence of Chinese into English while taking into consideration several factors to ensure an effective and appropriate translation. The source text is: “好看，但有点点小遗憾。” Given the context in which this sentence is being used, it is clear that the speaker is expressing their positive opinion of something they have just seen or experienced, but also acknowledging a minor flaw or disappointment that slightly
+
+diminishes their overall satisfaction. With this in mind, it is important to maintain the same level of nuance and complexity in the translation to effectively convey the same level of emotional and intellectual engagement expected in professional and polished English communication. After careful consideration of these factors and the specific requirements of the translation context, I have decided to render the source text as follows: "It looks good, but there's a tiny bit of disappointment." This translation not only accurately conveys the same level of appreciation and critique as the source text but also does so in a professional and tactful manner that respects the sensitivity and complexity of the situation at hand.
+
+<translate>
+
+"It looks good, but there's a tiny bit of disappointment." </translate>
+
+User：北交所融资融券交易的启动是持续推进北交所市场制度建设、完善交易机制、提升二级市场定价功能的重要举措。
+
+# Assistant: <think>
+
+When approaching the task of translating the given Chinese text into English, my primary goal is to create a faithful and engaging rendition that maintains the structure, tone, and level of detail of the original Chinese. I will focus on preserving the balance between providing all relevant information and maintaining readability, while also considering the natural flow of the English language. Additionally, I will pay attention to the following aspects:
+
+1. Structure: I will aim to maintain the overall structure of the Chinese text, including the use of paragraphs and sentence order, to ensure that the English translation remains coherent and easy to follow.
+
+2. Tone and voice: I will strive to capture the tone and voice of the original text, which appears to be informative and objective, while also considering the appropriate level of formality for the target audience.
+
+3. Terminology: I will use accurate and up-to-date terminology related to the stock exchange, financing, and trading mechanisms, ensuring that all terms are clear and easily understood by the English-speaking audience.
+
+4. Cultural and linguistic nuances: I will pay attention to any cultural or linguistic nuances present in the Chinese text and find suitable ways to convey them in English, where possible, without compromising the clarity or readability of the translation.
+
+5. Number and date formats: I will convert any numbers and date formats from the Chinese system to the corresponding English system, as necessary, to ensure that the translation is both accurate and locally appropriate.
+
+6. Punctuation and formatting: I will use standard English punctuation and formatting conventions, while taking into account any specific requirements or preferences for the target audience or publication platform.
+
+By considering these aspects, I aim to produce a high-quality English translation that effectively communicates the content and intent of the original Chinese text. </think>
+
+<translate>
+
+"The launch of financing and margin trading on the Beijing Stock Exchange is an important step in advancing the development of the exchange's market systems, improving trading mechanisms, and enhancing the secondary market's pricing capabilities." </translate>
+
+# User：加拿大也报告了类似的结果，只是数字较低。
+
+# Assistant: <think>
+
+When approaching the task of translating the given Chinese text into English, my primary goal is to create a faithful and engaging rendition that maintains the structure, tone, and level of detail of the original Chinese while ensuring it reads smoothly and naturally in English. I will focus on preserving the balance between the two sentences, maintaining the use of parallel structure where appropriate, and accurately conveying the nuances of the Chinese phrasing, including the use of numbers and the comparative nature of the information presented. </think> <translate> "Canada also reported similar results, although the numbers were lower." </translate>
+
+User: to divide the 412.5 meters long, 12.9 meters wide, with a total weight of 15,000 tons of continuous beams into 96 sections.
+
+Assistant: <think> To translate the given English sentence into Chinese, I would follow these steps to ensure accuracy, fidelity to the original text, and naturalness in the target language (Chinese):
+
+1. Understand the context and meaning of the original text: I would first read and re-read the sentence to grasp its full meaning, paying close attention to technical terms, measurements, and the overall structure of the information being conveyed. This ensures a proper translation that maintains the original message.
+
+2. Identify key elements and vocabulary: I would identify important numbers, measurements, and specific terms that need to be accurately translated, such as "412.5 meters," "12.9 meters," and "15,000 tons." I would also note any technical terms related to engineering or construction, such as "continuous beams" and "sections."
+
+3. Consider cultural and contextual nuances: While this particular sentence doesn't contain cultural references that would require adaptation, I would generally be aware of how to translate phrases or expressions that might have different connotations or usages in Chinese culture compared to English.
+
+# 4. Choose an appropriate translation
+
+strategy: For this type of technical information, I would likely use a combination of literal translation and necessary modifications to fit Chinese syntax and readability. This might involve reordering words or phrases to sound more natural in Chinese while still maintaining the original meaning.
+
+# 5. Translate while considering the
+
+audience: Since this translation is presumably for a Chinese-speaking audience, I would ensure that the translation aligns with how Chinese speakers might express similar ideas, using idiomatic expressions and terminology that would be familiar to them.
+
+6. Account for register and tone: Depending on the intended use of the translation (e.g., technical manual, news report, etc.), I would adjust the formality and wording to match the appropriate register. For a technical report, the language would be more formal and precise.
+
+7. Translate numbers and measurements: I would carefully translate the numbers and measurements given (412.5 meters, 12.9 meters, 15,000 tons) to their equivalent in Chinese, ensuring that the units of measurement are accurately reflected in the translation (e.g., using "米" for meters and "吨" for tons).
+
+8. Ensure coherence and flow in the translation: After translating each part of the sentence, I would review and refine the translation to ensure that it flows naturally in Chinese, making any necessary adjustments to grammar, syntax, and phrasing to create a smooth and coherent translation.
+
+# 9. Review and make final adjustments:
+
+Finally, I would review the entire translated sentence for accuracy, fluency, and naturalness in Chinese. I would make any necessary adjustments to improve the translation, checking again against the original text to ensure that all information is preserved and that there are no errors or misunderstandings.
+
+By following these steps, I can provide a translation that is accurate, faithful to the original text, and reads naturally in Chinese.
+
+</think><translate>将总长412.5米、宽12.9米、总重15000吨的连续梁分成96节。</translate>
+
+Ⅲ
+
+Figure 11: Qualitative examples (I-V) showcasing the diverse thinking patterns generated by MT-R1-Zero models (Finding 2). In particular, we observe patterns ranging from structured multi-step decomposition (Types I-III) to more colloquial processing (Types IV-V).

EMNLP/2025/MT-R1-Zero_ Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ea4b4da3ee6f743fcd9651de08c1610508bf2d8e2b0e29d6904a4517bada9e1
+size 1343381

EMNLP/2025/MT-R1-Zero_ Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1d16d644b90d791c6f18123f8494d4d69b49b5e3be0e6c6d2e3a96859ec73d5
+size 559734

EMNLP/2025/MT-RewardTree_ A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling/c063f64e-107f-4de3-a3ba-2701c0b18208_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9cdae2d6c0fd67e179b3b55525d03a425c9808457e6c6762e566b6164ace9a8
+size 86910

EMNLP/2025/MT-RewardTree_ A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling/c063f64e-107f-4de3-a3ba-2701c0b18208_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f591dd886151e02c241212f9e1d201a6ed2d38a45527aa1db9d44b4848ad50c
+size 105804

EMNLP/2025/MT-RewardTree_ A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling/c063f64e-107f-4de3-a3ba-2701c0b18208_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0602fd4c3dde3e626d2853e88ff1fd7e902d2919db4335c5b624c0351b6f769c
+size 2061888

EMNLP/2025/MT-RewardTree_ A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling/full.md

@@ -0,0 +1,356 @@
+# MT-RewardTree: A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling
+
+Zhaopeng Feng $^{1\spadesuit}$  Jiahan Ren $^{1\spadesuit}$  Jiayuan Su $^{1\spadesuit}$  Jiamei Zheng $^{1}$
+
+Hongwei Wang $^{1\dagger}$  Zuozhu Liu $^{1\dagger}$
+
+<sup>1</sup>Zhejiang University
+
+{zhaopeng.23, jiahan.24, jiayuan.23, jiamei.24}@int1.zju.edu.cn
+
+{hongweiwang,zuozhuliu}@intl.zju.edu.cn
+
+# Abstract
+
+Process reward models (PRMs) have shown success in complex reasoning tasks for large language models (LLMs). However, their application to machine translation (MT) remains underexplored due to the lack of systematic methodologies and evaluation benchmarks. To address this gap, we introduce MT-RewardTree, a comprehensive framework for constructing, evaluating, and deploying process reward models in MT. Unlike traditional vanilla preference pair construction, we propose a novel method for automatically generating token-level preference pairs using approximate Monte Carlo Tree Search (MCTS), which mitigates the prohibitive cost of human annotation for fine-grained steps. Then, we establish the first MT-specific reward model benchmark and provide a systematic comparison of different reward modeling architectures, revealing that token-level supervision effectively captures fine-grained preferences. Experimental results demonstrate that our MT-PRM-Qwen-2.5-3B achieves state-of-the-art performance in both token-level and sequence-level evaluation given the same input prefix. Furthermore, we showcase practical applications where MT-PRMs successfully identify token-level translation differences and enable test-time alignment for LLMs without additional alignment training. Our work provides valuable insights into the role of reward models in MT research. Our code and data are released in https://sabijun.github.io/MT_RewardTreePage.
+
+# 1 Introduction
+
+The next-token prediction process in large language models (LLMs) is often modeled as a Markov Decision Process (MDP) and has achieved remarkable success across various domains, largely attributed
+
+![](images/1347664b8163a16a21fda12a5bf26029cf3bad87d4d1dd52747c437571744dc7.jpg)  
+Figure 1: Components of MT-RewardTree.
+
+to reinforcement learning (RL) and the scaling of test-time compute (Snell et al., 2024; Zeng et al., 2024; DeepSeek-AI et al., 2025; Team, 2025; Xiang et al., 2025). Reward models are central to these advancements. Outcome Reward Models (ORMs), which are designed to evaluate full responses, have been widely adopted; however, due to the sparsity of outcome rewards, ORMs often yield suboptimal performance and struggle with stability and efficiency during RL training (Lightman et al., 2024; Cao et al., 2024; Chan et al., 2024). In contrast, Process Reward Models (PRMs) evaluate intermediate steps to provide fine-grained guidance during both training and inference. PRMs have proven particularly effective in tasks such as mathematics and coding by guiding stepwise decision-making (Wang et al., 2024a; Chen et al., 2024; Luo et al., 2024; Qi et al., 2024; Guan et al., 2025).
+
+Machine translation (MT) naturally aligns with token-level MDP frameworks, as each translation decision corresponds directly to token generation. However, there is still a lack of systematic methodologies for constructing and evaluating PRMs in MT, which has hindered progress relative to advancements in general-domain LLMs.
+
+Developing effective PRMs is challenging. Although Lightman et al. (2024) demonstrate that
+
+process supervision with human annotators improves PRM performance in mathematical tasks, this methodology requires domain-expert annotators, resulting in prohibitive costs and practical limitations for translation tasks. Recently, some studies suggest that a PRM can be automatically learned during Direct Preference Optimization (DPO) training (Rafailov et al., 2024b,a; Yuan et al., 2024). However, existing vanilla preference pair datasets provide only sequence-to-sequence preference data, rather than token-level preferences, which raises concerns about their applicability for token-level alignment. Additionally, evaluating PRMs remains a significant challenge. In mathematical tasks, evaluation is often done using a Best-of-N (BoN) sampling strategy—selecting the highest-scored response from N candidates based on a PRM (Lightman et al., 2024; Wang et al., 2024b; Luo et al., 2024)—or by having the PRM identify errors or verify correctness in the steps (Zheng et al., 2024; Zhang et al., 2025). Since each step in mathematics has a deterministic answer, these methods do not directly translate to PRM evaluation in MT.
+
+In this paper, we introduce MT-RewardTree, a comprehensive framework for constructing, evaluating, and deploying PRMs in machine translation. We propose an approximate Monte Carlo Tree Search (MCTS) method (Kocsis and Szepesvári, 2006; Silver et al., 2016) to generate the token-level preference pair dataset. This dataset is then split into a training set for reward model development and a benchmark for reward evaluation. We provide a systematic comparison of different reward modeling methods and test on both token-level and sequence-level performance. Furthermore, we demonstrate two practical applications of PRMs, offering valuable insights for future MT research. Our main contributions are as follows:
+
+- We introduce MT-RewardTree, a comprehensive framework for the construction, evaluation, and deployment of PRMs in MT. We establish the first dedicated reward benchmark - MTPRMBench. Our experiments demonstrate that MT-PRMs achieve competitive performance on both token-level and sequence-level evaluations.  
+- Comprehensive experiments indicate that our token-level preference pairs, generated through an approximate MCTS method, significantly outperform vanilla preference pairs in process reward model training. Furthermore, our analysis validates that supervising PRMs using
+
+preference-based signals is more effective than direct supervision with absolute value estimates.
+
+- We demonstrate that our MT-PRMs can directly identify token-level translation differences and facilitate test-time alignment for LLM-based MT without the need for additional alignment training, offering valuable practical insights for the application of reward models in MT.
+
+# 2 Background
+
+# 2.1 Token-level Markov Decision Process
+
+LLMs' autoregressive generation can be naturally formulated as a Markov Decision Process, where each token generation is treated as an action. At each time step  $t$ , an action  $a_{t}$  corresponds to the generation of a new token, and the state  $\mathbf{s}_t$  is represented as the sequence of tokens generated up to that point. For tasks that do not involve interaction with an external environment—such as translation—the state is defined as
+
+$$
+\mathbf {s} _ {t} = (x _ {0}, \dots , x _ {L}, y _ {0}, \dots , y _ {t - 1}),
+$$
+
+where  $(x_0,\ldots ,x_L)$  represents the input prompt and  $(y_0,\dots ,y_{t - 1})$  is the sequence of generated tokens until time step  $t - 1$ . The state transition function  $f$  is deterministic and updates the state by concatenating the newly generated token:
+
+$$
+\mathbf {s} _ {t + 1} = f (\mathbf {s} _ {t}, a _ {t}) = \mathbf {s} _ {t} \mid a _ {t},
+$$
+
+with | denoting concatenation.
+
+Within this token-level MDP framework, the reward function  $r(\mathbf{s}_t, a_t)$  is typically designed to provide feedback only at the terminal time step  $T$ , reflecting the overall correctness of the generated sequence or the successful completion of the task. To optimize the policy  $\pi_{\theta}$  based on this reward, Reinforcement Learning with Human Feedback (RLHF) (Ouyang et al., 2022) typically maximizes a KL-constrained objective:
+
+$$
+\mathbb {E} _ {(s _ {0}, \dots , s _ {T}) \sim \rho_ {\pi}} \left[ \sum_ {t = 0} ^ {T} \left(r (s _ {t}, a _ {t}) - \beta \log \frac {\pi \left(a _ {t} \mid s _ {t}\right)}{\pi_ {\mathrm {r e f}} \left(a _ {t} \mid s _ {t}\right)}\right) \right], \tag {1}
+$$
+
+where  $\pi_{\mathrm{ref}}$  is a pre-trained reference policy,  $\beta$  controls the strength of the KL penalty and  $\rho_{\pi}$  denotes the trajectory distribution induced by policy  $\pi$ .
+
+In practice, classical RLHF applies the reward solely at the terminal state. Specifically, the reward function used in Proximal Policy Optimization (PPO) (Schulman et al., 2017) is defined as:
+
+$$
+r \left(s _ {t}, a _ {t}\right) = \left\{ \begin{array}{l l} \beta \log \pi_ {\text {r e f}} \left(a _ {t} \mid s _ {t}\right), & \text {i f} s _ {t + 1} \text {i s n o n - t e r m i n a l ,} \\ r (x, y) + \beta \log \pi_ {\text {r e f}} \left(a _ {t} \mid s _ {t}\right), & \text {i f} s _ {t + 1} \text {i s t e r m i n a l .} \end{array} \right. \tag {2}
+$$
+
+# 2.2 Reward Modeling in RLHF
+
+Reward modeling is the cornerstone of RLHF, enabling LLMs to align their outputs with human preferences. In this section, we distinguish between typical (sequence-level) reward modeling and the more fine-grained token-level approach.
+
+Sequence-level Reward Modeling. In classical RLHF, the reward function is learned from human feedback on prompt-response pairs  $(\mathbf{x},\mathbf{y})$ . The reward model is formulated as a contextual bandit, where a scalar reward is assigned only at the terminal state—i.e., once the full response sequence has been generated. This formulation, known as Outcome Reward Modeling, follows the Bradley-Terry (Bradley and Terry, 1952) preference model to define the probability of preferring one response over another:
+
+$$
+p ^ {*} (\mathbf {y} ^ {w} \succeq \mathbf {y} ^ {l}) = \frac {\exp \left(r _ {\phi} (\mathbf {x} , \mathbf {y} ^ {w})\right)}{\exp \left(r _ {\phi} (\mathbf {x} , \mathbf {y} ^ {w})\right) + \exp \left(r _ {\phi} (\mathbf {x} , \mathbf {y} ^ {l})\right)}. \tag {3}
+$$
+
+To train the reward model  $r_{\phi}$ , we construct a preference dataset  $\mathcal{D}$ , where each prompt  $x$  is paired with two candidate responses,  $y$  and  $y'$ . Human annotators or heuristics determine the preferred response  $y_{w}$  and the rejected response  $y_{l}$ . The reward model is then optimized to maximize the likelihood of these human preferences:
+
+$$
+\left. \max  _ {\phi} \mathbb {E} _ {(x, y _ {w}, y _ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(r _ {\phi} (x, y _ {w}) - r _ {\phi} (x, y _ {l})\right) \right], \right. \tag {4}
+$$
+
+where  $\sigma$  is the logistic function. By training  $r_{\phi}$  in this manner, we ensure that the model assigns higher rewards to preferred responses, effectively capturing human-like quality judgments for sequence-level evaluation.
+
+Token-level Reward Modeling. While the sequence-level approach treats the entire generated response as a single action, it fails to capture the fine-grained decision-making process inherent in token generation. Token-level reward modeling addresses this limitation by evaluating rewards at each token-generation step. This approach corresponds to a form of Process Reward Models. The cumulative reward for a trajectory  $\tau$  is computed as the sum of per-token rewards, and the corresponding preference probability between two trajectories,  $\tau^w$  and  $\tau^l$ , is given by:
+
+$$
+p ^ {*} (\tau^ {w} \succeq \tau^ {l}) = \frac {\exp \left(\sum_ {i = 1} ^ {N} r \left(\mathbf {s} _ {i} ^ {w} , a _ {i} ^ {w}\right)\right)}{\exp \left(\sum_ {i = 1} ^ {N} r \left(\mathbf {s} _ {i} ^ {w} , a _ {i} ^ {w}\right)\right) + \exp \left(\sum_ {i = 1} ^ {M} r \left(\mathbf {s} _ {i} ^ {l} , a _ {i} ^ {l}\right)\right)}. \tag {5}
+$$
+
+Although token-level reward modeling offers finer-grained feedback, obtaining effective PRMs is more challenging to obtain and deploy (Lightman et al., 2024; Cao et al., 2024).
+
+# 3 MT-RewardTree
+
+In this section, we introduce the components of the MT-RewardTree. We first describe how we construct token-level preference pairs using an MCTS-based method. Next, we review several approaches employed for reward modeling.
+
+# 3.1 Constructing Token-level Preference Pairs
+
+Prior studies have investigated translation preference pair construction (Xu et al., 2024; Agrawal et al., 2024; Feng et al., 2024a), yet a standardized token-level preference pair dataset for PRMs in MT remains absent. MQM (Freitag et al., 2021) datasets depend on manual error annotation, which is both cost-prohibitive and incapable of producing granular token-level preference pairs.
+
+Drawing inspiration from MCTS, we propose a token-centric approach that quantifies token quality based on its potential to contribute to higher-quality translations. This method aligns with Monte Carlo-based PRMs construction techniques in mathematics, where step-wise quality is determined by its incremental contribution to deriving correct answers (Wang et al., 2024b; Guan et al., 2025).
+
+The MCTS process consists of four main steps (depicted in Figure 2): Selection, Expansion, Simulation (Evaluation), and Back-propagation.
+
+1. Selection: The first phase involves selecting a portion of the existing tree that is most promising for further expansion. Starting from the root node, a standard approach would traverse the tree down to a leaf using the PUCT algorithm (Rosin, 2011; Silver et al., 2017). Since our goal is to construct token-level preference pairs rather than achieving global optimality, we automatically select the existing prompt and previously generated tokens as the prefix  $y_{<t}$ .  
+2. Expansion: If the selected leaf node is not an EOS (end-of-sentence) token—i.e. if it is not a terminal state—the node is expanded by generating  $k$  candidate children. This is achieved by decoding one additional step using the language model and selecting the top-  $k$  tokens as the new children. We select the top-2 candidate tokens  $a_{tj}$  (with  $j \in \{1, 2\}$ ) that have the highest logits. Preliminary experiments demonstrate that tokens outside of the top-2 yield significantly lower translation quality during the Simulation phase. These top-2 tokens, sharing the same
+
+![](images/84cbf34c658affb565b7ab21e0d2bc18d9f93aef30302930bb36609e92739040.jpg)  
+Figure 2: The construction process of token-level preference pairs. We utilize TowerInstruct-7B-v0.2 to generate candidate tokens. A token-level preference pair comprises two translations that share an identical prefix.
+
+prefix  $y_{< t}$ , form the basis for our token-level preference pair.
+
+3. Simulation (Evaluation): From each expanded node  $a$ , we generate  $n$  complete translation rollouts until an EOS token is reached. We then evaluate the quality (or groundedness) of the full translation sequence, denoted by  $g(y,n)$ . In our framework, we use COMETKiwi (Rei et al., 2022) to estimate the quality of all  $n$  full rollouts. These scores are averaged and further assigned as the value of node  $a$ , i.e.,  $V(a)$ .  
+4. Back-propagation: Since our objective is to construct token-level preference pairs, we compare the values  $V(a_{t1})$  and  $V(a_{t2})$  to determine which expanded token is superior. Finally, we retain the node with the higher  $V$  value. This node, along with its corresponding prefix  $y_{<t}$ , is then used as the starting point in the next simulation cycle, beginning again at Step 1.
+
+These four steps are repeated until the EOS token appears during the Selection phase. We retain one rollout from the superior token and one from the inferior to construct our token-level preference pair. We use COMETKiwi to guarantee the score gap lies between 0.04 and 0.4 to control the quality.
+
+# 3.2 Implicit Process Reward Modeling
+
+Unlike ORMs, which assign a single reward to the entire response, PRMs aim to assign rewards
+
+at a finer granularity, such as at each step or token. However, traditional PRMs training requires step-level annotations, which are costly to obtain. Recent studies (Rafailov et al., 2024a; Zhong et al., 2024) show that ORMs can be trained with implicit reward modeling, enabling PRMs to emerge naturally without the need for explicit step labels.
+
+Consider an ORM where the reward is parameterized by the log-likelihood ratio of two causal language models:
+
+$$
+r _ {\theta} (\mathbf {y}) := \beta \log \frac {\pi_ {\theta} (\mathbf {y})}{\pi_ {\mathrm {r e f}} (\mathbf {y})} \tag {6}
+$$
+
+where  $\pi_{\theta}$  represents the trained model's probability distribution, and  $\pi_{\mathrm{ref}}$  is a reference model. We define the cumulative reward up to step  $t$  as:
+
+$$
+q _ {\theta} ^ {t} \left(\mathbf {y} _ {<   t}, y _ {t}\right) := \sum_ {i = 1} ^ {t} \beta \log \frac {\pi_ {\theta} \left(y _ {i} \mid \mathbf {y} _ {<   i}\right)}{\pi_ {\text {r e f}} \left(y _ {i} \mid \mathbf {y} _ {<   i}\right)} \tag {7}
+$$
+
+which serves as an exponential moving average of  $r_{\theta}$  across steps. The expected process reward at step  $t$  can then be expressed as:
+
+$$
+q _ {\theta} ^ {t} (\mathbf {y} _ {<   t}, y _ {t}) = \beta \log \mathbb {E} _ {\pi_ {\text {r e f}} (\mathbf {y} | \mathbf {y} _ {\leq t})} \left[ e ^ {\frac {1}{\beta} r _ {\theta} (\mathbf {y})} \right] \tag {8}
+$$
+
+This formulation shows that  $q_{\theta}^{t}$  is an exact expectation of the outcome reward  $r_{\theta}$  at step  $t$ , making it analogous to a Q-value in reinforcement learning.
+
+By defining the process reward  $r_{\theta}^{t}$  as the difference between successive Q-values:
+
+$$
+r _ {\theta} ^ {t} := q _ {\theta} ^ {t} - q _ {\theta} ^ {t - 1} = \beta \log \frac {\pi_ {\theta} \left(y _ {t} \mid \mathbf {y} <   t\right)}{\pi_ {\mathrm {r e f}} \left(y _ {t} \mid \mathbf {y} <   t\right)} \tag {9}
+$$
+
+<table><tr><td rowspan="3">Model</td><td colspan="6">MT-PRMBench</td></tr><tr><td colspan="3">Sequence-level</td><td colspan="3">Token-level</td></tr><tr><td>EN→XX</td><td>XX→EN</td><td>Avg.</td><td>EN→XX</td><td>XX→EN</td><td>Avg.</td></tr><tr><td colspan="7">Baselines</td></tr><tr><td>Skywork-Reward-LLaMA-3.1-8B</td><td>0.857</td><td>0.773</td><td>0.815</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MT-Ranker-base</td><td>0.785</td><td>0.787</td><td>0.786</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MT-Ranker-large</td><td>0.847</td><td>0.873</td><td>0.860</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan="7">PRMs</td></tr><tr><td>MT-PRM-LLaMA-3.2-3B</td><td>0.777</td><td>0.775</td><td>0.776</td><td>0.542</td><td>0.615</td><td>0.578</td></tr><tr><td>MT-PRM-Qwen-2.5-3B</td><td>0.867</td><td>0.858</td><td>0.863</td><td>0.637</td><td>0.685</td><td>0.660</td></tr></table>
+
+Table 1: Accuracy results on MT-PRMBench. Skywork-Reward-LLaMA-3.1-8B is an advanced ORM for general domains, while MT-Ranker represents the SoTA non-metric reference-free translation quality estimation model.  
+
+<table><tr><td rowspan="2">Translation Direction</td><td colspan="2">Token-level Preference Pairs</td></tr><tr><td>Train</td><td>MT-PRMBench</td></tr><tr><td>DE-EN</td><td>1,255</td><td>200</td></tr><tr><td>EN-DE</td><td>2,059</td><td>200</td></tr><tr><td>RU-EN</td><td>1,219</td><td>200</td></tr><tr><td>EN-RU</td><td>1,711</td><td>200</td></tr><tr><td>ZH-EN</td><td>1,232</td><td>200</td></tr><tr><td>EN-ZH</td><td>1,176</td><td>200</td></tr></table>
+
+Table 2: Data Statistics.
+
+We see that PRMs can be derived directly from an ORM trained on response-level data, without requiring explicit step-wise labels. This insight suggests that training an ORM inherently leads to the learning of a Q-function, enabling step- or token-level reward modeling without requiring additional supervision. A typical example of this is DPO (Rafailov et al., 2024b), which optimizes the following objective:
+
+$$
+L _ {\mathrm {D P O}} \left(\pi_ {\theta}; \pi_ {\text {r e f}}\right) = - \mathbb {E} _ {\left(x, y _ {w}, y _ {l}\right) \sim \mathcal {D}} \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(y _ {w} \mid x\right)}{\pi_ {\text {r e f}} \left(y _ {w} \mid x\right)} - \beta \log \frac {\pi_ {\theta} \left(y _ {l} \mid x\right)}{\pi_ {\text {r e f}} \left(y _ {l} \mid x\right)}\right) \tag {10}
+$$
+
+This formulation shows that optimizing  $\pi_{\theta}$  implicitly optimizes a reward model, as described in Eq. 4. Moreover, Yuan et al. (2024) demonstrated that this approach is agnostic to the specific training objective (i.e., not limited to DPO). It can be instantiated using various training objectives (e.g., KTO (Ethayarajh et al., 2024)), with the only modification being the substitution of  $r_{\theta}(\mathbf{y})$  with  $\beta \log \frac{\pi_{\theta}(\mathbf{y})}{\pi_{\mathrm{ref}}(\mathbf{y})}$ .
+
+Moreover, our implicit PRMs can seamlessly be converted into ORMs using weighted implicit rewards:
+
+$$
+r _ {\text {s e q u e n c e}} \left(y _ {1: T}\right) = \sum_ {k = 0} ^ {T - 1} w _ {t} \log \frac {\pi_ {\theta} \left(y _ {t} \mid y _ {<   t}\right)}{\pi_ {\text {r e f}} \left(y _ {t} \mid y _ {<   t}\right)} \tag {11}
+$$
+
+where the positional weights  $w_{t} = \frac{1}{|y < t|}$  are used
+
+to balance the contributions of each token.  
+
+<table><tr><td>Preference Pair Type</td><td>Training Strategy</td><td>Avg.</td></tr><tr><td>Token-level</td><td>DPO</td><td>0.660</td></tr><tr><td>Vanilla</td><td>DPO</td><td>0.574</td></tr><tr><td>Token-level</td><td>KTO</td><td>0.644</td></tr><tr><td>Vanilla</td><td>KTO</td><td>0.562</td></tr></table>
+
+Table 3: Ablation study on the effect of training preference data and implicit reward training objectives. The backbone model is Qwen-2.5-3B-Instruct and we test these variants on MT-PRMBench (Token-level).
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+Datasets. We explore four languages—English (EN), German (DE), Chinese (ZH), and Russian (RU)—and six translation directions: EN→XX and XX→EN. Our raw corpus consists of test sets from WMT17 to WMT20, supplemented with development and test sets from the Flores (Costa-jussa et al., 2022). We use the TowerInstruct-7B-v0.2 model with a temperature of 0.95 and apply the MCTS-based approach described earlier. During the Simulation step, we sample three candidate hypotheses for each node. Our token-level preference pairs is divided into train and test set (MTPRMBench). Detailed statistics are in Table 2.
+
+Training Details. We take LLaMA-3.2-3B-Instruct and Qwen-2.5-3B-Instruct as the backbone models for training. For the DPO training, the higher-scored sentence is designated as the chosen response, while the lower-scored sentence is labeled as the rejected response. For the KTO training, the higher-scored sentence is treated as the positive sample, and the lower-scored sentence as the negative sample. We set  $\beta$  as 0.1.
+
+Reward Evaluation. We evaluate reward models by framing the task as a classification problem, similar to prior work on reward model benchmarks in the general domain (Lambert et al., 2024; Liu et al., 2024). For sequence-level evaluation, given a tuple  $(x,y_{c},y_{r})$ , where  $x$  is the prompt,  $y_{c}$  is the chosen response, and  $y_{r}$  is the rejected response, the reward model predicts whether  $y_{c}$  is better than  $y_{r}$ . If the reward model assigns a higher reward to  $y_{c}$  than to  $y_{r}$ , the prediction is correct; otherwise, it is incorrect. We use accuracy as the evaluation metric, computed as follows:
+
+$$
+\text {A c c u r a c y} = \frac {1}{| D |} \sum_ {\left(x, y _ {c}, y _ {r}\right) \in D} I \left[ R _ {\theta} \left(x, y _ {c}\right) > R _ {\theta} \left(x, y _ {r}\right) \right] \tag {12}
+$$
+
+where  $I(\cdot)$  is the indicator function, and  $D$  denotes the evaluation dataset.
+
+For token-level evaluation, we use tuples of the form  $(x,y_{< t} + a_c,y_{< t} + a_r)$ , where  $y_{< t}$  is the generated tokens before,  $a_{c}$  is the next chosen token, and  $a_{r}$  is the rejected token. Similarly, we compute accuracy as the evaluation score: if the PRM assigns a higher reward to  $a_{c}$  than to  $a_{r}$ , the prediction is correct; otherwise, it is incorrect.
+
+MT-PRMBench. MT-PRMBench comprises two distinct subsets for evaluation: Token-level and Sequence-level. The Token-level subset is designed for assessing preferences between immediate next-token candidates that follow an identical input prefix. In contrast, the Sequence-level facilitates the comparison of entire generated sequence completions that also originate from a shared input prefix.
+
+# 4.2 Evaluation Results
+
+Token-level Performance. From Table 1, we can observe that our MT-PRM-LLaMA-3.2-3B and MT-PRM-Qwen-2.5-3B models achieved accuracies of 0.578 and 0.66 respectively on the token-level MT-PRMBench. As shown in Table 3, we systematically compare models trained with vanilla sequence-level preference pairs versus our token-level preference pairs, while evaluating both DPO and KTO training objectives. The results demonstrate that token-level preference pairs significantly improve discrimination accuracy: implicit PRMs trained with token-level preference pairs outperform vanilla sequence-level baselines by  $+8.6\%$  (DPO) and  $+11.5\%$  (KTO). This performance gap highlights the critical advantage of token-level preference pairs in helping capture fine-grained translation quality distinctions.
+
+Sequence-level Performance. We also convert our
+
+PRMs to sequence-level scoring through weighted DPO rewards (as shown in Eq. 11). We can observe that our MT-PRM-Qwen-2.5-3B achieves the highest performance among all models in the Prefix fixed set, with an average score of 0.863, outperforming both Skywork-Reward-LLaMA-3.1-8B and the MT-Ranker (Moosa et al., 2024) variants. This demonstrates the effectiveness of our token-level supervision framework even when adapted to sequence-level scoring.
+
+# 5 Analysis and Practical Insight
+
+# 5.1 Modeling Advantage versus Value as the PRM Training Signal
+
+We have explored the impact of vanilla preference pairs versus MCTS-generated token-level preference pairs on the performance of implicit PRMs in the previous experiments (Table 3). This section shifts focus to the nature of the supervisory signal used for training PRMs. Specifically, we analyze our choice of implicit reward modeling—derived from token-level preference pairs (termed "Supervised by Preference (SP)")—against the alternative of directly using MCTS-backpropagated values  $V(a)$  as a dense supervisory signal (termed "Supervised by Value (SV)").
+
+Table 4 demonstrates that SP yields markedly superior performance in token-level discrimination, where SP achieved an average accuracy of 0.66, while SV achieved 0.52. This significantly higher accuracy underscores the effectiveness of our implicit PRM when trained with preference-based supervision. The SV approach, directly regressing on MCTS-backpropagated  $V(a)$  values, tasks the PRM with learning a value function. However,  $V(a)$  as a dense, token-level reward faces limitations: Monte Carlo estimates can be noisy, and the absolute value of a partial translation may offer an indirect and insufficiently discriminative signal for the most recent token's quality, especially in complex MT scenarios or with imperfect rollouts—a recognized challenge in process supervision literature (Lightman et al., 2024). We also find that these  $V(a)$  values often clustered within a narrow numerical range (e.g., 0.7-0.8) further illustrates why this SV approach struggles. Such clustering severely hampers a regression model's ability to discern fine-grained distinctions, as the supervisory signal becomes subtle.
+
+<table><tr><td>MT-PRMBench (Token-level)</td><td>DE-EN</td><td>RU-EN</td><td>ZH-EN</td><td>EN-DE</td><td>EN-RU</td><td>EN-ZH</td><td>Avg.</td></tr><tr><td>Supervised by Value (SV)</td><td>0.52</td><td>0.56</td><td>0.48</td><td>0.50</td><td>0.51</td><td>0.55</td><td>0.52</td></tr><tr><td>Supervised by Preference (SP)</td><td>0.64</td><td>0.74</td><td>0.68</td><td>0.58</td><td>0.68</td><td>0.66</td><td>0.66</td></tr></table>
+
+Table 4: Experimental results of two training methods on the Token-level Benchmark  
+
+<table><tr><td>Source</td><td colspan="10">在橘红色背景墙映衬下格外鲜亮的广告牌上，它们各自标出了不菲的价码。</td></tr><tr><td>Reference</td><td colspan="10">On the bright advertising board against a tangerine colored wall, they listed their respective exorbitant price tags.</td></tr><tr><td>Translation A</td><td colspan="10">Against an orange background, the bright billboards listed their respective, hefty prices.</td></tr><tr><td>Translation B</td><td colspan="10">Against an orange background, the dark billboards listed their respective, hefty prices.</td></tr><tr><td>Translation C</td><td colspan="10">Against an orange background, the bright billboards listed their respective, bargain prices.</td></tr><tr><td>Translation A Reward</td><td>‘Against’ -2.3</td><td>‘an’ -2.3</td><td>...</td><td>‘the’ 0.07</td><td>‘bright’ 1.28</td><td>‘bill’ -2.3</td><td>...</td><td>‘hefty’ 0.03</td><td>‘prices’ 0.09</td><td>Weighted Implicit Rewards(↑) -3.43</td></tr><tr><td>Translation B Reward</td><td>‘Against’ -2.3</td><td>‘an’ -2.3</td><td>...</td><td>‘the’ 0.07</td><td>‘dark’ -2.3</td><td>‘bill’ -2.3</td><td>...</td><td>‘hefty’ 0.09</td><td>‘prices’ 0.09</td><td>Weighted Implicit Rewards(↑) -4.13</td></tr><tr><td>Translation C Reward</td><td>‘Against’ -2.3</td><td>‘an’ -2.3</td><td>...</td><td>‘the’ 0.07</td><td>‘bright’ 1.28</td><td>‘bill’ -2.3</td><td>...</td><td>‘bargain’ -0.95</td><td>‘prices’ -2.3</td><td>Weighted Implicit Rewards(↑) -3.99</td></tr></table>
+
+Table 5: Case study illustrating token-level credit assignment by our Qwen-PRM.
+
+In contrast, the SP approach, using DPO/KTO objectives, excels by directly learning a representation reflecting the advantage of one token choice over another. This implicitly shapes a reward function where the per-step process reward,  $r_{\theta}^{t} = \beta \log \frac{\pi_{\theta}(y_t|y_{< t})}{\pi_{ref}(y_t|y_{< t})}$  (Eq. 9), quantifies the localized, step-wise advantage of selecting token  $y_{t}$ . This preference-based optimization strategy is well-supported by prior work (Yuan et al., 2024; Rafailov et al., 2024a; Wang et al., 2024a). Crucially, DPO's mechanism, focusing on the log-probability ratio between sequences (Eq. 4), is inherently more sensitive to relative differences than absolute value regression. This makes it adept at capturing preferences even when underlying  $V(a)$  scores from the SV context are close, explaining its superior performance in training PRMs for nuanced token-level discrimination in our MT setting.
+
+# 5.2 Per-token Credit Assignment
+
+Our PRM can identify token-level translation differences, with the per-step process reward defined as Eq. 9 quantifying the reward for generating token  $y_{t}$  at step  $t$ . These individual token rewards are then aggregated into a final weighted sequence score (Eq. 11), allowing for a comprehensive evaluation that originates from fine-grained assessments.
+
+Table 5 provides a case study illustrating these capabilities. For instance, in Translation B, the token "dark", which semantically contradicts the source "鲜亮" (bright), receives a significantly negative reward (-2.3) from our PRM. In Translation C, the incorrect token "bargain", used where "不菲的" (exorbitant/costly) is implied, is substantially penalized (-0.95). In contrast, contextually appropriate tokens in Translation A, such as
+
+"bright" (1.28) and "hefty" (0.03), secure relatively higher rewards. Furthermore, the final weighted sequence scores computed by our PRM demonstrate strong alignment with automatic metrics like COMETKiwi. Translation A, the highest quality hypothesis (COMETKiwi 0.80), also achieves the most favorable weighted PRM score (-3.43). This case study thus substantiates our PRM's effective token-level credit assignment and the consistency of its fine-grained assessments with established sequence-level quality metrics.
+
+# 5.3 Test-time Alignment
+
+Test-time alignment, also known as decoding-time alignment (Huang et al., 2024; Rashid et al., 2024), refers to the process of adjusting an LM's output during inference to better align with human preferences, without additional training or fine-tuning. Its application in MT remains underexplored.
+
+In the context of MT, given the prior context  $s_{<t}$  and timestamp  $t$ , we define the reward-guided scoring function for a candidate token  $a$  as:
+
+$$
+s (a, s _ {<   t}) = \operatorname {L M} (a \mid s _ {<   t}) + w \cdot P (r ([ s _ {<   t}, a ])) \tag {13}
+$$
+
+where  $\mathrm{LM}(a\mid s_{< t})$  represents the LM's predicted probability for token  $a$  given the preceding context  $s_{< t}$ .  $r([s_{< t},a])$  denotes the reward signal for token  $a$ , conditioned on the prior context  $s_{< t}$ . The softmax function is applied over the reward signal  $r([s_{< t},a])$ , computed over the top  $k$  candidate tokens (with  $k$  being a window size), normalizing the reward value, which we label as  $P(r([s_{< t},a])$ . The scaling factor  $w$  adjusts the relative weight of the reward signal, allowing it to contribute effectively without overpowering the LM's probability. Compared to standard decoding strategies, this ap
+
+proach offers a more refined scoring function, as it encourages the generated text to: 1) Maintain semantic coherence and relevance with the prior context, and 2) Align more closely with reward-based criteria and human preferences. Test-time alignment also substantially reduces the need for the extensive resources typically required for LM alignment training.
+
+![](images/7a779f6c31be042bb9d7c6d5e4f6e28a8c105bc4593f93b8ad283816c4a138c4.jpg)
+
+![](images/3f382772d10e41bde93c7e0659e46ccfd6fde52d9f90cac4d03e55fbcc83da33.jpg)  
+Figure 3: Results of test-time alignment across WMT 23 ZH-EN and EN-RU. MT-PRMs with less parameters can assist in aligning Qwen-2.5-14B-Instruct.
+
+We use Qwen2.5-14B-Instruct<sup>3</sup> for generating tokens and leverage MT-PRM-LLaMA-3.2-3B and MT-PRM-Qwen-2.5-3B as the models for providing token-level rewards. We randomly sample 500 cases from the WMT 2023 testset. As shown in Figure 3, the reward-guided decoding methods outperform the standard greedy decoding in both EN-RU and ZH-EN translation tasks, evaluated by the COMET (Rei et al., 2020), COMETKiwi (Rei et al., 2022), and XCOMET-XL (Guerreiro et al., 2024) metrics. For instance, using the XCOMET-XL metric, LLaMA PRM and Qwen PRM outperform the standard greedy decoding by  $17.5\%$  and  $17.9\%$  in the EN-RU task respectively. Additionally, Qwen PRM slightly outperforms LLaMA PRM in both translation tasks and across all metrics, which aligns with the results in Table 1, where Qwen PRM achieves better token-level reward performance. These findings highlight the effectiveness of reward-guided decoding strategies in improving MT outcomes.
+
+A crucial aspect of reward-guided alignment is the trade-off with output diversity. Our framework explicitly manages this balance via the weighting parameter  $w$  in the reward-guided scoring function (Eq. 13). The parameter  $w$  controls the influence of the PRM: a higher  $w$  prioritizes the learned preferences for stronger alignment, while a lower  $w$  preserves more of the base model's original output distribution, thereby maintaining diversity.
+
+<table><tr><td>Model</td><td>Task</td><td>w=0</td><td>w=0.3</td><td>w=0.5</td><td>w=0.7</td></tr><tr><td rowspan="2">MT-PRM-Llama Guided</td><td>ZH-EN</td><td>0.7167</td><td>0.8759</td><td>0.8420</td><td>0.8677</td></tr><tr><td>EN-RU</td><td>0.5984</td><td>0.7459</td><td>0.7619</td><td>0.7143</td></tr><tr><td rowspan="2">MT-PRM-Qwen Guided</td><td>ZH-EN</td><td>0.7167</td><td>0.8723</td><td>0.8448</td><td>0.8671</td></tr><tr><td>EN-RU</td><td>0.5984</td><td>0.7444</td><td>0.7569</td><td>0.6721</td></tr></table>
+
+The quantitative results in Table 6 show that a range of  $w$  values provides a significant boost over the baseline ( $w = 0$ ), allowing users to choose a balance that suits their needs. For example, with the Llama-PRM on EN-RU,  $w = 0.5$  yields the best XCOMET score, while  $w = 0.3$  is optimal for the MT-PRM-Qwen. Furthermore, the qualitative case study in Table 7 illustrates that even under PRM guidance, different models and weights produce varied, high-quality outputs, demonstrating that diversity can be maintained.
+
+Table 6: XCOMET scores showing the impact of the weighting parameter  $w$  on test-time alignment. The baseline is Qwen2.5-14B-Instruct ( $w = 0$ ). Optimal values vary, allowing for tunable performance.  
+
+<table><tr><td>System</td><td>Translation</td></tr><tr><td>Source (Refer-ence)</td><td>放入充电盒中，耳机自动关机。(Put it in the charging case, the head-phones will turn off automatically.)</td></tr><tr><td>Qwen2.5-14B-Instruct (w=0)</td><td>Put them in the charging case, and the headphones will shut off automat-ically.</td></tr><tr><td>MT-PRM-Llama (w=0.3)</td><td>Put them in the charging case, and the headphones will shut off automat-ically.</td></tr><tr><td>MT-PRM-Qwen (w=0.5)</td><td>When placed in the charging case, the headphones shut off automatically.</td></tr><tr><td>MT-PRM-Llama (w=0.7)</td><td>Place the earbuds in the charging case, and they will shut down automatically.</td></tr></table>
+
+Table 7: Qualitative case study illustrating output diversity under PRM guidance.
+
+# 6 Related Work
+
+Token-Level Feedback Mechanisms. Fine-grained feedback has been recognized for its ability to help models capture potential errors more
+
+precisely (Lightman et al., 2024). In the context of mathematical reasoning, process supervision using Monte Carlo methods has shown significant promise (Wang et al., 2024b; Qi et al., 2024; Guan et al., 2025). Furthermore, developments in general-domain have demonstrated that DPO can implicitly learn token-level rewards through policy optimization, a process referred to implicit reward learning (Rafailov et al., 2024a; Wang et al., 2024a; Yuan et al., 2024). Despite these advancements, these approaches have yet to be tested in the context of MT. The translation community has long acknowledged the value of granular feedback, with early attempts relying on binary error markings from human annotations (Kreutzer et al., 2020), reference-based heuristics (Petrushkov et al., 2018), or LLM (Feng et al., 2024b).
+
+Alignment Paradigms in Machine Translation. Alignment techniques in neural machine translation have evolved from Minimum Risk Training (Shen et al., 2015) to more sophisticated reinforcement learning approaches (Dang et al., 2024). While PPO-based RLHF has achieved success in general-domain alignment, its application to MT presents unique challenges, particularly due to the need for fine-grained quality signals rather than the bandit reward. Recent works like He et al. (2024) and Xu et al. (2024) have investigated the use of automatic metrics to select better translations or construct preference pairs to improve the LLM, while Zhao et al. (2024) explored scaling test-time compute to further enhance translation performance. Recently, Ramos et al. (2024) pioneered the use of xCOMET as a dynamic reward signal during RL training. However, these methods remain limited to sequence-level guidance or binary approximations of the reward process, failing to provide the fine-grained token-level feedback required for more accurate translation alignment.
+
+# 7 Conclusion
+
+In this work, we propose MT-RewardTree, a comprehensive framework for constructing, evaluating, and deploying process reward models in MT. Our framework leverages an automatic token-level preference pair generation approach inspired by approximate Monte Carlo Tree Search, effectively addressing the challenge of large-scale fine-grained supervision annotation. Extensive experiments on both sequence-level and token-level benchmarks demonstrate that our MT-PRM achieves advanced
+
+performance in reward modeling in MT, surpassing traditional sequence-level preference pairs. Our exploration of token credit assignment and test-time alignment provide valuable insights for the application of reward models in MT.
+
+# Limitations
+
+Although we have developed the first comprehensive framework for process reward models in the field of machine translation, several important challenges remain to be addressed. Our work primarily focuses on synthesizing token-level data to leverage its fine-grained benefits. However, methods like Token-level DPO, RTO which optimize training algorithms, also show promise in further improving PRM performance. Additionally, our current framework includes only a limited set of high-resource languages, and expanding to multilingual settings, especially for low-resource languages, is a crucial direction for future work. While we have demonstrated the potential applications of reward models in test-time alignment and hypothesis ensembling, their integration into reinforcement learning training remains an important area for exploration.
+
+# Acknowledgments
+
+This work is supported by the National Key R&D Program of China (Grant No. 2024YFC3308304), the "Pioneer" and "Leading Goose" R&D Program of Zhejiang (Grant no. 2025C01128), the National Natural Science Foundation of China (Grant No. 62476241), the Natural Science Foundation of Zhejiang Province, China (Grant No. LZ23F020008), the Zhejiang Zhejiang Key Laboratory of Medical Imaging Artificial Intelligence.
+
+# References
+
+Sweta Agrawal, José G. C. De Souza, Ricardo Rei, Antonio Farinhas, Gonçalo Faria, Patrick Fernandes, Nuno M Guerreiro, and Andre Martins. 2024. Modeling user preferences with automatic metrics: Creating a high-quality preference dataset for machine translation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14503-14519, Miami, Florida, USA. Association for Computational Linguistics.
+
+Ralph Allan Bradley and Milton E. Terry. 1952. Rank analysis of incomplete block designs. Biometrika.
+
+Meng Cao, Lei Shu, Lei Yu, Yun Zhu, Nevan Wichers, Yinxiao Liu, and Lei Meng. 2024. Enhancing
+
+reinforcement learning with dense rewards from language model critic. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 9119-9138, Miami, Florida, USA. Association for Computational Linguistics.  
+Alex James Chan, Hao Sun, Samuel Holt, and Mihaela van der Schaar. 2024. Dense reward for free in reinforcement learning from human feedback. In *Forty-first International Conference on Machine Learning*.  
+Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2024. Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070.  
+Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.  
+John Dang, Arash Ahmadian, Kelly Marchisio, Julia Kreutzer, Ahmet Üstün, and Sara Hooker. 2024. RLHF can speak many languages: Unlocking multilingual preference optimization for LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 13134-13156, Miami, Florida, USA. Association for Computational Linguistics.  
+DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghai Lu, Shangyan Zhou, Shanhuang Chen Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu Shengfeng Ye Tao Yun Tian Pei Tianyu Sun T. Wang, Wangding Zeng Wanjia Zhao Wen Liu Wenfeng Liang Wenjun Gao Wenqin Yu Wentao Zhang W. L. Xiao Wei An Xiaodong Liu Xiaohan Wang Xiaokang Chen Xiaotao Nie Xin Cheng Xin Liu Xin Xie Xingchao Liu Xinyu Yang Xinyuan Li Xuecheng Su Xuheng Lin X. Q. Li Xiangyue Jin
+
+Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.  
+Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306.  
+Zhaopeng Feng, Ruizhe Chen, Yan Zhang, Zijie Meng, and Zuozhu Liu. 2024a. Ladder: A model-agnostic framework boosting LLM-based machine translation to the next level. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15377-15393, Miami, Florida, USA. Association for Computational Linguistics.  
+Zhaopeng Feng, Yan Zhang, Hao Li, Wenqiang Liu, Jun Lang, Yang Feng, Jian Wu, and Zuozhu Liu. 2024b. Improving llm-based machine translation with systematic self-correction. arXiv preprint arXiv:2402.16379.  
+Markus Freitag, George Foster, David Grangier, Viresh Ratnakar, Qijun Tan, and Wolfgang Macherey. 2021. Experts, errors, and context: A large-scale study of human evaluation for machine translation. Transactions of the Association for Computational Linguistics, 9:1460-1474.  
+Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. rstar-math: Small llms can master math reasoning with self-evolved deep thinking. arXiv preprint arXiv:2501.04519.  
+Nuno M Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and Andre FT Martins. 2024. xcomet: Transparent machine translation evaluation through fine-grained error detection. Transactions of the Association for Computational Linguistics, 12:979-995.  
+Zhiwei He, Xing Wang, Wenxiang Jiao, Zhuosheng Zhang, Rui Wang, Shuming Shi, and Zhaopeng Tu. 2024. Improving machine translation with human feedback: An exploration of quality estimation as a reward model. arXiv preprint arXiv:2401.12873.
+
+James Y Huang, Sailik Sengupta, Daniele Bonadiman, Yi-an Lai, Arshit Gupta, Nikolaos Pappas, Saab Mansour, Katrin Kirchhoff, and Dan Roth. 2024. Deal: Decoding-time alignment for large language models. arXiv preprint arXiv:2402.06147.  
+Levente Kocsis and Csaba Szepesvári. 2006. Bandit based monte-carlo planning. In Machine Learning: ECML 2006, pages 282-293, Berlin, Heidelberg. Springer Berlin Heidelberg.  
+Julia Kreutzer, Nathaniel Berger, and Stefan Riezler. 2020. Correct me if you can: Learning from error corrections and markings. arXiv preprint arXiv:2004.11222.  
+Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. 2024. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787.  
+Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let's verify step by step. In *The Twelfth International Conference on Learning Representations*.  
+Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. 2024. Rm-bench: Benchmarking reward models of language models with subtlety and style. arXiv preprint arXiv:2410.16184.  
+Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, et al. 2024. Improve mathematical reasoning in language models by automated process supervision. arXiv preprint arXiv:2406.06592.  
+Ibraheem Muhammad Moosa, Rui Zhang, and Wenpeng Yin. 2024. Mt-ranker: Reference-free machine translation evaluation by inter-system ranking. arXiv preprint arXiv:2401.17099.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744.  
+Pavel Petrushkov, Shahram Khadivi, and Evgeny Matusov. 2018. Learning from chunk-based feedback in neural machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 326-331, Melbourne, Australia. Association for Computational Linguistics.  
+Zhenting Qi, Mingyuan Ma, Jiahang Xu, Li Lyna Zhang, Fan Yang, and Mao Yang. 2024. Mutual reasoning makes smaller llms stronger problem-solvers. arXiv preprint arXiv:2408.06195.
+
+Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. 2024a. From $r$ to $q^*$: Your language model is secretly a q-function. In First Conference on Language Modeling.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024b. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36.  
+Miguel Moura Ramos, Tomás Almeida, Daniel Vareta, Filipe Azevedo, Sweta Agrawal, Patrick Fernandes, and André FT Martins. 2024. Fine-grained reward optimization for machine translation using error severity mappings. arXiv preprint arXiv:2411.05986.  
+Ahmad Rashid, Ruotian Wu, Julia Grosse, Agustinus Kristiadi, and Pascal Poupart. 2024. A critical look at tokenwise reward-guided text generation. arXiv preprint arXiv:2406.07780.  
+Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. Comet: A neural framework for mt evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702.  
+Ricardo Rei, Marcos Treviso, Nuno M Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José GC De Souza, Taisiya Glushkova, Duarte Alves, Luísca Coheur, et al. 2022. Cometkiwi: Ist-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 634-645.  
+Christopher D. Rosin. 2011. Multi-armed bandits with episode context. Annals of Mathematics and Artificial Intelligence, 61(3):203-230.  
+John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.  
+Shiqi Shen, Yong Cheng, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. 2015. Minimum risk training for neural machine translation. arXiv preprint arXiv:1512.02433.  
+David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, L. Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Vedavyas Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy P. Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. 2016. Mastering the game of go with deep neural networks and tree search. Nature, 529:484-489.  
+David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy Lillicrap, Fan Hui,
+
+Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. 2017. Mastering the game of go without human knowledge. Nature, 550(7676):354-359.  
+Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling ltm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314.  
+Kimi Team. 2025. Kimi k1.5: Scaling reinforcement learning with llms.  
+Huaijie Wang, Shibo Hao, Hanze Dong, Shenao Zhang, Yilin Bao, Ziran Yang, and Yi Wu. 2024a. Offline reinforcement learning for llm multi-step reasoning. arXiv preprint arXiv:2412.16145.  
+Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. 2024b. Math-shepherd: Verify and reinforce LLMs step-by-step without human annotations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9426-9439, Bangkok, Thailand. Association for Computational Linguistics.  
+Violet Xiang, Charlie Snell, Kanishk Gandhi, Alon Albalak, Anikait Singh, Chase Blagden, Duy Phung, Rafael Rafailov, Nathan Lile, Dakota Mahan, et al. 2025. Towards system 2 reasoning in llms: Learning how to think with meta chain-of-though. arXiv preprint arXiv:2501.04682.  
+Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024. Contrastive preference optimization: Pushing the boundaries of LLM performance in machine translation. In *Forty-first International Conference on Machine Learning*.  
+Lifan Yuan, Wendi Li, Huayu Chen, Ganqu Cui, Ning Ding, Kaiyan Zhang, Bowen Zhou, Zhiyuan Liu, and Hao Peng. 2024. Free process rewards without process labels. arXiv preprint arXiv:2412.01981.  
+Zhiyuan Zeng, Qinyuan Cheng, Zhangyue Yin, Bo Wang, Shimin Li, Yunhua Zhou, Qipeng Guo, Xuanjing Huang, and Xipeng Qiu. 2024. Scaling of search and learning: A roadmap to reproduce o1 from reinforcement learning perspective. arXiv preprint arXiv:2412.14135.  
+Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2025. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301.  
+Yu Zhao, Huifeng Yin, Bo Zeng, Hao Wang, Tianqi Shi, Chenyang Lyu, Longyue Wang, Weihua Luo, and Kaifu Zhang. 2024. Marco-o1: Towards open reasoning models for open-ended solutions. Preprint, arXiv:2411.14405.
+
+Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2024. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559.  
+Han Zhong, Guhao Feng, Wei Xiong, Xinle Cheng, Li Zhao, Di He, Jiang Bian, and Liwei Wang. 2024. Dpo meets ppo: Reinforced token optimization for rlhf. arXiv preprint arXiv:2404.18922.

EMNLP/2025/MT-RewardTree_ A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96609c7b69e5e0d114a3c1a95e740c779f59d63d3ce22b7c1847af21c98c85c6
+size 515148

EMNLP/2025/MT-RewardTree_ A Comprehensive Framework for Advancing LLM-Based Machine Translation via Reward Modeling/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57225a1890153444e8373f598030aad27e710d1102b02f6175f89e51536947a0
+size 426344

EMNLP/2025/MTabVQA_ Evaluating Multi-Tabular Reasoning of Language Models in Visual Space/dc67d43a-5509-4812-9e04-eb08fe821868_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ce7db6bf940fd7ace4e22837bcffdd202323a9910b7f9ef4bb69b5543a56727
+size 148897

EMNLP/2025/MTabVQA_ Evaluating Multi-Tabular Reasoning of Language Models in Visual Space/dc67d43a-5509-4812-9e04-eb08fe821868_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae8ca42752622ddad732af3a1916ed0f710f039f199359ee499022bf789b0bcf
+size 186303

EMNLP/2025/MTabVQA_ Evaluating Multi-Tabular Reasoning of Language Models in Visual Space/dc67d43a-5509-4812-9e04-eb08fe821868_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04d6324fe1e0233dc3512e4422e3de74977bcd067172caccb1628b37c7f3a55b
+size 3630061

EMNLP/2025/MTabVQA_ Evaluating Multi-Tabular Reasoning of Language Models in Visual Space/full.md

@@ -0,0 +1,777 @@
+# MTabVQA: Evaluating Multi-Tabular Reasoning of Language Models in Visual Space
+
+Anshul Singh<sup>1</sup>, Chris Biemann<sup>2</sup>, Jan Strich<sup>2</sup>
+
+$^{1}$ Department of Information Technology, Panjab University, India  $^{2}$ Language Technology Group, Universität Hamburg, Germany
+
+Correspondence: {firstname_lastname} @ uni-hamburg.de
+
+# Abstract
+
+Vision-Language Models (VLMs) have demonstrated remarkable capabilities in interpreting visual layouts and text. However, a significant challenge remains in their ability to interpret and reason robustly over multi-tabular data presented as images, a common occurrence in real-world scenarios such as web pages and digital documents. Existing benchmarks typically address single tables or non-visual data (text/structured). This leaves a critical gap: they do not assess the ability to parse diverse table images, correlate information across them, and perform multi-hop reasoning on the combined visual data. We introduce MTabVQA, a novel benchmark specifically designed for multi-tabular visual question answering to bridge this gap. MTabVQA comprises 3,745 complex question-answer pairs that necessitate multi-hop reasoning across several visually rendered table images. We provide extensive benchmark results for state-of-the-art VLMs on MTabVQA, revealing significant performance limitations. We further investigate posttraining techniques to enhance the multitabular reasoning abilities of vision-language models and release MTabVQA-Instruct, a large-scale instruction-tuning dataset. Our experiments show that fine-tuning VLMs with MTabVQA-Instruct substantially improves their performance on visual multi-tabular reasoning. Code and dataset are available online<sup>1</sup>.
+
+# 1 Introduction
+
+In recent years, vision language models (VLMs) and multimodal systems have demonstrated remarkable capabilities in interpreting complex visual layouts and text (Luo et al., 2024), enabling tasks ranging from document understanding (Zhang et al., 2025), visual information extraction (Cao et al., 2023), and structured data QA (Antol et al., 2015) to interactive processes like autonomous web navigation (He et al., 2024).
+
+![](images/5bf086400d9f7ab1387dba252fac26790ce611660522be411694419d7312b16b.jpg)  
+Figure 1: MTabVQA Benchmark, illustrative example showing three tables (Customers, Reviews, Services), a question requiring multi-table reasoning, the reasoning steps involved, and the final answer derived by a vision-language model.
+
+Yet, as these models become increasingly capable of acting as visual agents that can browse screen data and execute complex instructions, a new challenge has emerged: the interpretation and reasoning over multi-tabular data presented as images (Deng et al., 2024; Zheng et al., 2024). This challenge is particularly relevant in real-world scenarios, where tables often appear as images on web pages or digital documents. For instance, with approximately  $24\%$  of PDF documents being image-only formats (Johnson, 2018), their tabular data is inaccessible to traditional text-based models. Extracting actionable insights from such sources may require a model to reference multiple table images simultaneously.
+
+Traditional benchmarks (Yu et al., 2018; Chen et al., 2020; Zhong et al., 2017) in table understanding and question answering have primarily focused on single-table scenarios, often relying on textual or HTML representations (Zhu et al., 2021; Sui et al., 2024). However, such benchmarks are
+
+unable to evaluate model performance on visually complex, multi-tabular data, which requires interpreting layout and structure beyond simple text or HTML. In many practical applications, such as financial analysis, e-commerce, and scientific research (Lautert et al., 2013), key information is distributed across several tables, each with distinct layouts and visual structures. Current benchmarks (Wu et al., 2025b; Pal et al., 2023; Wu et al., 2025a; Li et al., 2025b), rooted in single-table, non-visual formats (like text/HTML or relational databases), fail to assess critical capabilities: (1) understanding diverse visual table layouts presented as images, (2) parsing and correlating information across multiple, physically separate tables, and (3) executing multi-hop reasoning grounded in visual data.
+
+To bridge this gap, we propose Multi-Tabular Visual Question Answering (MTabVQA), a novel benchmark specifically designed for assessing the visual reasoning capabilities of models on multi-tabular data represented as images. Distinct from prior benchmarks that primarily focus on single tables (Pasupat and Liang, 2015; Zhong et al., 2017; Zheng et al., 2024) or utilize non-visual (textual, structured) formats for multi-table reasoning (Wu et al., 2025a; Yu et al., 2018; Li et al., 2023a), MTabVQA uniquely evaluates the integration of information across multiple tables. Our benchmark, comprising 3,745 question-answer pairs, challenges models with complex queries across 14 distinct reasoning categories. These pairs are designed to necessitate multi-hop reasoning (e.g., involving aggregation, comparison, or fact-checking) by integrating information from two to five table images. MTabVQA enables a targeted evaluation of how well current models handle the process of extracting information from multiple table images and performing the multi-hop reasoning necessary to synthesize answers. Our main contributions are:
+
+- We introduce MTabVQA, a novel benchmark designed to evaluate multi-hop reasoning over multiple tables presented as images, addressing a gap in existing table QA benchmarks.  
+- We release MTabVQA-Instruct, an instruction-focused training dataset for multi-table QA.  
+- We trained TableVision, a VLM fine-tuned on MTabVQA-Instruct, which shows significant improvements on visual multi-tabular reasoning.
+
+- We provide extensive benchmark results for SOTA open-source, proprietary VLMs and TableVision on MTabVQA, revealing significant challenges posed by this task.
+
+# 2 Related Work
+
+Research in table understanding (Wu et al., 2025b) and multimodal reasoning (Zheng et al., 2024) has made significant advancements. Initial efforts often centered on converting tables into text-based representations, such as Markdown or HTML (Li et al., 2024; Zhang et al., 2024a), enabling text-only language models to process them. While effective in controlled environments, this approach encounters limitations in real-world settings where tables frequently appear only as images within documents or web interfaces. Processing visually rendered tables through multi-stage text-conversion pipelines (Nassar et al., 2022) presents inherent limitations.
+
+The pipelines are often complex and susceptible to OCR errors; they also tend to discard essential visual layout cues (e.g., merged cells, alignment), and risk compounding inaccuracies across stages. This highlights a critical need for models capable of interpreting and reasoning over tables directly from pixel data. Moreover, most systems rely on OCR combined with LLMs, which makes them more prone to errors than developing a single unified model. Our work focuses entirely on the challenge of extracting information directly from visual table data and drawing conclusions from it, addressing the complexities inherent in image-based table structures.
+
+# 2.1 Table Understanding and Extraction
+
+Effective reasoning over visual tables fundamentally relies on accurate underlying table understanding, including tasks like detection, segmentation, and structure interpretation (Bonfitto et al., 2021). Such foundational challenges were often addressed by specialized methods leveraging object detection and OCR, exemplified by systems like TableFormer (Nassar et al., 2022), which improved the extraction of cell structures from images. Despite the advances, such methods frequently encounter difficulties with complex visual layouts and the semantic alignment crucial for interpreting elements like multirow headers or merged cells.
+
+Although recent large-scale datasets like MMTab (Zheng et al., 2024) have significantly advanced benchmarking for table extraction and
+
+<table><tr><td>Benchmark</td><td>Question Format</td><td># Tables/Databases</td><td># QA Pairs</td><td>Task</td><td>Modality</td></tr><tr><td>WTQ (Pasupat and Liang, 2015)</td><td>NL Questions</td><td>2,108</td><td>22,033</td><td>Single-table QA</td><td>Text</td></tr><tr><td>SQA (Iyyer et al., 2017)</td><td>NL Questions</td><td>N/A</td><td>17,553</td><td>Single-Table QA</td><td>Text</td></tr><tr><td>WikiSQL (Zhong et al., 2017)</td><td>SQL Query</td><td>24,241</td><td>80k</td><td>Single-table QA</td><td>Text</td></tr><tr><td>Spider (Yu et al., 2018)</td><td>NL Questions &amp; SQL Query</td><td>200</td><td>10,181</td><td>Text-to-SQL</td><td>Text</td></tr><tr><td>HybridQA (Chen et al., 2020)</td><td>NL Questions</td><td>13,000</td><td>70k</td><td>Table-text QA</td><td>Text</td></tr><tr><td>FeTaQA (Nan et al., 2022)</td><td>NL Questions</td><td>10,330</td><td>10k</td><td>Single tables</td><td>Text</td></tr><tr><td>BIRD (Li et al., 2023a)</td><td>NL Questions &amp; SQL Query</td><td>95</td><td>12,751</td><td>Text-to-SQL</td><td>Text</td></tr><tr><td>TableBench (Wu et al., 2025b)</td><td>NL Questions</td><td>3,681</td><td>886</td><td>Single Table</td><td>Text</td></tr><tr><td>SPINACH (Liu et al., 2024)</td><td>NL Questions &amp; SQL Query</td><td>N/A</td><td>320</td><td>Text-to-SQL</td><td>Text</td></tr><tr><td>MMQA (Wu et al., 2025a)</td><td>NL Questions &amp; SQL Query</td><td>3,312</td><td>3,312</td><td>Text-to-SQL, Multi-table QA</td><td>Text</td></tr><tr><td>MMTab (Zheng et al., 2024)</td><td>NL Questions</td><td>23K</td><td>49K</td><td>Single-Table QA</td><td>Images</td></tr><tr><td>MTabVQA (ours)</td><td>NL Questions</td><td>8499</td><td>3,745</td><td>Multi-Table QA</td><td>Images</td></tr></table>
+
+Table 1: Differences between our MTabVQA and previous table QA benchmarks. We here abbreviate NL = Natural Language and SQL = Structured Query Language.
+
+understanding from table images, they primarily focus on single-table scenarios. The challenge of integrating information and reasoning across multiple visually presented tables, which MTabVQA addresses, remains less explored.
+
+# 2.2 Multimodal Question Answering
+
+Early benchmarks in table QA, such as WikiTable-Questions (Pasupat and Liang, 2015) and WikiSQL (Zhong et al., 2017) established the task but focused on single-table scenarios with text-based representations. More recent work like MMQA (Wu et al., 2025a) extends to multi-table and multi-hop reasoning but still relies on text, not raw images.
+
+In parallel, multimodal QA has made significant progress with general-purpose models like LLaVA (Li et al., 2025a), BLIP-2 (Li et al., 2023b), and GPT-4.1 (OpenAI et al., 2024) demonstrating strong capabilities on image-based tasks. While current models excel in general visual understanding, their capacity for reasoning across multiple tables presented as images remains largely unexplored by existing benchmarks.
+
+# 2.3 Multi-Tabular Reasoning
+
+Reasoning across multiple tables demands correlating information from potentially disparate structures via multi-hop operations, a known challenge for current models (Pal et al., 2023). While prior work explored multi-table QA (Pal et al., 2023), summarization (Zhang et al., 2024b), and text-to-SQL (Wu et al., 2025a), these efforts predominantly relied on textual or structured data representations. They often bypassed the complexities of interpreting combined visual table layouts, a critical requirement for agents interacting with screen data.
+
+MTabVQA directly addresses this research gap by focusing on multi-tabular visual reasoning. As in Table 1, prominent prior benchmarks like WTQ (Pasupat and Liang, 2015), WikiSQL (Zhong et al., 2017), and even multi-table focused ones such as Spider (Yu et al., 2018) and MMQA (Wu et al., 2025a), primarily operate on textual or structured representations of tables. While MMTab (Zheng et al., 2024) introduced image-based tables, its focus remained on single-table scenarios.
+
+In contrast, MTabVQA specifically requires models to answer complex, multi-hop questions by integrating information presented across multiple table images. This necessitates visual parsing of diverse table layouts from images, a capability not comprehensively evaluated by existing benchmarks that are either non-visual or single-table centric. Thus, MTabVQA's unique combination of multi-table reasoning and image-based input directly targets this underexplored area.
+
+# 3 MTabVQA Dataset
+
+We introduce Multi-Tabular Visual Question Answering (MTabVQA), a new dataset designed to evaluate and improve multi-hop reasoning over visually rendered tables. This dataset comprises two distinct and complementary components: MTabVQA, a benchmark with 3,745 QA pairs for evaluating model performance, and MTabVQA-Instruct, a large-scale instruction-tuning dataset with 15,853 examples. The MTabVQA benchmark is further divided into four sub-datasets based on the primary source of the underlying table data. The detailed composition of both the sub-datasets and their sources is shown in Table 2.
+
+<table><tr><td>Dataset Split</td><td>Source</td><td>Sub-dataset</td><td>#QA Pairs</td><td>#Tables</td><td>Proportion (%)</td></tr><tr><td rowspan="5">MTabVQA</td><td>QFMTS (Zhang et al., 2024b)</td><td>MTabVQA-Query</td><td>2456</td><td>5541</td><td>65.7%</td></tr><tr><td>Spider (Yu et al., 2018)</td><td>MTabVQA-Spider</td><td>1048</td><td>2363</td><td>27.9%</td></tr><tr><td>Atis (Dahl et al., 1994)</td><td>MTabVQA-Atis</td><td>112</td><td>429</td><td>3.0%</td></tr><tr><td>MiMoTable (Li et al., 2025b)</td><td>MTabVQA-Mimo</td><td>129</td><td>166</td><td>3.4%</td></tr><tr><td>Total Eval Set</td><td></td><td>3745</td><td>8499</td><td>100.0%</td></tr><tr><td rowspan="6">MTabVQA-Instruct</td><td>MultiTabQA (Pal et al., 2023)</td><td>-</td><td>10,990</td><td>21,976</td><td>69.3%</td></tr><tr><td>Spider (Yu et al., 2018)</td><td>-</td><td>2395</td><td>5845</td><td>15.2%</td></tr><tr><td>BIRD (Li et al., 2023a)</td><td>-</td><td>1572</td><td>3144</td><td>9.9%</td></tr><tr><td>Atis (Dahl et al., 1994)</td><td>-</td><td>384</td><td>1780</td><td>2.4%</td></tr><tr><td>MiMoTable (Li et al., 2025b)</td><td>-</td><td>512</td><td>719</td><td>3.2%</td></tr><tr><td>Full Instruct Set</td><td></td><td>15,853</td><td>33,464</td><td>100.0%</td></tr></table>
+
+Table 2: Detailed composition of the MTabVQA and MTabVQA-Instruct datasets. The table shows the original data sources and provides statistics for each sub-dataset, including the number of QA pairs and unique tables.
+
+To ensure a strict and fair separation between training and evaluation, these components are constructed from entirely disjoint data sources. The MTabVQA benchmark is built primarily from the development and test splits of its source datasets, while MTabVQA-Instruct is sourced exclusively from training splits. This prevents data leakage and guarantees that models are evaluated on unseen data structures and instances. The remainder of this section details the multi-stage pipeline (illustrated in Figure 2) used to construct both datasets, which includes data sourcing, relational sampling, image rendering, QA pair generation, and rigorous verification.
+
+# 3.1 Tabular Data Collection
+
+MTabVQA utilizes tabular data from BIRD (Li et al., 2023a), Spider (Yu et al., 2018), MiMoTable (Li et al., 2025b), QFMTS (Zhang et al., 2024b), and ATIS (Dahl et al., 1994). We prioritized text-to-SQL datasets as their associated complex SQL queries often involve multi-table joins, naturally lending themselves to multi-table reasoning tasks.
+
+To ensure our benchmark targets multi-table reasoning, we first identified relevant database subsets (Figure 2, Step 1). We parsed SQL queries from the source datasets, specifically selecting those requiring multi-table join operations. This analysis confirmed rich inter-table dependencies suitable for our task. Based on this query analysis, we extracted data instances for the MTabVQA split: 1,048 multi-join queries from Spider (Yu et al., 2018) forming MTabVQA-Spider, 2,578 multi-table instances from QFMTS (Zhang et al., 2024b), and 112 and
+
+129 multi-table pairs from ATIS (Dahl et al., 1994) and MiMoTable (Li et al., 2025b), respectively. The large and complex BIRD (Li et al., 2023a) dataset, over 7,200 join queries across 69 databases, was primarily used to generate MTabVQA-Instruct. This query-driven selection ensures that the underlying data inherently necessitates multi-table reasoning.
+
+# 3.2 Data Extraction and Preprocessing
+
+Following the identification of relevant database subsets (Section 3.1), we employed a pipeline to process the data. For each subset, the pipeline extracted the database schemata, including table definitions, column types, primary keys, and foreign key relationships defining inter-table links, and converted the relational data from its native storage (e.g., SQLite) into JSON format. As full tables can be too large for visualization and efficient processing, we adopted a controlled sampling strategy. Tables with more than  $N_{max} = 50$  rows were sampled, reducing size while maintaining a balance between visual clarity and representativeness across datasets.
+
+To preserve crucial relational information between multiple tables during sampling, we utilized a graph-based approach detailed in Algorithm 1 (Appendix A). This method ensures referential integrity by preferentially sampling rows linked across related tables via foreign keys, focusing on connections relevant to the multi-table queries identified earlier. The final output for each instance consists of the sampled table data and corresponding schemata, serialized into JSON.
+
+![](images/ba5297c2d50041b86016583b47a5c40b05cd3a4b60c26eb0cdff2f65a5d5d653.jpg)  
+Step 1: Table Extraction & Relational Sampling  
+Step 3: Synthetic QA Generation & Multi-Stage Verification  
+Step 2: SQL-based QA Pair Generation  
+Figure 2: MTabVQA Construction Framework Overview. (1) Data Sourcing & Sampling: Identify multi-table relational data via SQL joins, extract tables, apply relational sampling. (2) Visual QA Generation: Generate multi-hop QA pairs via SQL-to-question conversion or LLM-guided generation from sampled tables/taxonomy; render tables as images. (3) Verification & Finalization: Apply automated (LLM) and human verification for quality and multi-table necessity.
+
+# 3.3 Visual Table Rendering
+
+To ensure MTabVQA evaluates visual reasoning over image-based inputs, the sampled tabular data for each QA pair was rendered into images. This step forces models to interpret visual layouts over structured text. We utilized a rendering pipeline employing dataframe_image² (with selenium or matplotlib backends) and custom Pillow scripts. This process introduced significant visual diversity by systematically varying structural aspects (e.g., column/row dimensions, relative table positioning) and appearance features (e.g., color schemes, typography, grid styles) across 10 distinct, randomly applied styling themes. This approach simulates the varied appearances of tables in real-world documents and web pages. Further details on the specific themes are provided in Appendix D.
+
+# 3.4 Multi-Hop QA Pair Generation
+
+The QA pairs of the MTabVQA benchmark are designed for multi-hop reasoning across table images, generated via two strategies (Figure 2, Steps 2-3):
+
+1. SQL-to-Question (Step 2): We converted complex, multi-table SQL queries (from Section 3.1)
+
+into natural language questions. For each SQL query, we executed it on sampled table subsets  $(S_A, S_B)$  for a ground-truth answer. We used Gemini Flash 2.0 (Hassabis et al., 2024) to paraphrase the SQL (given schemes and instructions; Figure 2, bottom-left prompt) into a question, creating QA pairs grounded in verifiable SQL logic.
+
+2. Taxonomy-Guided Generation (Step 3): To diversify reasoning types, an LLM generated novel QA pairs from sampled table subsets and a predefined question taxonomy. This taxonomy, adapted from (Wu et al., 2025b) to cover common multi-table reasoning patterns (e.g., multi-hop fact-checking, aggregation), guided the LLM (with few-shot examples; Figure 2, upper-right prompt) to create questions requiring data from  $\geq 2$  tables, plus answers and reasoning steps in structured JSON. Figure 3 shows the distribution of the question categories, showing that most of the questions are fact-checking, analysis, aggregation, or ranking.
+
+# 3.5 Verification and Filtering
+
+To ensure QA quality and multi-table focus, our verification process (Figure 2, Step 3) was done by automated assessment from three LLM agents (Hass
+
+![](images/9f697625bb8b54f9ed9803832f3587e2e38826b36f84f774e33ce3ff13843097.jpg)  
+Figure 3: Distribution of verified question categories in the MTabVQA benchmark. "Other" includes categories like Anomaly Detection, Arithmetic Calculation, and Multi-hop Numerical Reasoning ( $N = 3,745$  QA pairs). See Appendix F for detailed descriptions and examples of the primary categories.
+
+abis et al., 2024), guided by a verification prompt (Appendix C). These agents evaluated question validity, multi-hop needs, answer accuracy, reasoning soundness, and multi-table necessity ( $\geq 2$  tables). LLM outputs (JSON with scores/flags) were aggregated by majority vote.
+
+Pairs meeting criteria (majority valid, confirmed multi-table use, average score  $\geq 7.0$ ) advanced to human verification using a Streamlit app (Appendix E) for final checks on correctness, especially for complex cases. Human Validation was conducted by one annotator. Only pairs passing both automated and human checks were integrated into MTabVQA. This LLM-assisted human oversight yielded a high-quality benchmark by filtering invalid tables or incorrect QA pairs.
+
+# 4 Experiments
+
+This section details the experiments conducted to evaluate VLM capabilities on visual multi-tabular reasoning using our MTabVQA benchmark. Our experiments encompass three key areas:
+
+1. Benchmarking Current VLMs: We first establish baseline performance by evaluating leading open-source and proprietary VLMs on the MTab-VQA split and compare it with our fine-tuned model (Section 4.1).  
+2. Evaluating Post-Training Strategies: We investigate and compare the effectiveness of three distinct post-training techniques to improve VLM performance: Chain-of-Thought (CoT) prompting,
+
+Group Relative Policy Optimization (GRPO), and Supervised Fine-Tuning (SFT) (Shao et al., 2024) (Section 4.2).
+
+3. Analyzing Cross-Dataset Generalization: We investigate how the composition (i.e., source and scale) of the instruction-tuning data affects model performance. To do this, we fine-tune models on distinct subsets of MTabVQA-Instruct and evaluate their ability to generalize across the full, unseen MTabVQA benchmark (Section 4.3).
+
+# 4.1 Benchmarking
+
+We conducted a comprehensive benchmarking study on MTabVQA to establish baselines for multi-table visual reasoning. We evaluated leading proprietary VLMs (GPT-4.1 (OpenAI et al., 2024), Gemini Flash 2.0 (Hassabis et al., 2024)) and prominent open-source alternatives (Qwen2.5 (Team et al., 2025b), Gemma-3 (Team et al., 2025a), LLaVA-One-Vision (Li et al., 2025a), InternVL3 (Zhu et al., 2025), Phi-3.5 (Abdin et al., 2024)), alongside our fine-tuned TableVision model. We assessed models in a zero-shot setting across all four MTabVQA sub-datasets (Spider, Query, ATIS, and MiMo), instructing them to generate structured JSON (Appendix H.1). Generation parameters were set to a temperature of 1.0 and top-P of 1.0.
+
+Evaluation Metrics. We primarily use EM for its strict correctness assessment, especially suitable for factual answers from tables. To capture semantic similarity and partial correctness, we also report F1 score, precision (P), and recall (R), providing a more nuanced view of answer quality.
+
+The results (Table 3) highlight that visual multibular reasoning is a challenging task for current VLMs. Open-source VLMs like LLaVA-OneVision  $(2.2\%$  EM,  $16.7\%$  F1 overall) and Phi-3.5-Vision struggled significantly in zero-shot, with Gemma-3 being the strongest open-source baseline  $(11.8\%$  EM,  $40.1\%$  F1 overall). Even proprietary models like GPT-4.1  $(37.0\%$  EM,  $61.7\%$  F1 overall) did not achieve perfect scores and showed performance dips on certain splits. For example, GPT-4.1's score on ATIS  $(6.3\%$  EM) is particularly revealing. It's very high recall  $(86.3\%)$  but low precision indicates that the model often identifies the correct information but includes it within verbose text, thus failing the strict EM criterion.
+
+TableVision. To demonstrate the value of targeted instruction fine-tuning, we introduce TableVision. We used Qwen2.5-VL-7B (Team et al.,
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">MTabVQA-Spider</td><td colspan="4">MTabVQA-Query</td><td colspan="4">MTabVQA-ATIS</td><td colspan="4">MTabVQA-MiMo</td><td colspan="2">Overall</td></tr><tr><td>EM</td><td>F1</td><td>P</td><td>R</td><td>EM</td><td>F1</td><td>P</td><td>R</td><td>EM</td><td>F1</td><td>P</td><td>R</td><td>EM</td><td>F1</td><td>P</td><td>R</td><td>EM</td><td>F1</td></tr><tr><td colspan="19">Open-Source VLMs (Zero-Shot)</td></tr><tr><td>LLaVA-OV-Qwen2-7B</td><td>2.2</td><td>20.0</td><td>19.5</td><td>29.3</td><td>2.3</td><td>15.7</td><td>15.9</td><td>23.6</td><td>0.0</td><td>9.2</td><td>5.9</td><td>33.8</td><td>0.7</td><td>5.5</td><td>4.3</td><td>19.1</td><td>2.1</td><td>18.4</td></tr><tr><td>Phi-3.5-Vision-Instruct</td><td>2.9</td><td>26.1</td><td>25.9</td><td>39.6</td><td>2.4</td><td>22.0</td><td>22.3</td><td>34.7</td><td>1.8</td><td>15.0</td><td>15.3</td><td>24.8</td><td>0.8</td><td>3.2</td><td>3.6</td><td>3.3</td><td>2.5</td><td>22.3</td></tr><tr><td>InternVL3-8B-Instruct</td><td>6.1</td><td>32.4</td><td>33.0</td><td>39.1</td><td>5.2</td><td>24.8</td><td>26.9</td><td>29.6</td><td>3.6</td><td>20.3</td><td>19.5</td><td>31.9</td><td>7.0</td><td>19.1</td><td>22.3</td><td>21.3</td><td>5.4</td><td>26.6</td></tr><tr><td>Qwen2.5-VL-7B</td><td>8.0</td><td>39.8</td><td>40.4</td><td>44.0</td><td>7.8</td><td>33.9</td><td>34.8</td><td>38.0</td><td>6.3</td><td>32.6</td><td>29.0</td><td>48.6</td><td>9.3</td><td>22.2</td><td>25.9</td><td>22.8</td><td>7.8</td><td>35.1</td></tr><tr><td>Gemma-3-12B-IT</td><td>15.6</td><td>48.0</td><td>48.2</td><td>53.4</td><td>10.3</td><td>38.1</td><td>39.4</td><td>42.6</td><td>11.6</td><td>35.1</td><td>34.2</td><td>40.8</td><td>9.3</td><td>18.6</td><td>22.0</td><td>18.8</td><td>11.8</td><td>40.1</td></tr><tr><td colspan="19">Proprietary VLMs (Zero-Shot)</td></tr><tr><td>Gemini-2.0-Flash</td><td>42.9</td><td>68.5</td><td>69.2</td><td>71.2</td><td>31.4</td><td>57.3</td><td>58.2</td><td>60.5</td><td>22.3</td><td>36.0</td><td>37.2</td><td>37.5</td><td>24.0</td><td>42.3</td><td>49.2</td><td>41.2</td><td>34.1</td><td>59.3</td></tr><tr><td>GPT-4.1</td><td>49.0</td><td>74.3</td><td>74.7</td><td>76.6</td><td>34.2</td><td>58.5</td><td>59.2</td><td>60.8</td><td>6.3</td><td>39.9</td><td>30.0</td><td>86.3</td><td>20.2</td><td>39.6</td><td>44.9</td><td>38.8</td><td>37.0</td><td>61.7</td></tr><tr><td colspan="19">Fine-tuned Model (Ours)</td></tr><tr><td>TableVision (Ours)</td><td>32.4</td><td>64.3</td><td>66.6</td><td>66.1</td><td>49.8</td><td>72.6</td><td>74.0</td><td>73.5</td><td>33.0</td><td>45.9</td><td>48.4</td><td>47.8</td><td>20.1</td><td>36.2</td><td>40.8</td><td>36.4</td><td>43.4</td><td>68.2</td></tr></table>
+
+Table 3: Performance Comparison of VLMs on MTabVQA Sub-datasets  $(\%)$ , and Overall EM/F1  $(\%)$ . Models categorized and sorted by overall F1 score within categories. Overall scores are weighted averages. Best overall and best open-source zero-shot overall scores are bolded. EM denotes Exact Match, P Precision, and R Recall.
+
+2025b) as the base model and fine-tuned it on our MTabVQA-Instruct dataset using a parameter-efficient training approach with Low-Rank Adaptation (LoRA) (Hu et al., 2022) at a rank of 128. As shown in Table 3, TableVision achieved the highest overall performance (43.4% EM, 68.2% F1), surpassing all other models, including GPT-4.1, on the MTabVQA-Query (49.8% EM, 72.6% F1) and MTabVQA-ATIS sub-datasets. This result shows that targeted fine-tuning can enable smaller open-source models to outperform larger proprietary systems on complex visual multi-tabular reasoning, underscoring the effectiveness of MTabVQA-Instruct.
+
+# 4.2 Post-training VLMs for Multi-Table Visual Reasoning
+
+To identify the most effective methods for enhancing VLM performance on visual multi-tabular reasoning, we conducted a controlled comparison of several post-training techniques. This section details the experimental setup and the corresponding results.
+
+Experimental Setup. Our investigation compares three distinct post-training strategies. For these intensive experiments, we selected the Qwen2.5-VL-3B model (Team et al., 2025b) as our base VLM, primarily due to its manageable size, which is crucial for the significant computational requirements of advanced methods like GRPO (Shao et al., 2024). The training was conducted on a specific small subset of our MTabVQA-Instruct dataset: the 2,395 QA pairs derived from the Spider data source as described in Table 2. This subset was chosen for two key reasons: its data quality and complex join operations provide a chal
+
+lenging and high-quality reasoning task, and its effectiveness for fine-tuning is demonstrated in our analysis (Section 4.3). To ensure a direct and fair comparison, all evaluations were performed on the corresponding MTabVQA-Spider sub-dataset from our MTabVQA benchmark.
+
+Results and Analysis. First, we established a baseline by evaluating the zero-shot performance of the 3B model. Consistent with observations for larger models (Section 4.1), the base 3B model exhibited poor initial performance on this complex multi-hop reasoning task, achieving an EM of  $2.8\%$  and an F1 score of  $22.9\%$  (Figure 4). We then evaluated the efficacy of using step-by-step reasoning through CoT prompting (see Appendix H.2). While this approach encouraged structured responses, it resulted in only marginal improvements, with EM increasing slightly to  $3.0\%$  and F1 to  $24.5\%$ .
+
+Next, recognizing the reasoning-intensive nature of multi-tabular VQA, we investigated GRPO (Shao et al., 2024), a reinforcement learning-based post-training technique (training details in Appendix G). As shown in Figure 4, GRPO improved performance over the CoT baseline, achieving an EM of  $13.1\%$  and an F1 score of  $46.5\%$ .
+
+Subsequently, we performed SFT. For this, we employed LoRA (Hu et al., 2022) with a rank of 128 for parameter-efficient training. SFT yielded substantial performance gains over both CoT and GRPO, boosting EM to  $28.0\%$  and F1 to  $55.9\%$  (Figure 4). This demonstrates the strong effectiveness of targeted instruction tuning with SFT for this task in our experiments. While GRPO showed improvement, its gains did not surpass SFT with LoRA. We hypothesize that the effec
+
+![](images/3d14af98b39a13530d3dd4d377bbc67edad3273e89cfa2457fd107c4d2361e19.jpg)  
+Figure 4: Performance comparison of Qwen2.5-VL-3B on the MTabVQA with different post-training strategies.
+
+tiveness of GRPO in this context might be limited by the challenge of defining a more sophisticated reward function than a simple exact match/F1 score, which could better capture nuanced aspects of visual multi-tabular reasoning.
+
+# 4.3 Cross-Dataset Generalization and the Impact of Fine-Tuning Data
+
+In this section, we investigate a central question for instruction tuning: which data source provides the most effective training signal for achieving robust, generalizable performance? To answer this, we analyze how the source and scale of fine-tuning data influence a model's ability to generalize across the different sub-datasets present in our benchmark.
+
+Experimental Setup. We used Qwen2.5-VL-7B as our base VLM for all experiments. We created several fine-tuned model variants by training on different subsets of our MTabVQA-Instruct dataset. These subsets were chosen to isolate the effects of data scale and source diversity (see Table 2 for their origins):
+
+- MiMo+ATIS Subset: A small, diverse set (896 examples).  
+- Spider Subset: A medium-sized, high-quality set (2,395 examples).  
+- MultiTabQA Subset: A large but more narrowly-focused set (10,990 examples).
+
+- Full Instruct Set: The complete, diverse training dataset (15,853 examples), used to train our final TableVision model.
+
+Each of these fine-tuned models was then evaluated on the entire MTabVQA benchmark, testing its performance across all four sub-datasets (Spider, Query, ATIS, MiMo). This cross-dataset evaluation protocol allows us to measure how well training on one data source generalizes to others.
+
+Results and Analysis. The results, detailed in Table 4, reveal a complex relationship between finetuning data and model performance, highlighting that both data diversity and data alignment are more critical than raw data scale alone.
+
+First, we observe strong evidence of domain-specific alignment. The model trained on the Spider subset, for instance, demonstrated the best performance on the corresponding MTabVQA-Spider evaluation sub-dataset (64.3% F1). Similarly, the model fine-tuned on the MiMo+ATIS subset achieved the highest scores on the ATIS (46.5%) and MiMo (39.7%) evaluation sub-datasets. This shows that targeted training on a specific data source is effective at improving performance on in-domain tasks.
+
+However, the source of the fine-tuning data is critically important, as scale alone does not guarantee success. The most striking result is the poor performance of the model trained on the large MultiTabQA subset. Despite being the largest single-source training set (10,990 examples), it yielded the lowest overall F1 score  $(30.2\%)$  besides the zero-shot baseline. While the MultiTabQA data is extensive, its characteristics do not align well with our visually-rich, multi-domain benchmark, likely resulting in a domain shift that harms generalization. This demonstrates that a large volume of narrowly-focused or misaligned data can be less effective than smaller, but more diverse or better-aligned datasets.
+
+Ultimately, the experiments show that combining scale with diversity is the most effective strategy for generalization. The model trained on the full MTabVQA-Instruct dataset, TableVision, achieved the highest overall F1 score  $(68.2\%)$ . By combining multiple data sources, this dataset exposes the model to a wider variety of table structures, question types, and reasoning patterns. This diversity is crucial for building a model that can generalize effectively across the varied scenarios presented in the MTabVQA benchmark.
+
+<table><tr><td rowspan="2">Fine-tuning Subset (Source)</td><td rowspan="2"># Samples</td><td colspan="2">MTabVQA-Spider</td><td colspan="2">MTabVQA-Query</td><td colspan="2">MTabVQA-ATIS</td><td colspan="2">MTabVQA-MiMo</td><td colspan="2">Overall</td></tr><tr><td>EM</td><td>F1</td><td>EM</td><td>F1</td><td>EM</td><td>F1</td><td>EM</td><td>F1</td><td>EM</td><td>F1</td></tr><tr><td>Qwen2.5-VL-7B (Zero-Shot)</td><td>0</td><td>8.0</td><td>39.8</td><td>7.8</td><td>33.9</td><td>6.3</td><td>32.6</td><td>9.3</td><td>22.2</td><td>7.8</td><td>35.1</td></tr><tr><td>MiMo+ATIS Subset</td><td>896</td><td>13.7</td><td>45.7</td><td>11.5</td><td>37.5</td><td>35.7</td><td>46.5</td><td>17.1</td><td>39.7</td><td>13.0</td><td>40.0</td></tr><tr><td>Spider Subset</td><td>2,395</td><td>26.9</td><td>59.2</td><td>49.8</td><td>71.2</td><td>13.4</td><td>22.5</td><td>17.1</td><td>31.9</td><td>41.5</td><td>65.2</td></tr><tr><td>MultiTabQA Subset</td><td>10,990</td><td>10.1</td><td>33.2</td><td>8.7</td><td>28.6</td><td>16.1</td><td>41.9</td><td>11.6</td><td>25.5</td><td>9.4</td><td>30.2</td></tr><tr><td>MTabVQA-Instruct (Full)</td><td>15,853</td><td>32.4</td><td>64.3</td><td>49.8</td><td>72.6</td><td>33.0</td><td>45.9</td><td>20.2</td><td>36.2</td><td>43.4</td><td>68.2</td></tr></table>
+
+Table 4: Performance of fine-tuned models on dataset splits of MTabVQA-Instruct, measuring the influence of the dataset on the overall performance on MTabVQA. Performance is measured in EM and F1. Bold indicates the best overall performance. Underline indicates best performance for each MTabVQA sub-datasets.
+
+# 5 Conclusion
+
+In this work, we introduce MTabVQA, a novel and challenging benchmark specifically designed to evaluate the multi-tabular reasoning capabilities of vision-language models over tables presented as images. MTabVQA, comprising 3,745 QA pairs, focuses on a critical yet underexplored area of integrating and reasoning about information distributed across several table images. This benchmark significantly contributes to bridging the gap between existing table QA benchmarks, which often rely on single or non-visual tables. We evaluated a range of SOTA open-source and proprietary VLMs on MTabVQA, revealing substantial challenges these models face with visual multi-tabular reasoning. To address this, we also release MTabVQA-Instruct, a large-scale instruction-tuning dataset. Our experiments demonstrate that our fine-tuned model, TableVision on the MTabVQA-Instruct dataset, leads to considerable performance improvements on this task. Despite these advancements, the performance of VLMs on MTabVQA indicates significant room for growth, underscoring the complexities of robust visual multi-tabular reasoning and highlighting key areas for future research in developing more capable VLMs.
+
+In future work, we plan to explore more programmatically generated or real-world sourced table images exhibiting even greater visual diversity and degradation to more rigorously test VLM visual parsing and grounding capabilities.
+
+# Limitations
+
+While MTabVQA represents a significant step towards evaluating visual multi-tabular reasoning, we acknowledge several limitations.
+
+English-Only. The current iteration of MTab-VQA is primarily English-centric. Its underlying tabular data, generated questions, and answers
+
+are predominantly in English, which limits the benchmark's applicability for evaluating VLMs on multi-tabular reasoning in other languages. Extending MTabVQA to include multilingual tables and queries would be a valuable contribution, allowing for a more comprehensive assessment of VLM capabilities across diverse linguistic contexts and promoting research in multilingual visual document understanding.
+
+Synthetic Table Layout. While MTabVQA tasks require multi-hop reasoning across table images and incorporate varied visual renderings, the scope of this visual complexity could be further expanded. Real-world documents often contain tables with highly unconventional layouts, extensive cell merging/spanning, embedded charts or icons within cells, and varying image quality (e.g., scanned documents with noise), which makes the task even more challenging for LLMs.
+
+Limited Annotation. To verify that the QA pairs were correct, we used only one annotator to verify the judgments of the LLM agents. Although the annotation was carried out carefully, there may have been minor errors in the data annotation, as it was not double-checked by two people.
+
+# Ethical Considerations
+
+Our work is built upon publicly available academic datasets, and we did not collect any new private or personally identifiable information (PII). The MTabVQA benchmark is designed as a research tool to facilitate the community's evaluation and improvement of multitabular reasoning in vision-language models in a transparent and reproducible manner. We acknowledge that the source datasets, while standard in the field, may contain inherent societal biases that our benchmark could reflect. We encourage users of our dataset to be mindful of these potential issues.
+
+The human verification stage, crucial for ensuring data quality as described in Section 3.5, was conducted by one of the paper's authors and did not involve external crowdworkers. To minimize computational cost and environmental impact, our experiments prioritized parameter-efficient finetuning methods (LoRA) and utilized smaller model variants where appropriate.
+
+# Acknowledgement
+
+This work is supported by the Genial4KMU project, Universität Hamburg, funded by BMBF (grant no. 01IS24044B).
+
+# References
+
+Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sebastien Bubeck, Martin Cai, Qin Cai, Vishray Chaudhary, Dong Chen, Dongdong Chen, and 110 others. 2024. Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone. Preprint, arXiv:2404.14219.  
+Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. 2015. VQA: Visual Question Answering. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV '15, page 2425-2433, USA. IEEE Computer Society.  
+Sara Bonfitto, Elena Casiraghi, and Marco Mesiti. 2021. Table Understanding Approaches for Extracting Knowledge from Heterogeneous Tables. WIREs Data Mining Knowl. Discov., 11(4).  
+Panfeng Cao, Ye Wang, Qiang Zhang, and Zaiqiao Meng. 2023. GenKIE: Robust Generative Multimodal Document Key Information Extraction. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14702-14713, Singapore. Association for Computational Linguistics.  
+Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020. HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1026-1036, Online. Association for Computational Linguistics.  
+Deborah A. Dahl, Madeleine Bates, Michael Brown, William Fisher, Kate Hunicke-Smith, David Pallett, Christine Pao, Alexander Rudnicky, and Elizabeth Shriberg. 1994. Expanding the Scope of the ATIS Task: The ATIS-3 Corpus. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, pages 43-48, New Jersey.
+
+Naihao Deng, Zhenjie Sun, Ruiqi He, Aman Sikka, Yulong Chen, Lin Ma, Yue Zhang, and Rada Mihalcea. 2024. Tables as Texts or Images: Evaluating the Table Reasoning Ability of LLMs and MLLMs. In Findings of the Association for Computational Linguistics: ACL 2024, pages 407-426, Bangkok, Thailand. Association for Computational Linguistics.  
+Demis Hassabis, Koray Kavukcuoglu, and Google DeepMind. 2024. Introducing Gemini 2.0: Our New AI Model for the Agentic Era. Blog post, Google DeepMind.  
+Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. 2024. WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6864–6890, Bangkok, Thailand. Association for Computational Linguistics.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In The Tenth International Conference on Learning Representations, 2022, Virtual Event. ICLR 2022.  
+Mohit Iyyer, Wen-tau Yih, and Ming-Wei Chang. 2017. Search-based Neural Structured Learning for Sequential Question Answering. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1821-1831, Vancouver, Canada. Association for Computational Linguistics.  
+Duff Johnson. 2018. PDF Statistics - the Universe of Electronic Documents. PDF Association presentation.  
+Larissa R. Lautert, Marcelo M. Scheidt, and Carina F. Dorneles. 2013. Web Table Taxonomy and Formalization. SIGMOD Rec., 42(3):28-33.  
+Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2025a. LLaVA-OneVision: Easy Visual Task Transfer. Transactions on Machine Learning Research.  
+Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, Xuanhe Zhou, Ma Chenhao, Guoliang Li, Kevin Chang, Fei Huang, Reynold Cheng, and Yongbin Li. 2023a. Can LLM Already Serve as a Database Interface? A Big Bench for Large-Scale Database Grounded Text-to-SQLs. In Advances in Neural Information Processing Systems, volume 36, pages 42330-42357. Curran Associates, Inc.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023b. BLIP-2: Bootstrapping Language-Image Pretraining with Frozen Image Encoders and Large Language Models. In Proceedings of the 40th International Conference on Machine Learning, volume 202
+
+of Proceedings of Machine Learning Research, pages 19730-19742. PMLR.  
+Peng Li, Yeye He, Dror Yashar, Weiwei Cui, Song Ge, Haidong Zhang, Danielle Rifinski Fainman, Dongmei Zhang, and Surajit Chaudhuri. 2024. Table-GPT: Table Fine-tuned GPT for Diverse Table Tasks. Proc. ACM Manag. Data, 2(3):176.  
+Zheng Li, Yang Du, Mao Zheng, and Mingyang Song. 2025b. MiMoTable: A Multi-scale Spreadsheet Benchmark with Meta Operations for Table Reasoning. In Proceedings of the 31st International Conference on Computational Linguistics, pages 2548-2560, Abu Dhabi, UAE. Association for Computational Linguistics.  
+Shicheng Liu, Sina Semnani, Harold Triedman, Jialiang Xu, Isaac Dan Zhao, and Monica Lam. 2024. SPINACH: SPARQL-Based Information Navigation for Challenging Real-World Questions. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 15977-16001, Miami, Florida, USA. Association for Computational Linguistics.  
+Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. 2024. LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15630-15640, Seattle WA, USA.  
+Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kryscinski, Hailey Schoelkopf, Riley Kong, Xiangru Tang, Mutethia Mutuma, Ben Rosand, Isabel Trindade, Renusree Bandaru, Jacob Cunningham, Caiming Xiong, and Dragomir R. Radev. 2022. FeTaQA: Free-form Table Question Answering. Trans. Assoc. Comput. Linguistics, 10:35-49.  
+Ahmed Nassar, Nikolaos Livathinos, Maksym Lysak, and Peter Staar. 2022. TableFormer: Table Structure Understanding with Transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4604-4613, Los Alamitos, CA, USA. IEEE Computer Society.  
+OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, and 262 others. 2024. GPT-4 Technical Report. arXiv preprint. ArXiv:2303.08774.  
+Vaishali Pal, Andrew Yates, Evangelos Kanoulas, and Maarten de Rijke. 2023. MultiTabQA: Generating Tabular Answers for Multi-Table Question Answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6322–6334, Toronto, Canada. Association for Computational Linguistics.
+
+Panupong Pasupat and Percy Liang. 2015. Compositional Semantic Parsing on Semi-Structured Tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1470–1480, Beijing, China. Association for Computational Linguistics.  
+Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. Preprint, arXiv:2402.03300.  
+Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. 2024. Table Meets LLM: Can Large Language Models Understand Structured Table Data? A Benchmark and Empirical Study. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, WSDM '24, page 645-654, New York, NY, USA. Association for Computing Machinery.  
+Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Rame, Morgane Riviere, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, and 197 others. 2025a. Gemma 3 Technical Report. Preprint, arXiv:2503.19786.  
+Qwen Team, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, and 24 others. 2025b. Qwen2.5 Technical Report. Preprint, arXiv:2412.15115.  
+Jian Wu, Linyi Yang, Dongyuan Li, Yuliang Ji, Manabu Okumura, and Yue Zhang. 2025a. MMQA: Evaluating LLMs with Multi-Table Multi-Hop Complex Questions. In International Conference on Representation Learning, volume 2025, pages 48626-48643, Singapore.  
+Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xeron Du, Di Liang, Daixin Shu, Xi'anfu Cheng, Tianzhen Sun, Tongliang Li, Zhoujun Li, and Guanglin Niu. 2025b. TableBench: A Comprehensive and Complex Benchmark for Table Question Answering. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence, AAAI'25/IAAI'25/EAAI'25, pages 25497-25506. AAAI Press.  
+Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. Spider: A Large-Scale Human-Labeled
+
+Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911-3921, Brussels, Belgium. Association for Computational Linguistics.  
+Jiaxin Zhang, Wentao Yang, Songxuan Lai, Zecheng Xie, and Lianwen Jin. 2025. DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence, AAAI'25/IAAI'25/EAAI'25, pages 9923-9932. AAAI Press.  
+Tianshu Zhang, Xiang Yue, Yifei Li, and Huan Sun. 2024a. TableLlama: Towards Open Large Generalist Models for Tables. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6024-6044, Mexico City, Mexico. Association for Computational Linguistics.  
+Weijia Zhang, Vaishali Pal, Jia-Hong Huang, Evangelos Kanoulas, and Maarten de Rijke. 2024b. QFMTS: Generating Query-Focused Summaries over Multi-Table Inputs. In ECAI 2024 - 27th European Conference on Artificial Intelligence, Including 13th Conference on Prestigious Applications of Intelligent Systems (PAIS 2024), volume 392 of Frontiers in Artificial Intelligence and Applications, pages 3875-3882, Santiago de Compostela, Spain. IOS Press.  
+Mingyu Zheng, Xinwei Feng, Qingyi Si, Qiaoqiao She, Zheng Lin, Wenbin Jiang, and Weiping Wang. 2024. Multimodal Table Understanding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 9102-9124, Bangkok, Thailand3. Association for Computational Linguistics.  
+Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning. Preprint, arXiv:1709.00103.  
+Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and TatSeng Chua. 2021. TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3277-3287, Online. Association for Computational Linguistics.  
+Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, and 32
+
+others. 2025. InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models. Preprint, arXiv:2504.10479.
+
+# A Relational Table Sampling
+
+Algorithm 1 details our method for creating smaller, interconnected samples from large databases. We start by randomly selecting a limited number of rows (up to a maximum,  $N_{max} = 50$ ) from one initial table. Then, using the database's foreign keys, we identify other tables linked to this first one. When sampling from these linked tables, the crucial step is to find and prioritize rows that are directly related to the rows already chosen from the previous table. This is achieved by matching values in the specific columns that link the tables. This process of finding related data and sampling continues as the algorithm explores outwards to other connected tables, ensuring the final set of sampled tables forms a related subset of the original database.
+
+Algorithm 1 Relational Table Sampling
+
+Input:
+
+$\mathcal{D}$  : Input database (collection of tables)
+
+$\mathcal{R}$  : Set of foreign key relationships between tables in  $\mathcal{D}$
+
+$N_{max}$ : Maximum number of rows per sampled table
+
+$(V$ : Set of table identifiers derived from  $\mathcal{D}$ )
+
+$(G = (V,E)$  : Relationship graph derived from  $\mathcal{D}$  and  $\mathcal{R}$
+
+Output:
+
+$\mathcal{S}$ : Set of pairs  $(t, S_t)$ , where  $S_t$  is the sampled row subset for table  $t \in V$ .
+
+1:  $\mathcal{S}\gets \emptyset ;\mathcal{P}\gets \emptyset$
+
+$\triangleright S$ : Output samples,  $\mathcal{P}$ : Processed tables set
+
+2:  $t_{start} \gets \operatorname{SelectSeed}(V, G)$
+
+$\triangleright$  Select a starting table (e.g., highest degree)
+
+3:  $S_{t_{start}} \gets \text{Sample}(t_{start}, N_{max})$
+
+$\triangleright$  Sample initial rows for  $t_{start}$
+
+4:  $\mathcal{S} \gets \{(t_{start}, S_{t_{start}})\}$ ;  $\mathcal{P} \gets \{t_{start}\}$
+
+$\triangleright$  Update output set and processed set
+
+5: Initialize  $Q$ ;  $Q$ .Enqueue  $(t_{start})$
+
+$\triangleright Q$  : Queue for Breadth-First Search (BFS)
+
+6: while  $Q$  is not empty do
+
+$\triangleright$  Perform BFS traversal
+
+7:  $t_{\text{curr}} \gets Q.\text{Dequeue()}$
+
+$\triangleright t_{curr}$ : Current table being processed
+
+8: for each  $t_{rel} \in \mathrm{Neighbors}(t_{curr}, G) \setminus \mathcal{P}$  do
+
+$\triangleright t_{rel}$ : Related, unprocessed neighbor table
+
+9:  $R_{\text{linked}} \gets \text{GetLinkedRows}(t_{\text{rel}}, t_{\text{curr}}, S_{t_{\text{curr}}}, \mathcal{R})$
+
+$\triangleright$  Get rows in  $t_{rel}$  linked to sampled rows
+
+10:  $S_{t_{rel}} \gets \text{SampleSubset}(R_{\text{linked}}, N_{\text{max}})$
+
+a subset from linked rows, max size  $N_{max}$
+
+11:  $\mathcal{S}\gets \mathcal{S}\cup \{(t_{rel},S_{t_{rel}})\}$
+
+> Add the new sample to the output
+
+12:  $\mathcal{P}\gets \mathcal{P}\cup \{t_{rel}\} ;Q.\mathrm{Enqueue}(t_{rel})$
+
+$\triangleright$  Mark  $t_{rel}$  as processed and add to queue
+
+13: end for
+
+14: end while
+
+15: return S
+
+$\triangleright$  Return the final set of sampled table subsets
+
+# B Data Sourcing: Join and Filter Details
+
+This section provides a detailed breakdown of the process used to identify and filter data instances requiring multi-table join operations from the source datasets, as mentioned in Section 3.1. This formed the basis for constructing both the MTabVQA and MTabVQA-Instruct splits, ensuring a focus on multi-tabular reasoning. The primary method involved parsing SQL queries associated with text-to-SQL datasets to detect explicit join clauses (e.g., 'JOIN', 'INNER JOIN', 'LEFT JOIN'). For datasets without explicit SQL, we relied on provided metadata or question characteristics indicative of multi-table requirements.
+
+# B.1 Spider Dataset
+
+The Spider dataset (Yu et al., 2018) is a large-scale text-to-SQL benchmark. We analyzed its train, development (dev), and test splits to identify questions whose corresponding SQL queries involved joins.
+
+# - Train Split:
+
+Total Questions: 7,000  
+- Questions with SQL Joins: 2,771  
+- Selected for MTabVQA-Instruct (after filtering and processing): 2,395 instances.
+
+# - Development (Dev) Split:
+
+- Total Questions: 1,034  
+- Questions with SQL Joins: 408
+
+# Test Split:
+
+- Total Questions: 2,147  
+- Questions with SQL Joins: 862
+
+# - MTabVQA (from Spider Dev/Test):
+
+- Combined Join Questions from Dev & Test: 408 (Dev) + 862 (Test) = 1,270  
+- Selected for MTabVQA (MTabVQA-Spider-Eval split): 1,048 instances. These were chosen from the 1,270 join questions based on criteria ensuring clear multi-hop reasoning paths, unambiguous answers from sampled data, and visual representability.
+
+# B.2 QFMTS Dataset
+
+The QFMTS dataset (Zhang et al., 2024b) focuses on query-focused multi-document summarization with tables. We identified instances requiring information synthesis across multiple tables.
+
+- Total Questions/Instances: 4,908  
+- Instances Identified as Requiring Multi-Table Reasoning (e.g., via SQL joins or inherent task nature): 2,578  
+- Selected for MTabVQA (MTabVQA-Query-Eval split, primarily from QFMTS): 2,456 instances. Filtering ensured complexity and suitability for our visual QA benchmark.
+
+# B.3 BIRD Dataset
+
+BIRD (Li et al., 2023a) is another challenging text-to-SQL benchmark designed to evaluate robustness on large databases and complex queries.
+
+- Total Identified SQL Join Queries (approx.): 7,900  
+- Generated QA pairs for MTabVQA-Instruct: 1,572 instances. These were generated from a diverse selection of the join queries, focusing on creating complex multi-hop reasoning scenarios suitable for instruction tuning.
+
+# B.4 MultiTabQA Dataset
+
+The MultiTabQA dataset (Pal et al., 2023) is specifically designed for question answering over multiple tables.
+
+- Total QA pairs involving joins/multi-table lookups utilized: 10,990  
+- These were directly incorporated into the MTabVQA-Instruct dataset due to their inherent multi-table nature.
+
+# B.5 ATIS Dataset
+
+The Air Travel Information System (ATIS) dataset (Dahl et al., 1994) contains spoken language queries related to flight information, often mapped to relational database queries.
+
+Total Questions Analyzed: 496  
+- Instances identified/selected for MTabVQA (MTabVQA-Atis split): 112  
+- Instances selected/generated for MTabVQA-Instruct: 384 (See Table 2).
+
+# B.6 MiMoTable Dataset
+
+The MiMoTable dataset (Li et al., 2025b) focuses on multimodal table understanding.
+
+Total Questions/Instances: 1,636  
+- Questions Identified with Multi-Table Requirements (e.g., from problem descriptions or metadata indicating cross-table information needed): 641  
+- Selected for MTabVQA-Instruct: 512 instances.  
+- Selected for MTabVQA: 129 instances.
+
+# B.7 Overall Summary
+
+Across all source datasets, we identified approximately 26,826 potential questions or instances that involved multi-table join operations or inherently required multi-table reasoning. Through our processing, filtering, and generation pipeline, a total of 19,608 high-quality, multi-tabular visual question-answering instances were curated to form the MTabVQA (3,745 pairs) and MTabVQA-Instruct (15,853 pairs, with some overlap in underlying source tables but disjoint QA pairs) datasets. The filtering criteria included ensuring genuine multi-hop reasoning, clarity of questions and answers, visual representability of the involved tables, and overall quality for benchmarking and instruction tuning.
+
+# C Verification Prompt
+
+The following prompt was provided to the verification LLM-based verification agents during the automated assessment phase described in Section 3.5.
+
+```txt
+You are a verification agent for table-based question answering.   
+You need to verify if the answer and reasoning for the given   
+question are correct based ONLY on the provided table data.   
+[Tables Used] [Sampled Table Data (JSON Format)]   
+[Question-Answer Pair] Question: [Generated Question Text] Answer: [Generated Answer (JSON Format)] Reasoning Steps: [Generated Reasoning Steps] Question Type: [Designated Question Type]   
+Your task: 1. Check if the question is well-formed and genuinely requires multi-hop reasoning across MULTIPLE provided tables. Single-table questions are invalid.. 2. Verify if the answer is accurate based only on the information present in the given tables. If the answer is incorrect, 'is_valid' must be 'false'. 3. Check if the 'tables_used' field correctly lists relevant tables and if at least two tables were necessary.. 4. Validate if the reasoning steps are logical, coherent, and correctly lead from the table data to the answer..   
+Respond with ONLY a valid JSON object (no introductory text, markdown formatting, or code blocks outside the JSON structure) containing the following keys: {{ "is_valid": true/false, "verification_comments": "Your detailed verification comments explaining the validity/issues and multi-table requirement.", "score": <an integer score from 0 to 10, where 10 is perfect adherence to all criteria>, "uses_multiple_tables": true/false }
+```
+
+Figure 5: LLM prompt for automated QA pair verification. Placeholders like '[Generated Question Text]' represent the actual data provided to the model.
+
+# D Visual Table Rendering Details
+
+As described in Section 3.3, MTabVQA table images were generated with significant visual diversity to mimic real-world appearances. For each QA pair, the rendering process introduced controlled variations across several dimensions using 10 distinct styling themes, randomly selected per table. These themes systematically varied:
+
+# - Structure and Layout:
+
+- Column widths and row heights were adapted to content to ensure readability while introducing natural variations.  
+- The relative positioning of multiple table images within the final visual context presented to the model was also varied (e.g., tables rendered side-by-side, stacked vertically, or with other layout configurations).
+
+- Appearance (Themes, Fonts, Styles): The 10 distinct styling themes systematically manipulated the following:
+
+- Color Schemes: This included variations in header background colors (e.g., using specific hex codes like #4CAF50 (green), #1E88E5 (blue), #333 (dark grey)), cell background colors, text colors (e.g., white text on dark headers, black text on light backgrounds), and alternating row shading ('zebra striping' with colors like #f2f2f2).  
+- Typography: Different font families (e.g., common serif and sans-serif fonts) were used. Font weights were varied (e.g., bold headers, normal weight for cell content). Font sizes were adjusted within themes (e.g., a base size of 12pt in one theme, with relative adjustments for headers).  
+- Styling Elements: The presence, style, and color of grid lines were varied (e.g., solid lines, dashed lines, varying thickness, or minimalist themes with no grid lines). Cell padding was adjusted to control spacing within cells. Border styles for the overall table and individual cells were also diversified (e.g., 1px solid black, 2px solid #000, or no borders).
+
+This deliberate introduction of visual diversity is key to challenging models on robust OCR and layout understanding across varied presentations before they engage in multi-tabular reasoning.
+
+# E Human Verification Interface
+
+Figure 6 shows the interface of the Streamlit application used for the final human verification stage (Section 3.5). This tool displayed the rendered table images, the generated question, the LLM-generated answer and reasoning, and the automated verification scores, allowing reviewers to make the final acceptance decision.
+
+![](images/6781b9dbe53f17576110441ad2e81c66d54e16b1c40f6ae6d7cc684f6b111ac2.jpg)  
+(a) Initial login screen for evaluator identification.
+
+![](images/be4f425a71556fd90385d35f1478fb88d4cec5099fff83832de6ecfa7d51c551.jpg)  
+(b) Main evaluation screen displaying table images, question, predicted answer, and reviewer judgment options.  
+Figure 6: Screenshots of the Streamlit application interface used for human verification. Panel (a) shows the user login step, and panel (b) presents the core evaluation interface with table images and QA details.
+
+# F Taxonomy of Reasoning Categories and Performance Analysis
+
+To ensure a diverse and challenging benchmark, we employed a taxonomy-guided approach for generating the QA pairs. This taxonomy defines distinct reasoning skills that models must demonstrate. This section provides a detailed breakdown of the most prominent categories, complete with illustrative examples including rendered tables, and an analysis of how leading proprietary models perform on each.
+
+# F.1 Description of Reasoning Categories with Examples
+
+Below are descriptions and examples for the six most frequent reasoning categories in the MTabVQA benchmark.
+
+Descriptive Analysis. This is a broad category where the goal is to retrieve and present a set of data based on specified criteria. Unlike fact-checking, which often seeks a single value, descriptive analysis requires the model to gather multiple rows and columns of information, often by joining tables, and format them into a comprehensive list. It tests the model's core ability to parse what is being asked and retrieve a complete set of relevant, multi-column data.
+
+# Example: Descriptive Analysis
+
+Question: What are the names of players and the corresponding clubs that they are in? Golden Answer: ["Nick Price", "Arsenal"], ["Paul Azinger", "Blackburn Rovers"], ...]
+
+<table><tr><td colspan="3">Table: Player</td><td colspan="2">Table: Club</td></tr><tr><td>Player_ID</td><td>Name</td><td>Club_ID</td><td>Club_ID</td><td>Name</td></tr><tr><td>1</td><td>Nick Price</td><td>1</td><td>1</td><td>Arsenal</td></tr><tr><td>2</td><td>Paul Azinger</td><td>2</td><td>2</td><td>Blackburn Rovers</td></tr><tr><td>3</td><td>Greg Norman</td><td>5</td><td>3</td><td>Chelsea</td></tr><tr><td>4</td><td>Jim Gallagher, Jr.</td><td>1</td><td>4</td><td>Everton</td></tr><tr><td>5</td><td>David Frost</td><td>5</td><td>5</td><td>Fulham</td></tr></table>
+
+Multi-hop Fact Checking. This is a more complex form of fact-checking that requires a chain of two or more reasoning steps (hops). For instance, a model might need to join Table A to Table B to get an intermediate result, and then use that result to filter or join with Table C to find the final answer.
+
+# Example: Multi-hop Fact Checking
+
+Question: Which countries' TV channels are playing some cartoon written by Todd Casey? Golden Answer: ["United Kingdom"], ["Italy"]]
+
+Table: TV_Channel  
+
+<table><tr><td>id</td><td>Country</td></tr><tr><td>1</td><td>Italy</td></tr><tr><td>2</td><td>Poland</td></tr><tr><td>3</td><td>United Kingdom</td></tr></table>
+
+Table: Cartoon  
+
+<table><tr><td>id</td><td>Title</td><td>Writer</td><td>Channel</td></tr><tr><td>1</td><td>The Cow...</td><td>Michael...</td><td>2</td></tr><tr><td>2</td><td>Ben 10</td><td>Todd Casey</td><td>1</td></tr><tr><td>3</td><td>The Life...</td><td>Todd Casey</td><td>3</td></tr></table>
+
+Match-Based Fact Checking. These tasks require the model to perform straightforward lookups and joins across multiple tables to find a specific piece of information. The primary challenge is correctly identifying the keys to join the tables and retrieving the corresponding value without complex calculations.
+
+# Example: Match-Based Fact Checking
+
+Question: What is the name of the director who is in the "Dracula" program? Golden Answer: ["Hank Baskett"]
+
+Table : Director  
+
+<table><tr><td>Dir_ID</td><td>Name</td><td>Age</td></tr><tr><td>1</td><td>DeSean Jackson</td><td>45</td></tr><tr><td>2</td><td>Hank Baskett</td><td>48</td></tr><tr><td>3</td><td>Greg Lewis</td><td>52</td></tr></table>
+
+Table: Program  
+
+<table><tr><td>Prog_ID</td><td>Title</td><td>Dir_ID</td></tr><tr><td>1</td><td>The Angry Brigade</td><td>1</td></tr><tr><td>2</td><td>Dracula</td><td>2</td></tr><tr><td>3</td><td>Another Country</td><td>3</td></tr></table>
+
+Ranking. These tasks require the model to sort a set of results based on a specific criterion (e.g., highest value, most frequent occurrence) and often retrieve the top-N results. This involves both extraction and comparison logic.
+
+# Example: Ranking
+
+Question: Find the name of the director who is in charge of the most programs. Golden Answer: ["Greg Lewis"]
+
+Table: Director  
+
+<table><tr><td>ID</td><td>Name</td></tr><tr><td>1</td><td>DeSean Jackson</td></tr><tr><td>2</td><td>Hank Baskett</td></tr><tr><td>3</td><td>Greg Lewis</td></tr></table>
+
+Table: Program  
+
+<table><tr><td>Title</td><td>Director_ID</td></tr><tr><td>The Angry Brigade</td><td>1</td></tr><tr><td>Dracula</td><td>2</td></tr><tr><td>Another Country</td><td>3</td></tr><tr><td>Othello</td><td>3</td></tr></table>
+
+Aggregation. Aggregation tasks require the model to perform mathematical operations over a set of table rows, such as 'COUNT', 'SUM', or 'AVG'. This tests the model's ability to not only extract data but also to apply numerical reasoning to it.
+
+# Example: Aggregation
+
+Question: What is the average age of all the club leaders?
+
+Golden Answer: ["20.166..."]]
+
+Table: Member  
+
+<table><tr><td>Mem_ID</td><td>Age</td></tr><tr><td>1</td><td>18</td></tr><tr><td>2</td><td>20</td></tr><tr><td>3</td><td>21</td></tr><tr><td>...</td><td>...</td></tr></table>
+
+Table: Club_Leader  
+
+<table><tr><td>Club_ID</td><td>Mem_ID</td><td>Since</td></tr><tr><td>1</td><td>1</td><td>2018</td></tr><tr><td>2</td><td>2</td><td>2017</td></tr><tr><td>3</td><td>3</td><td>2018</td></tr><tr><td>…</td><td>…</td><td>…</td></tr></table>
+
+Counting. A specific and frequent type of aggregation, counting tasks require the model to enumerate the number of items that satisfy a certain condition. This can range from simple counts to more complex conditional counts after a join.
+
+# Example: Counting
+
+Question: How many movies are playing in theater Odeon?
+
+Golden Answer: [["2"]]
+
+Table: Movies  
+
+<table><tr><td>MovieID</td><td>Title</td></tr><tr><td>101</td><td>The Wizard of Oz</td></tr><tr><td>102</td><td>A Night at the Opera</td></tr><tr><td>103</td><td>North by Northwest</td></tr><tr><td>104</td><td>Citizen Kane</td></tr><tr><td>105</td><td>The Quiet Man</td></tr></table>
+
+Table: MovieTheaters  
+
+<table><tr><td>Theater</td><td>MovieID</td></tr><tr><td>Odeon</td><td>103</td></tr><tr><td>Imperial</td><td>104</td></tr><tr><td>Royale</td><td>106</td></tr><tr><td>Odeon</td><td>101</td></tr></table>
+
+# F.2 Performance by Reasoning Category
+
+To better understand the strengths and weaknesses of current VLMs, we analyzed the performance of GPT-4.1 and Gemini across the main reasoning categories. The overall performance of GPT-4.1 on the MTabVQA benchmark is strong, with a weighted F1 score of  $61.7\%$ . However, as shown in Figure 7, performance varies significantly by reasoning type.
+
+![](images/0eed8d5943d56ac2684bacf4dd41c63a7367371e1310a0840972489dbba151d6.jpg)  
+Figure 7: Performance (F1 Score) of GPT-4.1 and Gemini-2.0-Flash on the six most prominent reasoning categories within the MTabVQA benchmark. Both models excel at descriptive and fact-checking tasks but show weaker performance on numerical reasoning tasks like counting and aggregation.
+
+The analysis reveals several key trends. Both models demonstrate the highest proficiency on Descriptive Analysis and Fact Checking tasks, with F1 scores often exceeding 0.65. This indicates that current models are adept at understanding natural language queries, performing table joins, and retrieving textual data.
+
+Conversely, performance drops noticeably on tasks requiring explicit numerical reasoning. Aggregation and, particularly, Counting represent the most significant challenges, with F1 scores falling below 0.57 for both models. This suggests that while VLMs can successfully parse and correlate data, they are more prone to errors when required to perform precise mathematical operations. This highlights a key area for future research: improving the quantitative reasoning capabilities of VLMs in complex, multi-table visual contexts.
+
+# G GRPO Training Details
+
+This section provides additional details on the Group Relative Policy Optimization (GRPO) (Shao et al., 2024) experiments discussed in Section 4.2 for fine-tuning the Qwen2.5-VL-3B model. We utilized the EasyR1 framework<sup>3</sup> for these experiments, training for a total of 270 steps. The training was conducted on a subset of MTabVQA-Instruct derived from the Spider dataset (2,395 examples).
+
+![](images/744baf9f2d0149a8d7bafafc9f38bdd184de4fd3199c227ca975f53ecb4495be.jpg)  
+(a) Reward: Accuracy (EM)
+
+![](images/33e58a7564874eb0047ad0d1f2359b23f04f50b30fac343b2879cd828a27f986.jpg)  
+(b) Reward: F1 Score
+
+![](images/fa4e7bfd10ffd191959e7f63567f40449f99e39c87b8bd6fc2cdcd1adbac2e38.jpg)  
+(c) Reward: Structural Format
+
+![](images/bd0237837ca36104ba9ab2a20381935fdc523a2feef56b2d139b0ee1a6b76523.jpg)  
+(d) Reward: JSON Format
+
+![](images/651b9656d844fe79f0ae0764bbc0b3f73a57b650ada424a026c400070ba33acf.jpg)  
+(e) Reward: Overall  
+Figure 8: GRPO training reward component curves for Qwen2.5-VL-3B over 270 training steps. These plots illustrate the learning progress for content accuracy (EM, F1), structural format adherence, JSON validity, and the combined overall reward.
+
+Reward Function: The reward function for GRPO was designed to encourage both semantic correctness and proper output formatting. It was a composite score derived from:
+
+- Content Correctness: Assessed by the weighted sum of Exact Match (EM) and F1 score between the generated answer and the ground truth.  
+- Format Adherence: This included two components:
+
+- Structural Format Score: A binary score indicating whether the model's output correctly included the required ‘<think>' and ‘<answer>' tags.  
+- JSON Format Score: A binary score indicating whether the content within the `<answer>` tags was valid JSON.
+
+The overall reward signal aimed to maximize these components, guiding the model towards generating accurate and well-formatted responses.
+
+Figure 8 shows the progression of various reward components during the GRPO training process. The plots for 'reward/accuracy' (EM) and 'reward/f1' show a general upward trend, indicating learning of
+
+content correctness. The 'reward/format' and 'reward/json' plots demonstrate that the model quickly learned to adhere to the specified output structure. The 'reward/overall' plot reflects the combined learning signal. The final checkpoint used for evaluation was selected based on the highest 'reward/overall' achieved during training. These settings were chosen to balance training stability, computational efficiency, and exploration during the reinforcement learning process for the multi-tabular visual question answering task, aiming for both accurate content and correctly formatted output. Key GRPO training parameters are summarized in Table 5.
+
+<table><tr><td>Parameter</td><td>Value</td></tr><tr><td colspan="2">Core Algorithm</td></tr><tr><td>Advantage Estimator</td><td>GRPO</td></tr><tr><td>KL Coefficient (λKL)</td><td>0.01</td></tr><tr><td colspan="2">Training Setup</td></tr><tr><td>Base Model</td><td>Qwen/Qwen2.5-VL-3B-Instruct</td></tr><tr><td>Training Data</td><td>MTabVQA-Instruct (Spider Subset) (2,395 ex.)</td></tr><tr><td>Max Training Steps</td><td>270</td></tr><tr><td>Total Epochs</td><td>15</td></tr><tr><td>Rollout Batch Size</td><td>128</td></tr><tr><td colspan="2">Actor Model (Qwen2.5-VL-3B)</td></tr><tr><td>Learning Rate</td><td>1e-06</td></tr><tr><td>Optimizer</td><td>AdamW (BF16)</td></tr><tr><td>Global Update Batch Size</td><td>32</td></tr><tr><td colspan="2">Rollout Generation</td></tr><tr><td>Temperature (Training)</td><td>1.0</td></tr><tr><td>Top-p (Training)</td><td>0.99</td></tr><tr><td>Num. Generations per Prompt (n)</td><td>5</td></tr></table>
+
+Table 5: GRPO Hyperparameters for Qwen2.5-VL-3B Fine-tuning.
+
+# H Model Evaluation and Generation Prompts
+
+This section details the system prompts used for evaluating and generating responses from the Vision-Language Models (VLMs) in different experimental settings.
+
+# H.1 Standard Zero-Shot Evaluation Prompt
+
+For standard zero-shot evaluations of VLMs (Section 4.1), including proprietary models and open-source baselines before specific post-training, the following system prompt was used. This prompt instructs the model on how to interpret multi-tabular image data, reason about the question, and provide an answer strictly in the specified JSON format.
+
+# System Prompt: Zero-Shot Evaluation
+
+You are an intelligent assistant capable of understanding and reasoning about multi-tabular data given as images, each table is one image. You will be presented with one or more tables containing information on a specific topic.
+
+You will then be asked a question that requires you to analyze the data in the table(s) and provide a correct answer in strict required format.
+
+Your task is to:
+
+1. Carefully examine the provided table(s) Pay close attention to the column headers, the data types within each column, and the relationships between tables if multiple tables are given.  
+2. Understand the question being asked. Identify the specific information being requested and determine which table(s) and columns are relevant to answering the question.  
+3. Extract the necessary information from the table(s). Perform any required filtering, joining, aggregation, or calculations on the data to arrive at the answer.  
+4. Formulate a clear and concise answer in natural language. The answer should be directly responsive to the question and presented in a human-readable format. It may involve listing data, presenting a single value, or explaining a derived insight.  
+5. Do not include any SQL queries in the answer. But you can use it internally, to come up with answer. 6. Be accurate and avoid hallucinations. Your answer should be completely based on the data in the provided table(s). Do not introduce any external information or make assumptions not supported by the data.  
+7. Be specific and follow the instructions in the question. If the question ask to get specific columns, return only mentioned columns.  
+8. If the question is unanswerable based on the provided tables, state "The question cannot be answered based on the provided data.  
+9. Please provide only the answer which has been asked, without any additional text (try to use few tokens). However, take the time to think and reason before giving your answer. Also, try to provide an answer even if you are unsure.  
+10. Provide the answer in JSON format with given response schema as given [['ans1', 'ans2'], ['ans3', 'ans4']]. Respond only with valid JSON format.
+
+Take your time to understand the question. Break it down into smaller steps. Come up with an answer and examine your reasoning. Finally, verify your answer. you need to extract answers based on the given multi-hop question [Question] and given multiple tables [TABLE1], and [TABLE2]. Please only output the results without any other words. Return the answer in the following JSON format.
+
+```txt
+Return the answer in JSON schema: : { "type": "json_schema", "json_schema": { "name": "Response", "type": "object", "properties": { "data": { "type": "array", "items": {"type": "array", "items": {"type": "string"}}, }, }, "required": ["data"], "additionalProperties": False, },
+```
+
+# H.2 Chain-of-Thought (CoT) Evaluation Prompt
+
+For the Chain-of-Thought (CoT) prompting experiments (Section 4.2), a modified system prompt was used. This prompt explicitly instructs the model to first generate a step-by-step reasoning process (the chain of thought) and then provide the final answer.
+
+# System Prompt: Chain-of-Thought (CoT)
+
+You are an intelligent assistant capable of understanding and reasoning about multi-tabular data given as images, each table potentially being one image. You will be presented with one or more tables containing information on a specific topic. You will then be asked a question that requires you to analyze the data in the table(s) and provide a correct answer in the strictly required format.
+
+Your task is to:
+
+1. Carefully examine the provided table(s): Pay close attention to the column headers, the data types within each column, and the relationships between tables if multiple tables are given.  
+2. Understand the question being asked: Identify the specific information being requested and determine which table(s) and columns are relevant to answering the question.  
+3. Reason step-by-step (Chain of Thought): Before generating the final answer, formulate a clear chain of thought outlining how you identified the relevant data, performed necessary operations (filtering, joining, aggregation, calculations),  
+and arrived at the result. This reasoning is crucial and MUST be included in the final output.  
+4. Extract the necessary information from the table(s): Perform any required filtering, joining, aggregation, or calculations on the data based on your chain of thought to arrive at the answer. 5. Do not include any SQL queries in the final answer JSON. You can use SQL logic internally during your reasoning (Chain of Thought), but the final output should not contain raw SQL code.  
+6. Be accurate and avoid hallucinations: Your answer must be completely based on the data in the provided table(s).
+
+. Provide the output strictly in the specified JSON format: The output must be a single JSON object containing two keys: `chain_of Thought` (a string detailing your reasoning steps) and `data` (an array of arrays containing the answer).
+
+Your entire response must be ONLY a valid JSON string conforming to the schema below. JSON Schema:
+
+```txt
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+```
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``"
+``}
+```
+
+Take your time to understand the question and the data. Break the problem down using Chain of Thought. Construct the final JSON containing both your reasoning and the extracted data. Verify your answer and the format before outputting. Remember to output ONLY the JSON string.
+
+# H.3 GRPO Thinking Prompt
+
+For the Group Relative Policy Optimization (GRPO) training and generation (Section 4.2 and Appendix G), the prompt was used. This prompt is similar to the CoT prompt in that it requires an internal reasoning process ('<think>...</think>'') before the final answer, but it is specifically tailored for the GRPO framework, which often involves distinct markers for thought processes versus final outputs used in reward calculation. The final answer is expected within '<answer>...</answer>' tags in a specific JSON format.
+
+# System Prompt: GRPO Thinking Prompt
+
+You are an intelligent assistant capable of understanding and reasoning about multi-tabular data given as images, each table is one image. You will be presented with one or more tables containing information on a specific topic.
+
+You will then be asked a question that requires you to analyze the data in the table(s) and provide a correct answer in strict required format using multi-hop reasoning.
+
+Your task is to:
+
+1. Carefully examine the provided table(s) Pay close attention to the column headers, the data types within each column, and the relationships between tables if multiple tables are given.  
+2. Understand the question being asked. Identify the specific information being requested and determine which table(s) and columns are relevant to answering the question.  
+3. Extract the necessary information from the table(s). Perform any required filtering, joining, aggregation, or calculations on the data to arrive at the answer.  
+4. Formulate a clear and concise answer in natural language. The answer should be directly responsive to the question and presented in a human-readable format.
+
+It may involve listing data, presenting a single value, or explaining a derived insight.
+
+5. Do not include any SQL queries in the answer. But you can use it internally, to come up with answer.  
+6. Be accurate and avoid hallucinations. Your answer should be completely based on the data in the provided table(s). Do not introduce any external information or make assumptions not supported by the data.  
+7. Provide the answer in JSON format with given response schema as given [['ans1', 'ans2'], ['ans3', 'ans4']]. Respond only with valid JSON format, as shown in the example above.
+
+Strictly, Give answer in this format, using the example below as reference:
+
+You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. You will be presented with one or more tables containing information on a specific topic. You will then be asked a question that requires you to analyze the data in the table(s) and provide a correct answer. The final answer MUST BE put in <answer> </answer> in json format.
+
+Example JSON format inside <answer>{"data": [['ans1', 'ans2'], ['ans3', 'ans4']]}</answer>.

EMNLP/2025/MTabVQA_ Evaluating Multi-Tabular Reasoning of Language Models in Visual Space/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f1fbe61b75e691ab57e1166f3b879df9560dc4a633d3951464d3c4f257abe67
+size 869118

EMNLP/2025/MTabVQA_ Evaluating Multi-Tabular Reasoning of Language Models in Visual Space/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9036d51b5b40742651b864224cccb858b488c1396c099f600b7764b1e7be351c
+size 679488

EMNLP/2025/MUG-Eval_ A Proxy Evaluation Framework for Multilingual Generation Capabilities in Any Language/622f7be9-7bee-45f1-ad70-fd4f26d326bd_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4c181eacbb1c02befa41b826c4eddf28bfbde42181b63572b627a8c4b867ab6
+size 169436

EMNLP/2025/MUG-Eval_ A Proxy Evaluation Framework for Multilingual Generation Capabilities in Any Language/622f7be9-7bee-45f1-ad70-fd4f26d326bd_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3537ed5e9a7a809ae7fe3a9cefeab1c9ee48f0cfcec0ced8e2a661fd4bcb5ece
+size 189390