Daily Papers

Recent advancements in large language models (LLMs) have revolutionized their ability to handle single-turn tasks, yet real-world applications demand sophisticated multi-turn interactions. This survey provides a comprehensive review of recent advancements in evaluating and enhancing multi-turn interactions in LLMs. Focusing on task-specific scenarios, from instruction following in diverse domains such as math and coding to complex conversational engagements in roleplay, healthcare, education, and even adversarial jailbreak settings, we systematically examine the challenges of maintaining context, coherence, fairness, and responsiveness over prolonged dialogues. The paper organizes current benchmarks and datasets into coherent categories that reflect the evolving landscape of multi-turn dialogue evaluation. In addition, we review a range of enhancement methodologies under multi-turn settings, including model-centric strategies (contextual learning, supervised fine-tuning, reinforcement learning, and new architectures), external integration approaches (memory-augmented, retrieval-based methods, and knowledge graph), and agent-based techniques for collaborative interactions. Finally, we discuss open challenges and propose future directions for research to further advance the robustness and effectiveness of multi-turn interactions in LLMs. Related resources and papers are available at https://github.com/yubol-cmu/Awesome-Multi-Turn-LLMs.

On-policy distillation (OPD) has shown strong potential for transferring reasoning ability from frontier or domain-specific models to smaller students. While effective on static single-turn tasks, its behavior in multi-turn agent settings remains underexplored. In this work, we identify a key limitation of vanilla OPD in such settings, which we term Trajectory-Level KL Instability. Specifically, we observe that KL divergence increases together with a drop in success rate, and even after convergence, the KL remains high, leading to unstable training. This instability arises from inter-turn error compounding: as errors accumulate, the student is driven beyond the teacher's effective support, rendering the supervision signal unreliable. To address this, we propose TCOD (Temporal Curriculum On-Policy Distillation), a simple yet effective framework that controls the trajectory depth exposed to the student and progressively expands it from short to long with a curriculum schedule.Experimental results across four student-teacher pairs on three multi-turn agent benchmarks (ALFWorld, WebShop, ScienceWorld) show that TCOD mitigates KL escalation and enhances KL stability throughout training, improving agent performance by up to 18 points over vanilla OPD. Further evaluations show that TCOD can even surpass the teacher's performance and generalize to tasks on which the teacher fails.

Large Language Models (LLMs) excel at single-turn tasks such as instruction following and summarization, yet real-world deployments require sustained multi-turn interactions where user goals and conversational context persist and evolve. A recurring challenge in this setting is context drift: the gradual divergence of a model's outputs from goal-consistent behavior across turns. Unlike single-turn errors, drift unfolds temporally and is poorly captured by static evaluation metrics. In this work, we present a study of context drift in multi-turn interactions and propose a simple dynamical framework to interpret its behavior. We formalize drift as the turn-wise KL divergence between the token-level predictive distributions of the test model and a goal-consistent reference model, and propose a recurrence model that interprets its evolution as a bounded stochastic process with restoring forces and controllable interventions. We instantiate this framework in both synthetic long-horizon rewriting tasks and realistic user-agent simulations such as in tau-Bench, measuring drift for several open-weight LLMs that are used as user simulators. Our experiments consistently reveal stable, noise-limited equilibria rather than runaway degradation, and demonstrate that simple reminder interventions reliably reduce divergence in line with theoretical predictions. Together, these results suggest that multi-turn drift can be understood as a controllable equilibrium phenomenon rather than as inevitable decay, providing a foundation for studying and mitigating context drift in extended interactions.

End-to-end (E2E) spoken dialogue systems are increasingly replacing cascaded pipelines for voice-based human-AI interaction, processing raw audio directly without intermediate transcription. Existing benchmarks primarily evaluate these models on synthetic speech and single-turn tasks, leaving realistic multi-turn conversational ability underexplored. We introduce Audio MultiChallenge, an open-source benchmark to evaluate E2E spoken dialogue systems under natural multi-turn interaction patterns. Building on the text-based MultiChallenge framework, which evaluates Inference Memory, Instruction Retention, and Self Coherence, we introduce a new axis Voice Editing that tests robustness to mid-utterance speech repairs and backtracking. We further augment each axis to the audio modality, such as introducing Audio-Cue challenges for Inference Memory that require recalling ambient sounds and paralinguistic signals beyond semantic content. We curate 452 conversations from 47 speakers with 1,712 instance-specific rubrics through a hybrid audio-native agentic and human-in-the-loop pipeline that exposes model failures at scale while preserving natural disfluencies found in unscripted human speech. Our evaluation of proprietary and open-source models reveals that even frontier models struggle on our benchmark, with Gemini 3 Pro Preview (Thinking), our highest-performing model achieving a 54.65% pass rate. Error analysis shows that models fail most often on our new axes and that Self Coherence degrades with longer audio context. These failures reflect difficulty of tracking edits, audio cues, and long-range context in natural spoken dialogue. Audio MultiChallenge provides a reproducible testbed to quantify them and drive improvements in audio-native multi-turn interaction capability.

While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and Llama-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.

Large Vision-Language Models (VLMs) exhibit impressive multi-modal capabilities but suffer from prohibitive computational and memory demands, due to their long visual token sequences and massive parameter sizes. To address these issues, recent works have proposed training-free compression methods. However, existing efforts often suffer from three major limitations: (1) Current approaches do not decompose techniques into comparable modules, hindering fair evaluation across spatial and temporal redundancy. (2) Evaluation confined to simple single-turn tasks, failing to reflect performance in realistic scenarios. (3) Isolated use of individual compression techniques, without exploring their joint potential. To overcome these gaps, we introduce LLMC+, a comprehensive VLM compression benchmark with a versatile, plug-and-play toolkit. LLMC+ supports over 20 algorithms across five representative VLM families and enables systematic study of token-level and model-level compression. Our benchmark reveals that: (1) Spatial and temporal redundancies demand distinct technical strategies. (2) Token reduction methods degrade significantly in multi-turn dialogue and detail-sensitive tasks. (3) Combining token and model compression achieves extreme compression with minimal performance loss. We believe LLMC+ will facilitate fair evaluation and inspire future research in efficient VLM. Our code is available at https://github.com/ModelTC/LightCompress.

We present PLSQLBench, to our knowledge the first benchmark for evaluating whether LLMs can write executable PL/SQL programs, with correctness measured through execution-based tests. Existing LLM evaluations largely target general-purpose code generation or declarative text-to-SQL, leaving procedural database programming underexplored. PLSQLBench contains 2,865 instances: 2,594 single-turn tasks and 271 multi-turn conversations spanning 978 turns. The benchmark combines complex schema-grounded tasks over enterprise-style Spider 2 databases, simpler schema-grounded tasks derived from Spider, and MBPP-derived procedural problems, covering varying levels of database grounding and procedural complexity. Experiments with eight LLMs reveal recurring difficulties in schema grounding, PL/SQL dialect fidelity, procedural control flow, exception handling, and cross-turn consistency. Tool-augmented LLM agents improve performance on several schema-grounded evaluations, although substantial gaps remain. These results highlight procedural database programming capabilities not directly assessed by conventional code generation or text-to-SQL benchmarks. Our code is available at https://github.com/oracle-samples/plsqlbench.

Error-targeted learning gives a legible way to improve a model: identify specific weaknesses of the model, train the model to fix these weaknesses, and check whether the weaknesses are fixed. We study how to apply error-targeted learning to terminal agents, which solve long-horizon tasks specified in natural language by interacting with a command shell. Existing error-targeted methods are designed for single-turn tasks, and existing analyses of agent failures are used to evaluate and debug agents rather than to train them. We propose Misbehavior-driven RL (MiRL), which first profiles the errors of a policy to find generalizable misbehaviors: kinds of wrong steps that recur across many environments. MiRL then trains the policy with RL, starting rollouts at the steps where these misbehaviors occur and rewarding a rollout only if it passes the tests and avoids the misbehavior. On Qwen3.5-4B, eight misbehaviors account for 52.1% of wrong steps. After training with MiRL, the policy repeats the measured misbehaviors 43% to 57% less often on average on unseen prefixes. MiRL outperforms TMax-4B on Terminal-Bench Lite (51.8% vs. 47.4%) and matches it on Terminal-Bench 2.1 (21.3% vs. 20.6%), while using 35% of its rollouts. These results show that error-targeted learning can be applied to realistic, long-horizon terminal tasks.

Sampling is a basic operation in many inference-time algorithms of large language models (LLMs). To scale up inference efficiently with a limited compute, it is crucial to find an optimal allocation for sample compute budgets: Which sampling configurations (model, temperature, language, etc.) do we use? How many samples do we generate in each configuration? We formulate these choices as a learning problem and propose OSCA, an algorithm that Optimizes Sample Compute Allocation by finding an optimal mix of different inference configurations. Our experiments show that with our learned mixed allocation, we can achieve accuracy better than the best single configuration with 128x less compute on code generation and 25x less compute on 4 reasoning tasks. OSCA is also shown to be effective in agentic workflows beyond single-turn tasks, achieving a better accuracy on SWE-Bench with 3x less compute than the default configuration. Our code and generations are released at https://github.com/LeiLiLab/OSCA.

Large Language Models (LLMs) have achieved remarkable performance on single-turn tasks, yet their effectiveness deteriorates in multi-turn conversations. We define this phenomenon as cumulative contextual decay - a progressive degradation of contextual integrity caused by attention pollution, dilution, and drift. To address this challenge, we propose Rhea (Role-aware Heuristic Episodic Attention), a novel framework that decouples conversation history into two functionally independent memory modules: (1) an Instructional Memory (IM) that persistently stores high-fidelity global constraints via a structural priority mechanism, and (2) an Episodic Memory (EM) that dynamically manages user-model interactions via asymmetric noise control and heuristic context retrieval. During inference, Rhea constructs a high signal-to-noise context by applying its priority attention: selectively integrating relevant episodic information while always prioritizing global instructions. To validate this approach, experiments on multiple multi-turn conversation benchmarks - including MT-Eval and Long-MT-Bench+ - show that Rhea mitigates performance decay and improves overall accuracy by 1.04 points on a 10-point scale (a 16% relative gain over strong baselines). Moreover, Rhea maintains near-perfect instruction fidelity (IAR > 8.1) across long-horizon interactions. These results demonstrate that Rhea provides a principled and effective framework for building more precise, instruction-consistent conversational LLMs.

LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.

Recent advances in group-based reinforcement learning (RL) have driven frontier large language models (LLMs) in single-turn tasks like mathematical reasoning. However, their scalability to long-horizon LLM agent training remains limited. Unlike static tasks, agent-environment interactions unfold over many steps and often yield sparse or delayed rewards, making credit assignment across individual steps significantly more challenging. In this work, we propose Group-in-Group Policy Optimization (GiGPO), a novel RL algorithm that achieves fine-grained credit assignment for LLM agents while preserving the appealing properties of group-based RL: critic-free, low memory, and stable convergence. GiGPO introduces a two-level structure for estimating relative advantage: (i) At the episode-level, GiGPO computes macro relative advantages based on groups of complete trajectories; (ii) At the step-level, GiGPO introduces an anchor state grouping mechanism that retroactively constructs step-level groups by identifying repeated environment states across trajectories. Actions stemming from the same state are grouped together, enabling micro relative advantage estimation. This hierarchical structure effectively captures both global trajectory quality and local step effectiveness without relying on auxiliary models or additional rollouts. We evaluate GiGPO on two challenging agent benchmarks, ALFWorld and WebShop, using Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct. Crucially, GiGPO delivers fine-grained per-step credit signals and achieves performance gains of > 12\% on ALFWorld and > 9\% on WebShop over the GRPO baseline: all while maintaining the same GPU memory overhead, identical LLM rollout, and incurring little to no additional time cost.

Large language models are increasingly evolving from text generators into general agents capable of understanding user requests, invoking external tools, and completing complex tasks through interaction. However, existing agent benchmarks often focus on limited scenarios, tool ecosystems, or interaction formats, making it difficult to systematically characterize model capabilities across heterogeneous application settings. We introduce OmniaBench, a benchmark for evaluating general agents across diverse scenarios with explicit state spaces. We derive application-oriented scenario knowledge from app stores, product documents, industry resources, Web retrieval, and human refinement, forming a hierarchical taxonomy that spans ToC, ToB and ToE with 90 level-1 and 354 level-2 domains. Based on this taxonomy, we construct executable environments and synthesize single-turn and multi-turn tasks through four complementary routes: DAG, DAG-S, Solver, and Program. OmniaBench further introduces a ten-dimensional capability taxonomy and eight compositional atomic difficulty factors to support fine-grained evaluation and analysis. The resulting dataset contains 1,431 tasks, together with a challenging subset of 644 tasks designed to reduce evaluation cost and mitigate potential contamination of the full set after public release. The bench presents substantial challenges to current frontier models, with even Claude-Sonnet-5 and GPT-5.6-Sol achieving Overall Pass@1 scores of only 58.54 and 57.14, respectively. Further analyses reveal clear differences across domains and capabilities, as well as persistent limitations in planning, constraint maintenance, and adaptive correction. OmniaBench provides a broad and diagnostic benchmark for characterizing the capability boundaries of general agents.

Recent advancements in reinforcement learning (RL) for large language models (LLMs), exemplified by DeepSeek R1, have shown that even a simple question-answering task can substantially improve an LLM's reasoning capabilities. In this work, we extend this approach by modifying the task into a multi-attempt setting. Instead of generating a single response per question, the model is given multiple attempts, with feedback provided after incorrect responses. The multi-attempt task encourages the model to refine its previous attempts and improve search efficiency. Experimental results show that even a small LLM trained on a multi-attempt task achieves significantly higher accuracy when evaluated with more attempts, improving from 45.6% with 1 attempt to 52.5% with 2 attempts on the math benchmark. In contrast, the same LLM trained on a standard single-turn task exhibits only a marginal improvement, increasing from 42.3% to 43.2% when given more attempts during evaluation. The results indicate that, compared to the standard single-turn task, an LLM trained on a multi-attempt task achieves slightly better performance on math benchmarks while also learning to refine its responses more effectively based on user feedback. Full code is available at https://github.com/DualityRL/multi-attempt

Large Language Models (LLMs) have demonstrated remarkable capabilities, enabling language agents to excel at single-turn tasks. However, their application to complex, multi-step, and long-horizon tasks remains challenging. While reinforcement learning (RL) offers a promising avenue for addressing these challenges, mainstream approaches typically rely solely on sparse, outcome-based rewards, a limitation that becomes especially problematic for group-based RL algorithms lacking critic models, such as Group Relative Policy Optimization (GRPO). In such methods, uniformly rewarding or penalizing all actions within a trajectory can lead to training instability and suboptimal policies, because beneficial and detrimental actions are often entangled across multi-step interactions. To address this challenge, we propose SALT, a novel and lightweight framework that provides a finer-grained advantage assignment, derived solely from outcome rewards. We achieve this by constructing a graph from trajectories of the same prompt, which allows us to quantify the quality of each step and assign advantages accordingly. Crucially, SALT is designed as a plug-and-play module that seamlessly integrates with existing group-based RL algorithms, requiring no modifications to the rollout procedure and introducing negligible computational overhead. Extensive experiments on the WebShop, ALFWorld, and AppWorld benchmarks with various model sizes demonstrate that SALT consistently improves performance. We also conduct a thorough analysis to validate the design choices behind SALT and offer actionable insights.

Unified Multimodal Models (UMMs) integrate visual understanding, generation, and editing within a single unified framework. However, existing unified multimodal datasets and benchmarks remain dominated by single-turn tasks. Even in recent multi-turn datasets, many turns can be solved independently without tracking information across turns. Consequently, they provide limited supervision and evaluation for multimodal stateful interaction, which requires models to understand, maintain, update, and faithfully realize evolving dialogue states across turns. To address this gap, we introduce YANAMI, the first large-scale multimodal dialogue dataset and benchmark for stateful interaction, comprising YANAMI-400K and YANAMI-Bench. YANAMI-400K contains 400K multi-turn multimodal dialogues, with more than 2M images and 2.6M dialogue turns. The dialogues span four representative scenarios: Multi-Image Fusion, Image Editing, World Modeling, and Visual Reasoning, and require models to continuously track, preserve, and revise evolving multimodal dialogue states across turns. YANAMI-Bench comprises 400 multi-turn dialogues for systematically evaluating multimodal stateful interaction. It jointly evaluates Dialogue State Understanding and Visual Realization, characterizing the gap between models' ability to understand evolving dialogue states and their ability to faithfully realize them in visual outputs. Experiments show that fine-tuning BAGEL on YANAMI-400K improves image editing, generation, and multimodal stateful interaction by 10.3%, 2.0%, and 13.23%, respectively, while preserving its general visual understanding capabilities. Evaluations on YANAMI-Bench further reveal that existing UMMs still struggle to maintain, update, and visually realize evolving multimodal dialogue states.

The widespread use of chest X-rays (CXRs), coupled with a shortage of radiologists, has driven growing interest in automated CXR analysis and AI-assisted reporting. While existing vision-language models (VLMs) show promise in specific tasks such as report generation or abnormality detection, they often lack support for interactive diagnostic capabilities. In this work we present RadVLM, a compact, multitask conversational foundation model designed for CXR interpretation. To this end, we curate a large-scale instruction dataset comprising over 1 million image-instruction pairs containing both single-turn tasks -- such as report generation, abnormality classification, and visual grounding -- and multi-turn, multi-task conversational interactions. After fine-tuning RadVLM on this instruction dataset, we evaluate it across different tasks along with re-implemented baseline VLMs. Our results show that RadVLM achieves state-of-the-art performance in conversational capabilities and visual grounding while remaining competitive in other radiology tasks. Ablation studies further highlight the benefit of joint training across multiple tasks, particularly for scenarios with limited annotated data. Together, these findings highlight the potential of RadVLM as a clinically relevant AI assistant, providing structured CXR interpretation and conversational capabilities to support more effective and accessible diagnostic workflows.

LLM agents have demonstrated remarkable capabilities in software development, but their performance is hampered by long interaction contexts, which incur high API costs and latency. While various context compression approaches such as LongLLMLingua have emerged to tackle this challenge, they typically rely on fixed metrics such as PPL, ignoring the task-specific nature of code understanding. As a result, they frequently disrupt syntactic and logical structure and fail to retain critical implementation details. In this paper, we propose SWE-Pruner, a self-adaptive context pruning framework tailored for coding agents. Drawing inspiration from how human programmers "selectively skim" source code during development and debugging, SWE-Pruner performs task-aware adaptive pruning for long contexts. Given the current task, the agent formulates an explicit goal (e.g., "focus on error handling") as a hint to guide the pruning targets. A lightweight neural skimmer (0.6B parameters) is trained to dynamically select relevant lines from the surrounding context given the goal. Evaluations across four benchmarks and multiple models validate SWE-Pruner's effectiveness in various scenarios, achieving 23-54% token reduction on agent tasks like SWE-Bench Verified and up to 14.84x compression on single-turn tasks like LongCodeQA with minimal performance impact.

As LLMs become more capable, they are increasingly deployed as collaborative agents, taking on user-delegated tasks through iterative interaction. Yet genuine interaction is inherently dynamic: users rarely specify their intent upfront, instead disclosing, revising, and reshaping it as the conversation unfolds. Despite this, LLMs are still predominantly evaluated or trained in single-turn, fully-specified settings, leaving open a fundamental question: how well do LLMs track and act on user intent as it evolves over the course of a conversation? To study this, we introduce a framework that transforms static, single-turn tasks into dynamic multi-turn conversations in which the user's intent evolves across turns--incrementally revealed, revised, and at times redirected mid-conversation--while preserving each task's original evaluation protocol, enabling existing benchmarks to be reused as controlled testbeds without new annotation. Across multiple tasks, we surface a consistent phenomenon: strong static-setting performance does not transfer to the evolving-intent setting, with substantial drops across model families. Our findings point to a fundamental gap: today's LLMs do not yet faithfully track and act on the user's evolving intent, a capability invisible to static evaluation yet critical for future collaborative agents.

Agent skills are often injected in full on every request, increasing token cost. We compare four content-preserving loading methods: Full, Skill Block, Reference, and Hybrid. Across SearchQA, SpreadsheetBench, ALFWorld, ScienceWorld, and SynthProc, we measure token usage using raw input for single-turn tasks and cache-correct effective input for multi-turn tasks. Results show no universal winner. Hybrid reduces input by 27.4% on SearchQA and 39.8% on SpreadsheetBench. On large multi-turn skills, Skill Block and Hybrid achieve substantial reductions, reaching 62.5% and 52.8% on ScienceWorld and 73.0% and 66.6% on SynthProc. ALFWorld shows smaller gains because procedures are short and repeatedly needed. Paired outcome tests detect no quality differences, though they do not establish equivalence. Overall, conditional loading is most beneficial when large portions of a skill are not needed on every turn.

Recent advancements in generative models have enabled high-fidelity text-to-image generation. However, open-source image-editing models still lag behind their proprietary counterparts, primarily due to limited high-quality data and insufficient benchmarks. To overcome these limitations, we introduce ImgEdit, a large-scale, high-quality image-editing dataset comprising 1.2 million carefully curated edit pairs, which contain both novel and complex single-turn edits, as well as challenging multi-turn tasks. To ensure the data quality, we employ a multi-stage pipeline that integrates a cutting-edge vision-language model, a detection model, a segmentation model, alongside task-specific in-painting procedures and strict post-processing. ImgEdit surpasses existing datasets in both task novelty and data quality. Using ImgEdit, we train ImgEdit-E1, an editing model using Vision Language Model to process the reference image and editing prompt, which outperforms existing open-source models on multiple tasks, highlighting the value of ImgEdit and model design. For comprehensive evaluation, we introduce ImgEdit-Bench, a benchmark designed to evaluate image editing performance in terms of instruction adherence, editing quality, and detail preservation. It includes a basic testsuite, a challenging single-turn suite, and a dedicated multi-turn suite. We evaluate both open-source and proprietary models, as well as ImgEdit-E1, providing deep analysis and actionable insights into the current behavior of image-editing models. The source data are publicly available on https://github.com/PKU-YuanGroup/ImgEdit.

Text-to-image generation tasks have driven remarkable advances in diverse media applications, yet most focus on single-turn scenarios and struggle with iterative, multi-turn creative tasks. Recent dialogue-based systems attempt to bridge this gap, but their single-agent, sequential paradigm often causes intention drift and incoherent edits. To address these limitations, we present Talk2Image, a novel multi-agent system for interactive image generation and editing in multi-turn dialogue scenarios. Our approach integrates three key components: intention parsing from dialogue history, task decomposition and collaborative execution across specialized agents, and feedback-driven refinement based on a multi-view evaluation mechanism. Talk2Image enables step-by-step alignment with user intention and consistent image editing. Experiments demonstrate that Talk2Image outperforms existing baselines in controllability, coherence, and user satisfaction across iterative image generation and editing tasks.

LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.

Reinforcement learning for long-horizon agents relies on purely retrospective training signals: credit is assigned only after observing environmental consequences, leaving the agent's belief at action time invisible to the gradient. We introduce Prospective Hindsight (PH), a self-calibrating training principle that augments any retrospective base method with a signal derived from the gap between the agent's prospective prediction (before feedback) and the retrospective evaluation (after feedback). This per-rollout surprise identifies samples where the agent's self-model is most inaccurate and amplifies their gradient contribution through a stop-gradient surprise-weighted advantage. Since the prospective predictor shares parameters with the policy, the two co-evolve, progressively shifting focus to the agent's remaining blind spots. We connect this principle to a privileged-information gap and show that minimizing the surprise residual provides a descent pathway on the agent's miscalibration rate; calibration thus emerges as a byproduct of optimization rather than from an added objective. On single-turn verifiable tasks and a multi-turn personal-agent task (under GRPO, on-policy distillation, and their combination), PH improves both task performance and calibration, with consistent gains across model scales. Notably, the dominant miscalibration mode shifts structurally between regimes, overconfident failures in single-turn, underconfident successes in multi-turn, yet the same training principle addresses both successfully.

LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.

Adapting one's thought process based on corrective feedback is an essential ability in human learning, particularly in collaborative settings. In contrast, the current large language model training paradigm relies heavily on modeling vast, static corpora. While effective for knowledge acquisition, it overlooks the interactive feedback loops essential for models to adapt dynamically to their context. In this work, we propose a framework that treats this interactive in-context learning ability not as an emergent property, but as a distinct, trainable skill. We introduce a scalable method that transforms single-turn verifiable tasks into multi-turn didactic interactions driven by information asymmetry. We first show that current flagship models struggle to integrate corrective feedback on hard reasoning tasks. We then demonstrate that models trained with our approach dramatically improve the ability to interactively learn from language feedback. More specifically, the multi-turn performance of a smaller model nearly reaches that of a model an order of magnitude larger. We also observe robust out-of-distribution generalization: interactive training on math problems transfers to diverse domains like coding, puzzles and maze navigation. Our qualitative analysis suggests that this improvement is due to an enhanced in-context plasticity. Finally, we show that this paradigm offers a unified path to self-improvement. By training the model to predict the teacher's critiques, effectively modeling the feedback environment, we convert this external signal into an internal capability, allowing the model to self-correct even without a teacher.

Conversational search, unlike single-turn retrieval tasks, requires understanding the current question within a dialogue context. The common approach of rewrite-then-retrieve aims to decontextualize questions to be self-sufficient for off-the-shelf retrievers, but most existing methods produce sub-optimal query rewrites due to the limited ability to incorporate signals from the retrieval results. To overcome this limitation, we present a novel framework RetPO (Retriever's Preference Optimization), which is designed to optimize a language model (LM) for reformulating search queries in line with the preferences of the target retrieval systems. The process begins by prompting a large LM to produce various potential rewrites and then collects retrieval performance for these rewrites as the retrievers' preferences. Through the process, we construct a large-scale dataset called RF collection, containing Retrievers' Feedback on over 410K query rewrites across 12K conversations. Furthermore, we fine-tune a smaller LM using this dataset to align it with the retrievers' preferences as feedback. The resulting model achieves state-of-the-art performance on two recent conversational search benchmarks, significantly outperforming existing baselines, including GPT-3.5.

Length scaling during reinforcement-learning (RL) post-training is often viewed as a sign of improved reasoning ability, especially on difficult problems, but may also make responses to already-solved problems unnecessarily verbose. We quantify this side effect as the length-scaling tax (LST): excess response length on already-solved queries without a commensurate accuracy gain. To mitigate LST, we propose Length Self-Distillation (LSD), which routes solved prompts to on-policy distillation and retains the original RL objective for unsolved prompts. LSD uses an exponential moving average of the online policy as its teacher, requiring no external model. We find that LSD achieves comparable or better performance than RL across multiple variants, while substantially curbing response-length growth on easy queries. LSD reduces LST from 19.0% to -3.7% on single-turn reasoning and from 31.4% to 13.7% on multi-turn agentic tasks, demonstrating that LSD effectively preserves concise response patterns on easy queries while supporting efficient exploration on difficult queries during RL post-training.

We introduce SWE-Interact, a new testbed for evaluating coding agents on multi-turn, interactive, user-driven software engineering tasks. Existing frontier SWE benchmarks typically provide complete requirements upfront and evaluate agents on autonomous implementation. In contrast, SWE-Interact places agents in a realistic developer workflow: a carefully designed user simulator starts with vague or incomplete instructions, progressively reveals requirements, inspects the agent's workspace, and provides targeted feedback, revisions, and new constraints until the full task goal has been handed off. Grounded in large-scale studies of real coding-agent interactions, this setup tests whether agents can discover user intent, adapt to evolving requirements, and build on their own prior work. Across a suite of frontier and open-weight models, we find that strong performance on single-turn SWE tasks does not reliably transfer to multi-turn, user-driven workflows: the best-performing models solve roughly 50% of single-turn baseline tasks but only 25% of the corresponding SWE-Interact tasks. The strongest models in our evaluation, including Opus 4.8 and GPT 5.5, start strong even in the face of vague initial instructions, persevere until all the requirements are surfaced by the user, integrate them better and write clean code. However, they still suffer from over-agentic coding, forgetting requirements and technical mistakes. Weaker models start poorly under ambiguity, give up early, forget or ignore instructions and rework their code more. Overall, SWE-Interact measures an orthogonal, real-world capability axis for frontier model development: interactive goal discovery and iterative refinement with a user in the loop.

Large-scale reinforcement learning with verifiable rewards (RLVR) has demonstrated its effectiveness in harnessing the potential of large language models (LLMs) for single-turn reasoning tasks. In realistic reasoning scenarios, LLMs can often utilize external tools to assist in task-solving processes. However, current RL algorithms inadequately balance the models' intrinsic long-horizon reasoning capabilities and their proficiency in multi-turn tool interactions. To bridge this gap, we propose Agentic Reinforced Policy Optimization (ARPO), a novel agentic RL algorithm tailored for training multi-turn LLM-based agents. Through preliminary experiments, we observe that LLMs tend to exhibit highly uncertain behavior, characterized by an increase in the entropy distribution of generated tokens, immediately following interactions with external tools. Motivated by this observation, ARPO incorporates an entropy-based adaptive rollout mechanism, dynamically balancing global trajectory sampling and step-level sampling, thereby promoting exploration at steps with high uncertainty after tool usage. By integrating an advantage attribution estimation, ARPO enables LLMs to internalize advantage differences in stepwise tool-use interactions. Our experiments across 13 challenging benchmarks in computational reasoning, knowledge reasoning, and deep search domains demonstrate ARPO's superiority over trajectory-level RL algorithms. Remarkably, ARPO achieves improved performance using only half of the tool-use budget required by existing methods, offering a scalable solution for aligning LLM-based agents with real-time dynamic environments. Our code and datasets are released at https://github.com/dongguanting/ARPO

LLM-based agents score well on search benchmarks, yet real users consistently find results unsatisfying, revealing a persistent evaluation-experience gap. We attribute this gap to existing benchmarks' reliance on over-specified queries, single-turn interactions, and fixed-schema evaluation, none of which reflect real search behavior where users and agents collaboratively refine vague intent through multi-turn dialogue. We term this paradigm VibeSearch and introduce VibeSearchBench, a benchmark comprising 200 manually curated bilingual (Chinese and English) tasks across 20 domains, split into VibeSearch-Pro (professional) and VibeSearch-Daily (daily-life) subsets. Each task pairs a user persona with a schema-free ground-truth knowledge graph, and is evaluated through a progressive-disclosure user simulator and a graph-matching evaluation framework. We benchmark seven frontier models under both the ReAct framework and the OpenClaw agent harness. Results show that all models remain substantially inadequate for VibeSearch (best F1: 30.30), highlighting the need for fundamental advances in long-context reasoning, proactive intent elicitation, and structured knowledge construction.

Video understanding is a fundamental task for evaluating the capabilities of multimodal large language models (MLLMs). However, existing leading models have already achieved approximately 90% accuracy on the Video-MME leaderboard, suggesting that conventional single-turn video understanding tasks are becoming increasingly saturated and insufficient for assessing the intelligence of advanced MLLMs. Towards this end, we introduce VideoGAIA, an agentic video understanding benchmark for general artificial intelligence (AI) assistants. Moving beyond one-shot video question answering, VideoGAIA formulates video understanding as a multi-turn, tool-augmented interaction process, where models must iteratively perceive videos, invoke external tools, gather complementary information, and integrate multimodal evidence across turns. VideoGAIA contains 271 model-human co-designed tasks covering diverse and complex real-world scenarios. Each video-question-answer instance is independently verified by three human experts to ensure both correctness and appropriate difficulty. All evaluated MLLMs, including frontier models such as GPT-5.5 and Kimi-K3, achieve less than 60% accuracy on VideoGAIA, highlighting its value as a high-quality and timely benchmark for evaluating next-generation MLLMs. We hope that VideoGAIA will facilitate the transition from conventional video understanding toward agentic video understanding.

We present a single-turn agent for graphical user interface (GUI) interaction tasks, using Vision-Language Model Florence-2-Base. The agent's primary task is identifying the screen coordinates of the UI element corresponding to the user's command. It demonstrates strong performance on Screenspot and OmniAct, while maintaining a compact size of 0.27B parameters and minimal latency. Relevant improvement comes from multi-task training and MLLM-based data augmentation. Manually annotated corpora are scarce, but we show that MLLM augmentation might produce better results. On Screenspot and OmniAct, our model outperforms both GUI-specific models (e.g., SeeClick) and MLLMs (e.g., GPT-4V).

Large Language Models (LLMs) have demonstrated remarkable capabilities in knowledge acquisition, reasoning, and tool use, making them promising candidates for autonomous agent applications. However, training LLM agents for complex multi-turn task planning faces significant challenges, including sparse episode-wise rewards, credit assignment across long horizons, and the computational overhead of reinforcement learning in multi-turn interaction settings. To this end, this paper introduces a novel approach that transforms multi-turn task planning into single-turn task reasoning problems, enabling efficient policy optimization through Group Relative Policy Optimization (GRPO) with dense and verifiable reward from expert trajectories. Our theoretical analysis shows that GRPO improvement on single-turn task reasoning results in a lower bound of the multi-turn success probability under the minimal turns, as well as the generalization to subtasks with shorter horizons. Experimental evaluation on the complex task planning benchmark demonstrates that our 1.5B parameter model trained with single-turn GRPO achieves superior performance compared to larger baseline models up to 14B parameters, with success rates of 70% for long-horizon planning tasks.

Large Language Models (LLMs) have achieved remarkable success at tasks like summarization that involve a single turn of interaction. However, they can still struggle with multi-turn tasks like dialogue that require long-term planning. Previous works on multi-turn dialogue extend single-turn reinforcement learning from human feedback (RLHF) methods to the multi-turn setting by treating all prior dialogue turns as a long context. Such approaches suffer from covariate shift: the conversations in the training set have previous turns generated by some reference policy, which means that low training error may not necessarily correspond to good performance when the learner is actually in the conversation loop. In response, we introduce REgressing the RELative FUture (REFUEL), an efficient policy optimization approach designed to address multi-turn RLHF in LLMs. REFUEL employs a single model to estimate Q-values and trains on self-generated data, addressing the covariate shift issue. REFUEL frames the multi-turn RLHF problem as a sequence of regression tasks on iteratively collected datasets, enabling ease of implementation. Theoretically, we prove that REFUEL can match the performance of any policy covered by the training set. Empirically, we evaluate our algorithm by using Llama-3.1-70B-it to simulate a user in conversation with our model. REFUEL consistently outperforms state-of-the-art methods such as DPO and REBEL across various settings. Furthermore, despite having only 8 billion parameters, Llama-3-8B-it fine-tuned with REFUEL outperforms Llama-3.1-70B-it on long multi-turn dialogues. Implementation of REFUEL can be found at https://github.com/ZhaolinGao/REFUEL/, and models trained by REFUEL can be found at https://huggingface.co/Cornell-AGI.

Scientific and engineering progress is fundamentally a long-horizon iterative process: proposing changes, running experiments, measuring outcomes, and continuously refining artifacts. Yet existing benchmarks for frontier models primarily evaluate either single-turn responses or short-horizon agent trajectories, failing to capture the challenges of sustained iterative improvement over extended time horizons. To address this gap, we introduce AutoLab, a new benchmark for ultra long-horizon closed-loop optimization. AutoLab consists of 36 realistic, expert-curated tasks spanning four diverse domains: system optimization, puzzle & challenge, model development, and CUDA kernel optimization. Each task begins with a correct but deliberately suboptimal baseline and challenges agents to improve it within a strict wall-clock budget. Evaluating 17 state-of-the-art models reveals the dominant predictor of success is not the quality of an agent's initial attempt, but its persistence in repeatedly benchmarking, editing, and incorporating empirical feedback. While claude-opus-4.6 exhibits strong long-horizon optimization capabilities, most frontier models, including several proprietary ones, either terminate prematurely or exhaust their budgets with minimal progress. These results underscore the importance of time awareness and persistent iteration in autonomous agents. We open-source the full benchmark, evaluation harness, and task artifacts, to accelerate research toward truly capable long-horizon agents.

Recent studies have shown that large language models' (LLMs) mathematical problem-solving capabilities can be enhanced by integrating external tools, such as code interpreters, and employing multi-turn Chain-of-Thought (CoT) reasoning. While current methods focus on synthetic data generation and Supervised Fine-Tuning (SFT), this paper studies the complementary direct preference learning approach to further improve model performance. However, existing direct preference learning algorithms are originally designed for the single-turn chat task, and do not fully address the complexities of multi-turn reasoning and external tool integration required for tool-integrated mathematical reasoning tasks. To fill in this gap, we introduce a multi-turn direct preference learning framework, tailored for this context, that leverages feedback from code interpreters and optimizes trajectory-level preferences. This framework includes multi-turn DPO and multi-turn KTO as specific implementations. The effectiveness of our framework is validated through training of various language models using an augmented prompt set from the GSM8K and MATH datasets. Our results demonstrate substantial improvements: a supervised fine-tuned Gemma-1.1-it-7B model's performance increased from 77.5% to 83.9% on GSM8K and from 46.1% to 51.2% on MATH. Similarly, a Gemma-2-it-9B model improved from 84.1% to 86.3% on GSM8K and from 51.0% to 54.5% on MATH.

A broad use case of large language models (LLMs) is in goal-directed decision-making tasks (or "agent" tasks), where an LLM needs to not just generate completions for a given prompt, but rather make intelligent decisions over a multi-turn interaction to accomplish a task (e.g., when interacting with the web, using tools, or providing customer support). Reinforcement learning (RL) provides a general paradigm to address such agent tasks, but current RL methods for LLMs largely focus on optimizing single-turn rewards. By construction, most single-turn RL methods cannot endow LLMs with the ability to intelligently seek information over multiple turns, perform credit assignment, or reason about their past actions -- all of which are critical in agent tasks. This raises the question: how can we design effective and efficient multi-turn RL algorithms for LLMs? In this paper, we develop a framework for building multi-turn RL algorithms for fine-tuning LLMs, that preserves the flexibility of existing single-turn RL methods for LLMs (e.g., proximal policy optimization), while accommodating multiple turns, long horizons, and delayed rewards effectively. To do this, our framework adopts a hierarchical RL approach and runs two RL algorithms in parallel: a high-level off-policy value-based RL algorithm to aggregate reward over utterances, and a low-level RL algorithm that utilizes this high-level value function to train a token policy within each utterance or turn. Our hierarchical framework, Actor-Critic Framework with a Hierarchical Structure (ArCHer), can also give rise to other RL methods. Empirically, we find that ArCHer significantly improves efficiency and performance on agent tasks, attaining a sample efficiency of about 100x over existing methods, while also improving with larger model capacity (upto the 7 billion scale that we tested on).

Large language models (LLMs) produce responses rated as highly empathic in single-turn settings (Ayers et al., 2023; Lee et al., 2024), yet they are also known to be formulaic generators that reuse the same lexical patterns, syntactic templates, and discourse structures across tasks (Jiang et al., 2025; Shaib et al., 2024; Namuduri et al., 2025). Less attention has been paid to whether this formulaicity extends to the level of discourse moves, i.e., what a response does for the person it is addressing. This question is especially consequential for empathic dialogue, where effective support demands not just a kind response at one moment but varied strategies as a conversation unfolds (Stiles et al., 1998). Indeed, prior work shows that LLMs reuse the same tactic sequences more than human supporters in single-turn settings (Gueorguieva et al., 2026). We extend this analysis to multi-turn conversations and find that the rigidity compounds: once a tactic appears in a supporter turn, LLMs reuse it in the next at nearly double the rate of humans (0.50-0.56 vs. 0.27). This pattern holds across LLMs serving as supporters in real emotional support conversations, and is invisible to standard similarity metrics. To address this gap, we introduce MINT (Multi-turn Inter-tactic Novelty Training), the first reinforcement learning framework to optimize discourse move diversity across multi-turn empathic dialogue. The best MINT variant combines an empathy quality reward with a cross-turn tactic novelty signal, improving aggregate empathy by 25.3% over vanilla across 1.7B and 4B models while reducing cross-turn discourse move repetition by 26.3% on the 4B model, surpassing all baselines including quality-only and token-level diversity methods on both measures. These results suggest that what current models lack is not empathy itself, but the ability to vary their discourse moves across a conversation.

Large Language Models (LLMs) are conversational interfaces. As such, LLMs have the potential to assist their users not only when they can fully specify the task at hand, but also to help them define, explore, and refine what they need through multi-turn conversational exchange. Although analysis of LLM conversation logs has confirmed that underspecification occurs frequently in user instructions, LLM evaluation has predominantly focused on the single-turn, fully-specified instruction setting. In this work, we perform large-scale simulation experiments to compare LLM performance in single- and multi-turn settings. Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks. Analysis of 200,000+ simulated conversations decomposes the performance degradation into two components: a minor loss in aptitude and a significant increase in unreliability. We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that *when LLMs take a wrong turn in a conversation, they get lost and do not recover*.

Multimodal large language models (MLLMs), built on large-scale pre-trained vision towers and language models, have shown great capabilities in multimodal understanding. However, most existing MLLMs are trained on single-turn vision question-answering tasks, which do not accurately reflect real-world human conversations. In this paper, we introduce MMDiag, a multi-turn multimodal dialogue dataset. This dataset is collaboratively generated through deliberately designed rules and GPT assistance, featuring strong correlations between questions, between questions and images, and among different image regions; thus aligning more closely with real-world scenarios. MMDiag serves as a strong benchmark for multi-turn multimodal dialogue learning and brings more challenges to the grounding and reasoning capabilities of MLLMs. Further, inspired by human vision processing, we present DiagNote, an MLLM equipped with multimodal grounding and reasoning capabilities. DiagNote consists of two modules (Deliberate and Gaze) interacting with each other to perform Chain-of-Thought and annotations respectively, throughout multi-turn dialogues. We empirically demonstrate the advantages of DiagNote in both grounding and jointly processing and reasoning with vision and language information over existing MLLMs.

Deep reasoning is fundamental for solving complex tasks, especially in vision-centric scenarios that demand sequential, multimodal understanding. However, existing benchmarks typically evaluate agents with fully synthetic, single-turn queries, limited visual modalities, and lack a framework to assess reasoning quality over multiple steps as required in real-world settings. To address this, we introduce Agent-X, a large-scale benchmark for evaluating vision-centric agents multi-step and deep reasoning capabilities in real-world, multimodal settings. Agent- X features 828 agentic tasks with authentic visual contexts, including images, multi-image comparisons, videos, and instructional text. These tasks span six major agentic environments: general visual reasoning, web browsing, security and surveillance, autonomous driving, sports, and math reasoning. Our benchmark requires agents to integrate tool use with explicit, stepwise decision-making in these diverse settings. In addition, we propose a fine-grained, step-level evaluation framework that assesses the correctness and logical coherence of each reasoning step and the effectiveness of tool usage throughout the task. Our results reveal that even the best-performing models, including GPT, Gemini, and Qwen families, struggle to solve multi-step vision tasks, achieving less than 50% full-chain success. These findings highlight key bottlenecks in current LMM reasoning and tool-use capabilities and identify future research directions in vision-centric agentic reasoning models. Our data and code are publicly available at https://github.com/mbzuai-oryx/Agent-X

Developing robust world model reasoning is crucial for large language model (LLM) agents to plan and interact in complex environments. While multi-turn interaction offers a superior understanding of environmental dynamics via authentic feedback, current approaches often impose a rigid reasoning process, which constrains the model's active learning, ultimately hindering efficient world model reasoning. To address these issues, we explore world-model internalization through efficient interaction and active reasoning (WMAct), which liberates the model from structured reasoning, allowing the model to shape thinking directly through its doing, and achieves effective and efficient world model reasoning with two key mechanisms: (1) a reward rescaling mechanism adjusting outcome reward based on action efficacy to incentivize redundancy reduction and purposeful interaction; (2) an interaction frequency annealing strategy to progressively reduce the maximum allowed interaction turns, which compels the model to condense its learning and internalize environmental dynamics rather than over-relying on environmental cues. Our experiments on Sokoban, Maze, and Taxi show that WMAct yields effective world model reasoning capable of resolving tasks in a single turn that previously required multiple interactions and fosters strong transferability to complex environments, improving performance on a suite of reasoning benchmarks.

LLM agents have incredible potential for scientific discovery applications. However, the performance of LLM agents on real-world, small molecule drug design (SMDD) tasks across diverse chemistries and targets is unclear. Current evaluation methods are either ad hoc, too simple for real-world discovery, limited in scale, or restricted to single-turn question answering. In effort to standardize the evaluation of LLM agents on small molecule design, we introduce SMDD-Bench, a challenging, multi-turn, long-horizon agentic benchmark consisting of 502 guaranteed-solvable task instances spanning 5 task types: 2D Pharmacophore Identification, Interaction Point Discovery, Scaffold Hopping, Lead Optimization, and Fragment Assembly. SMDD-Bench tasks span a wide region of chemical space and involve 102 unique protein targets. Completely solving the benchmark would require having strong chemical and biological reasoning and 3D intuition, understanding specialized tool use, and displaying planning expertise over a limited number of oracle calls. We benchmark 7 frontier open and closed source LLMs and find even the most performant LLM, GPT5.4, solves only 40.2\% of tasks. We hope SMDD-Bench provides a standardized testbed to invigorate the field towards training and evaluating LLM agents for fully autonomous computational drug design. We host a public leaderboard at smddbench.com .

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists' First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

The rapid development of large language models and multimodal large language models has accelerated the emergence of proactive agents capable of operating everyday tools and assisting users in real-world environments. However, existing benchmarks struggle to evaluate such agents effectively, as they often rely on sandboxed environments and single-turn evaluation paradigms. Moreover, their scenario-based task taxonomies mix multiple model capabilities within the same task category, making it difficult to identify the root causes of agent failures. To address these limitations, we introduce UniClawBench, the first capability-driven benchmark designed to evaluate proactive agents in dynamic, real-world settings. UniClawBench is built around five foundational model capabilities: Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, and Cross-Platform Coordination. Based on these capabilities, we design 400 bilingual real-world tasks. Unlike previous benchmarks that rely on static, pre-recorded answers, our benchmark evaluates agents in live Docker containers using fine-grained, step-by-step completion checkpoints. Furthermore, we design a closed-loop evaluation strategy comprising an executor agent, a hidden supervisor agent, and a user agent to simulate realistic multi-turn human feedback without leaking grading criteria. To disentangle base model capabilities from framework-level design choices, we evaluate state-of-the-art models under multiple agent frameworks. Through comprehensive comparisons across both models and frameworks, we show how base model capabilities and agent framework designs jointly shape performance in real-world environments. To facilitate future research, we make our benchmark and code publicly available at https://github.com/HKU-MMLab/UniClawBench.

Off-policy problems such as policy staleness and training-inference mismatch, has become a major bottleneck for training stability and further exploration for LLM RL. To enhance inference efficiency, the distribution gap between the inference and updated policy grows, leading to heavy-tailed importance ratios. Heavy-tailed ratios arise when the policy is locally sharp, which further inflates sharp gradients and can push updates outside the trust region. To address this, we propose Adaptive Layerwise Perturbation(ALP) by injecting small learnable perturbations into input hidden states of each layer during updates, which is used as the numerator of the importance ratio against the unchanged inference policy in the objective. Intuitively, by adding controlled noise to intermediate representations, ALP prevents the updated policy from deviating too sharply from the inference policy, and enlarges the policy family to cover the inference policy family with mismatch noises. Hence, the flattened distribution can naturally tighten the updated and inference policy gap and reduce the tail of importance ratios, thus maintaining training stability. This is further validated empirically. Experiments on single-turn math and multi-turn tool-integrated reasoning tasks show that ALP not only improves final performance, but also avoid blow up of importance ratio tail and KL spikes during iterative training, along with boosted exploration. Ablations show that representation-level perturbations across all layers are most effective, substantially outperforming partial-layer and logits-only variants.

Large language models (LLMs) have demonstrated remarkable prowess in language understanding and generation. Advancing from foundation LLMs to instructionfollowing LLMs, instruction tuning plays a vital role in aligning LLMs to human preferences. However, the existing LLMs are usually focused on English, leading to inferior performance in non-English languages. In order to improve the performance for non-English languages, it is necessary to collect language-specific training data for foundation LLMs and construct language-specific instructions for instruction tuning, both of which are heavy loads. To minimize human workload, we propose to transfer the capabilities of language generation and instruction following from English to other languages through an interactive translation task. We have developed BayLing, an instruction-following LLM by utilizing LLaMA as the foundation LLM and automatically constructing interactive translation instructions for instructing tuning. Extensive assessments demonstrate that BayLing achieves comparable performance to GPT-3.5-turbo, despite utilizing a considerably smaller parameter size of only 13 billion. Experimental results on translation tasks show that BayLing achieves 95% of single-turn translation capability compared to GPT-4 with automatic evaluation and 96% of interactive translation capability compared to GPT-3.5-turbo with human evaluation. To estimate the performance on general tasks, we created a multi-turn instruction test set called BayLing-80. The experimental results on BayLing-80 indicate that BayLing achieves 89% of performance compared to GPT-3.5-turbo. BayLing also demonstrates outstanding performance on knowledge assessment of Chinese GaoKao and English SAT, second only to GPT-3.5-turbo among a multitude of instruction-following LLMs. Demo, homepage, code and models of BayLing are available.

Embodied agents empowered by large language models (LLMs) have shown strong performance in household object rearrangement tasks. However, these tasks primarily focus on single-turn interactions with simplified instructions, which do not truly reflect the challenges of providing meaningful assistance to users. To provide personalized assistance, embodied agents must understand the unique semantics that users assign to the physical world (e.g., favorite cup, breakfast routine) by leveraging prior interaction history to interpret dynamic, real-world instructions. Yet, the effectiveness of embodied agents in utilizing memory for personalized assistance remains largely underexplored. To address this gap, we present MEMENTO, a personalized embodied agent evaluation framework designed to comprehensively assess memory utilization capabilities to provide personalized assistance. Our framework consists of a two-stage memory evaluation process design that enables quantifying the impact of memory utilization on task performance. This process enables the evaluation of agents' understanding of personalized knowledge in object rearrangement tasks by focusing on its role in goal interpretation: (1) the ability to identify target objects based on personal meaning (object semantics), and (2) the ability to infer object-location configurations from consistent user patterns, such as routines (user patterns). Our experiments across various LLMs reveal significant limitations in memory utilization, with even frontier models like GPT-4o experiencing a 30.5% performance drop when required to reference multiple memories, particularly in tasks involving user patterns. These findings, along with our detailed analyses and case studies, provide valuable insights for future research in developing more effective personalized embodied agents. Project website: https://connoriginal.github.io/MEMENTO

Large language models (LLMs) have demonstrated remarkable performance on single-turn text-to-SQL tasks, but real-world database applications predominantly require multi-turn interactions to handle ambiguous queries, execution errors, and evolving user requirements. Existing multi-turn benchmarks fall short by treating conversation histories as static context or limiting evaluation to read-only operations, failing to reflect production-grade database assistant challenges. We introduce BIRD-INTERACT, a benchmark that restores this realism through: (1) a comprehensive interaction environment coupling each database with a hierarchical knowledge base, metadata files, and a function-driven user simulator, enabling models to solicit clarifications, retrieve knowledge, and recover from errors without human supervision; (2) two evaluation settings consisting of a pre-defined conversational protocol (c-Interact) and an open-ended agentic setting (a-Interact) where models autonomously decide when to query the user simulator or explore the environment; (3) a challenging task suite covering the full CRUD spectrum for business-intelligence and operational use cases, guarded by executable test cases. Each task features ambiguous and follow-up sub-tasks requiring dynamic interaction. The suite comprises BIRD-INTERACT-FULL (600 tasks, up to 11,796 interactions) for comprehensive performance assessment, and BIRD-INTERACT-LITE (300 tasks with simplified databases) for detailed behavioral analysis and rapid method development. Our empirical results highlight BIRD-INTERACT's difficulty: GPT-5 completes only 8.67% of tasks in c-Interact and 17.00% in a-Interact. Analysis via memory grafting and Interaction Test-time Scaling validates the importance of effective interaction for complex, dynamic text-to-SQL tasks.

Designing tool-integrated reasoning agent systems for complex chemical tasks remains a fundamental challenge. Although recent chemical agents have demonstrated promising performance through the integration of external tools, they are still built on the autoregressive paradigm, which is inherently constrained by causal attention and left-to-right generation. As a result, they often struggle with bidirectional molecular understanding, global coordinated tool planning and multi-source evidence integration. To address this challenge, we introduce ChemDiffAgent, the first masked diffusion language agent for tool-integrated chemical reasoning. It reformulates chemical reasoning as iterative denoising over interleaved reasoning and tool-use trajectories, enabling tool-use decisions to be jointly determined under full-trajectory context and thereby better aligning the agent's decision boundary with its knowledge boundary. We further analyze how this paradigm benefits agentic reasoning through harder infilling subproblems during training and flexible arbitrary-order decoding during inference. Then we develop a two-stage post-training framework, comprising agentic supervised fine-tuning and variance-reduced preference optimization, to enhance tool-use and reasoning capabilities. We also construct a comprehensive benchmark covering six chemistry tasks across both single-turn and multi-turn tool-use scenarios. Under matched budgets and evaluation, ChemDiffAgent consistently outperforms autoregressive agents, achieving an overall improvement of approximately 20%. Notably, our 8B model achieves performance comparable to, in some cases surpassing, significantly larger frontier models, such as GPT 5.5 and Claude Opus 4.7. Code is available at https://github.com/bilemond/ChemDiffAgent.

We present KGConv, a large, conversational corpus of 71k conversations where each question-answer pair is grounded in a Wikidata fact. Conversations contain on average 8.6 questions and for each Wikidata fact, we provide multiple variants (12 on average) of the corresponding question using templates, human annotations, hand-crafted rules and a question rewriting neural model. We provide baselines for the task of Knowledge-Based, Conversational Question Generation. KGConv can further be used for other generation and analysis tasks such as single-turn question generation from Wikidata triples, question rewriting, question answering from conversation or from knowledge graphs and quiz generation.

The proliferation of LLM-based agents has led to increasing deployment of inter-agent collaboration for tasks like scheduling, negotiation, resource allocation etc. In such systems, privacy is critical, as agents often access proprietary tools and domain-specific databases requiring strict confidentiality. This paper examines whether LLM-based agents demonstrate an understanding of contextual privacy. And, if instructed, do these systems preserve inference time user privacy in non-adversarial multi-turn conversation. Existing benchmarks to evaluate contextual privacy in LLM-agents primarily assess single-turn, low-complexity tasks where private information can be easily excluded. We first present a benchmark - MAGPIE comprising 158 real-life high-stakes scenarios across 15 domains. These scenarios are designed such that complete exclusion of private data impedes task completion yet unrestricted information sharing could lead to substantial losses. We then evaluate the current state-of-the-art LLMs on (a) their understanding of contextually private data and (b) their ability to collaborate without violating user privacy. Empirical experiments demonstrate that current models, including GPT-4o and Claude-2.7-Sonnet, lack robust understanding of contextual privacy, misclassifying private data as shareable 25.2\% and 43.6\% of the time. In multi-turn conversations, these models disclose private information in 59.9\% and 50.5\% of cases even under explicit privacy instructions. Furthermore, multi-agent systems fail to complete tasks in 71\% of scenarios. These results underscore that current models are not aligned towards both contextual privacy preservation and collaborative task-solving.

Reliable multimodal interaction depends on selective use of conversational history: an earlier question may remain relevant while its previous answer is outdated, whereas a current request may depend on historical evidence despite conflicting new observations. Existing multi-turn evaluations rarely separate these history-use demands from underlying question difficulty. To address this gap, we introduce ReTurn, a benchmark of 7,000 base tasks spanning visual and audio evidence for evaluating selective history use. For task-carrying history, Reconfirm/Reground require applying a historical question to current media while varying historical agreement; for evidence-carrying history, Retrieve/Rebind require answering a current question using historical media while varying current-media competition. Each pair preserves the target question, media, and answer. Tasks support open-ended and multiple-choice evaluation, with matched single-turn counterparts serving as answerability references. Across 13 omni-modal, vision-language, and audio-language models, median model-level open-ended accuracy falls from 93.7% with direct input to 72.3% in conversation. Behavioral probes show that high question recall can coexist with weaker task application, while competing media can redirect answers away from historical targets. Supervised adaptation yields only partial gains. ReTurn provides a controlled framework for assessing whether multimodal models select and use the historical information required by each request.

While NLP research has made strides in conversational tasks, many approaches focus on single-turn responses with well-defined objectives or evaluation criteria. In contrast, coaching presents unique challenges with initially undefined goals that evolve through multi-turn interactions, subjective evaluation criteria, mixed-initiative dialogue. In this work, we describe and implement five multi-turn coaching agents that exhibit distinct conversational styles, and evaluate them through a user study, collecting first-person feedback on 155 conversations. We find that users highly value core functionality, and that stylistic components in absence of core components are viewed negatively. By comparing user feedback with third-person evaluations from health experts and an LM, we reveal significant misalignment across evaluation approaches. Our findings provide insights into design and evaluation of conversational coaching agents and contribute toward improving human-centered NLP applications.

While Instruction-based Image Editing (IIE) has achieved significant progress, existing benchmarks pursue task breadth via mixed evaluations. This paradigm obscures a critical failure mode crucial in professional applications: the inconsistent performance of models across tasks of varying semantic scales. To address this gap, we introduce Omni IIE Bench, a high-quality, human-annotated benchmark specifically designed to diagnose the editing consistency of IIE models in practical application scenarios. Omni IIE Bench features an innovative dual-track diagnostic design: (1) Single-turn Consistency, comprising shared-context task pairs of attribute modification and entity replacement; and (2) Multi-turn Coordination, involving continuous dialogue tasks that traverse semantic scales. The benchmark is constructed via an exceptionally rigorous multi-stage human filtering process, incorporating a quality standard enforced by computer vision graduate students and an industry relevance review conducted by professional designers. We perform a comprehensive evaluation of 8 mainstream IIE models using Omni IIE Bench. Our analysis quantifies, for the first time, a prevalent performance gap: nearly all models exhibit a significant performance degradation when transitioning from low-semantic-scale to high-semantic-scale tasks. Omni IIE Bench provides critical diagnostic tools and insights for the development of next-generation, more reliable, and stable IIE models.

Large language models (LLMs) have shown the potential to be integrated into human daily lives. Therefore, user preference is the most critical criterion for assessing LLMs' performance in real-world scenarios. However, existing benchmarks mainly focus on measuring models' accuracy using multi-choice questions, which limits the understanding of their capabilities in real applications. We fill this gap by proposing a comprehensive Chinese benchmark SuperCLUE, named after another popular Chinese LLM benchmark CLUE. SuperCLUE encompasses three sub-tasks: actual users' queries and ratings derived from an LLM battle platform (CArena), open-ended questions with single and multiple-turn dialogues (OPEN), and closed-ended questions with the same stems as open-ended single-turn ones (CLOSE). Our study shows that accuracy on closed-ended questions is insufficient to reflect human preferences achieved on open-ended ones. At the same time, they can complement each other to predict actual user preferences. We also demonstrate that GPT-4 is a reliable judge to automatically evaluate human preferences on open-ended questions in a Chinese context. Our benchmark will be released at https://www.CLUEbenchmarks.com

Large language models (LLMs) are increasingly being applied to programming tasks, ranging from single-turn code completion to autonomous agents. Current code agent designs frequently depend on complex, hand-crafted workflows and tool sets. However, this reliance on elaborate scaffolding presents several challenges: agent performance becomes overly dependent on prompt tuning and custom design choices, heavy human intervention obscures a model's true underlying capabilities, and intricate pipelines are costly to build and maintain. Furthermore, optimizing complex task prompts increases the risk of data leakage. Currently, when introducing new models, LLM providers like OpenAI and Anthropic often publish benchmark scores to demonstrate their models' coding proficiency, but keep their proprietary evaluation frameworks confidential. To address these limitations, we introduce Lita (Lite Agent), which operationalizes liteness, a principle of minimizing manual design while retaining the essential elements of a fully autonomous agent. Lita enables a more faithful and unified evaluation without elaborate scaffolding. Experiments on the Aider Polyglot and SWE-Bench with frontier models demonstrate that Lita achieves competitive or superior performance compared to workflow-based and agentic baselines. Crucially, Lita also consumes fewer tokens and requires significantly less design effort. Our results suggest that Lita is sufficient to reveal the underlying coding competence of modern LLMs. Finally, we propose the Agent Complexity Law: the performance gap between agents of varying complexity, from simple to sophisticated designs, will shrink as the core model improves, ultimately converging to a negligible difference.

Production large language-model (LLM) agents are increasingly deployed not as lone problem-solvers but as managers: a main model creates specialized subagents, delegates work, and orchestrates their parallel, asynchronous returns through dynamic workflows. Whether one model can actually run such a team is largely unmeasured: existing benchmarks score a policy's own task-solving or a fixed multi-agent system's emergent behavior, but none isolate the management ability of the single LLM acting as leader. We introduce ClawArena-Team, a benchmark of 41 multi-turn, multimodal, multi-directory scenarios spanning 258 evaluation rounds and 72 staged updates that measures this management ability. The main agent is deliberately constrained: it natively perceives only text and directly accesses only part of the workspace. It commands a fixed, locally served subagent pool, so score differences reflect management skill, not raw capability. All scoring is execution-based with no LLM judge: an overall score -- the Subagent-Management Score (SMS) -- multiplies task correctness by a least-privilege and modality-routing factor. Across twelve proprietary, community-hosted, and self-hosted models, experiments show that the management bottleneck is privilege granting rather than perception (no model exceeds 50% workspace-permission precision); that cost and management quality are decoupled (API cost spans over 100 times while the overall score spans under 4 times, with the cheapest open models on the Pareto frontier); and that most leaderboard scores cluster within a 9.9-point band while orchestration behaviors diverge by more than an order of magnitude. Code is available at https://github.com/aiming-lab/ClawArena.

Multimodal Large Language Models (MLLMs) are increasingly applied in real-world scenarios where user-provided images are often imperfect, requiring active image manipulations such as cropping, editing, or enhancement to uncover salient visual cues. Beyond static visual perception, MLLMs must also think with images: dynamically transforming visual content and integrating it with other tools to solve complex tasks. However, this shift from treating vision as passive context to a manipulable cognitive workspace remains underexplored. Most existing benchmarks still follow a think about images paradigm, where images are regarded as static inputs. To address this gap, we introduce VisualToolBench, a visual tool-use reasoning benchmark that rigorously evaluates MLLMs' ability to perceive, transform, and reason across complex visual-textual tasks under the think-with-images paradigm. VisualToolBench comprises 1,204 challenging, open-ended vision tasks (603 single-turn, 601 multi-turn) spanning across five diverse domains, each paired with detailed rubrics to enable systematic evaluation. Our evaluation shows that current MLLMs struggle with tasks requiring effective integration of vision and general-purpose tools. Even the strongest model (GPT-5-think) reaches only 18.68% pass rate. We further observe divergent tool-use behaviors, with OpenAI models benefiting from diverse image manipulations while Gemini-2.5-pro shows no improvement. By introducing the first benchmark centered on think with images, VisualToolBench offers critical insights for advancing visual intelligence in MLLMs.

In addressing the task of converting natural language to SQL queries, there are several semantic and syntactic challenges. It becomes increasingly important to understand and remedy the points of failure as the performance of semantic parsing systems improve. We explore semantic parse correction with natural language feedback, proposing a new solution built on the success of autoregressive decoders in text-to-SQL tasks. By separating the semantic and syntactic difficulties of the task, we show that the accuracy of text-to-SQL parsers can be boosted by up to 26% with only one turn of correction with natural language. Additionally, we show that a T5-base model is capable of correcting the errors of a T5-large model in a zero-shot, cross-parser setting.

Evaluating the alignment capabilities of large Vision-Language Models (VLMs) is essential for determining their effectiveness as helpful assistants. However, existing benchmarks primarily focus on basic abilities using nonverbal methods, such as yes-no and multiple-choice questions. In this paper, we address this gap by introducing AlignMMBench, a comprehensive alignment benchmark specifically designed for emerging Chinese VLMs. This benchmark is meticulously curated from real-world scenarios and Chinese Internet sources, encompassing thirteen specific tasks across three categories, and includes both single-turn and multi-turn dialogue scenarios. Incorporating a prompt rewrite strategy, AlignMMBench encompasses 1,054 images and 4,978 question-answer pairs. To facilitate the evaluation pipeline, we propose CritiqueVLM, a rule-calibrated evaluator that exceeds GPT-4's evaluation ability. Finally, we report the performance of representative VLMs on AlignMMBench, offering insights into the capabilities and limitations of different VLM architectures. All evaluation codes and data are available on https://alignmmbench.github.io.

Uncertainty quantification (UQ) for large language models (LLMs) is a key building block for safety guardrails of daily LLM applications. Yet, even as LLM agents are increasingly deployed in highly complex tasks, most UQ research still centers on single-turn question-answering. We argue that UQ research must shift to realistic settings with interactive agents, and that a new principled framework for agent UQ is needed. This paper presents the first general formulation of agent UQ that subsumes broad classes of existing UQ setups. Under this formulation, we show that prior works implicitly treat LLM UQ as an uncertainty accumulation process, a viewpoint that breaks down for interactive agents in an open world. In contrast, we propose a novel perspective, a conditional uncertainty reduction process, that explicitly models reducible uncertainty over an agent's trajectory by highlighting "interactivity" of actions. From this perspective, we outline a conceptual framework to provide actionable guidance for designing UQ in LLM agent setups. Finally, we conclude with practical implications of the agent UQ in frontier LLM development and domain-specific applications, as well as open remaining problems.

Large Language Models (LLMs) show promise as data analysis agents, but existing benchmarks overlook the iterative nature of the field, where experts' decisions evolve with deeper insights of the dataset. To address this, we introduce IDA-Bench, a novel benchmark evaluating LLM agents in multi-round interactive scenarios. Derived from complex Kaggle notebooks, tasks are presented as sequential natural language instructions by an LLM-simulated user. Agent performance is judged by comparing its final numerical output to the human-derived baseline. Initial results show that even state-of-the-art coding agents (like Claude-3.7-thinking) succeed on < 50% of the tasks, highlighting limitations not evident in single-turn tests. This work underscores the need to improve LLMs' multi-round capabilities for building more reliable data analysis agents, highlighting the necessity of achieving a balance between instruction following and reasoning.

LLM agents accumulate histories that outgrow their context windows, motivating a growing literature on memory systems. Yet most existing designs are tuned to a single scenario (multi-session chat or a single trajectory format), and there is little evidence that they generalize across the heterogeneous trajectories agents encounter in deployment. We revisit eight memory systems plus an agentic harness for search problems, on five scenarios: single-turn QA, multi-session chat, agentic-trajectory QA, memory stress tests, and long-horizon agentic tasks. The harness, which self-manages flat text-file storage via tool calls, achieves the best cross-task ranking, suggesting that memory performance hinges on giving the agent active control over storage and retrieval rather than on a passive store behind a fixed pipeline. We instantiate this insight in AutoMEM, an agentic memory harness with a self-managed tool interface that achieves the best cross-scenario generality among the systems we evaluate.

A central piece in enabling intelligent agentic behavior in foundation models is to make them capable of introspecting upon their behavior, reasoning, and correcting their mistakes as more computation or interaction is available. Even the strongest proprietary large language models (LLMs) do not quite exhibit the ability of continually improving their responses sequentially, even in scenarios where they are explicitly told that they are making a mistake. In this paper, we develop RISE: Recursive IntroSpEction, an approach for fine-tuning LLMs to introduce this capability, despite prior work hypothesizing that this capability may not be possible to attain. Our approach prescribes an iterative fine-tuning procedure, which attempts to teach the model how to alter its response after having executed previously unsuccessful attempts to solve a hard test-time problem, with optionally additional environment feedback. RISE poses fine-tuning for a single-turn prompt as solving a multi-turn Markov decision process (MDP), where the initial state is the prompt. Inspired by principles in online imitation learning and reinforcement learning, we propose strategies for multi-turn data collection and training so as to imbue an LLM with the capability to recursively detect and correct its previous mistakes in subsequent iterations. Our experiments show that RISE enables Llama2, Llama3, and Mistral models to improve themselves with more turns on math reasoning tasks, outperforming several single-turn strategies given an equal amount of inference-time computation. We also find that RISE scales well, often attaining larger benefits with more capable models. Our analysis shows that RISE makes meaningful improvements to responses to arrive at the correct solution for challenging prompts, without disrupting one-turn abilities as a result of expressing more complex distributions.

When prompting a language model (LM), users frequently expect the model to adhere to a set of behavioral principles across diverse tasks, such as producing insightful content while avoiding harmful or biased language. Instilling such principles into a model can be resource-intensive and technically challenging, generally requiring human preference labels or examples. We introduce SAMI, a method for teaching a pretrained LM to follow behavioral principles that does not require any preference labels or demonstrations. SAMI is an iterative algorithm that finetunes a pretrained LM to increase the conditional mutual information between constitutions and self-generated responses given queries from a datasest. On single-turn dialogue and summarization, a SAMI-trained mistral-7b outperforms the initial pretrained model, with win rates between 66% and 77%. Strikingly, it also surpasses an instruction-finetuned baseline (mistral-7b-instruct) with win rates between 55% and 57% on single-turn dialogue. SAMI requires a "principle writer" model; to avoid dependence on stronger models, we further evaluate aligning a strong pretrained model (mixtral-8x7b) using constitutions written by a weak instruction-finetuned model (mistral-7b-instruct). The SAMI-trained mixtral-8x7b outperforms both the initial model and the instruction-finetuned model, achieving a 65% win rate on summarization. Our results indicate that a pretrained LM can learn to follow constitutions without using preference labels, demonstrations, or human oversight.

Aligning the behavior of Large language models (LLMs) with human intentions and values remains a critical challenge. Reinforcement learning from human feedback (RLHF) aligns LLMs by training a reward model (RM) on human preferences and fine-tuning the LLMs to maximize RM feedback. Despite its effectiveness and popularity, RLHF is prone to biased local optimization. It means RM fails to provide feedback that accurately aligns with human preference, causing LLMs to explore unexpected generalizations, and failing to achieve alignment objectives. To mitigate this issue, we propose a novel sequence-to-sequence (seq2seq) reward modeling method. Its key insight is that learning from language feedback rather than scalar feedback improves RLHF without additional annotations. We replaced the reward modeling target from binary maximum likelihood estimation (MLE) with sequence MLE. This method enables richer and fine-grained language feedback without additional annotations, models, or training stages. Our experiments demonstrated its effectiveness, specifically, reducing the refusal-to-response paradigm in single-turn safety dialogues and the long-response bias in text summarization tasks. We provide further analysis that seq2seq RM improves RLHF performance across 2B and 7B LLMs on 3 NLP tasks, achieving an average win rate of 76.9\%. We further show that seq2seq RM can still improve the performance of RLHF under out-of-distribution prompts.

Large Language Models (LLMs) have significantly advanced natural language processing, demonstrating strong capabilities in tasks such as text generation, summarization, and reasoning. Recently, their potential for automating precise text editing tasks across specialized domains, such as programming code, LaTeX, and structured database languages, has gained attention. However, current state-of-the-art LLMs still struggle with executing precise, instruction-driven edits, particularly when structural accuracy and strict adherence to domain conventions are required. To address these challenges, we introduce InstrEditBench, an automated benchmark dataset comprising over 30,000 structured editing tasks spanning diverse domains, including Wikipedia articles, LaTeX documents, source code, and database languages. Using this benchmark, we develop FineEdit, a specialized editing model explicitly trained for accurate, context-aware text modifications. Experimental evaluations demonstrate that FineEdit outperforms state-of-the-art models, achieving improvements of approximately 10% over Gemini models on single-turn edits, up to 30% over Llama-3.2-3B, and exceeding Mistral-7B-OpenOrca performance by over 40% on direct editing tasks. FineEdit also effectively generalizes to realistic multi-turn editing scenarios, highlighting its practical applicability.

Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, a baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes (N=32) with only N=4, reducing token consumption by over 65% across single-turn and tool-integrated reasoning tasks.

Large Language Models (LLMs) have displayed massive improvements in reasoning and decision-making skills and can hold natural conversations with users. Recently, many tool-use benchmark datasets have been proposed. However, existing datasets have the following limitations: (1). Insufficient evaluation scenarios (e.g., only cover limited tool-use scenes). (2). Extensive evaluation costs (e.g., GPT API costs). To address these limitations, in this work, we propose a multi-granularity tool-use benchmark for large language models called MTU-Bench. For the "multi-granularity" property, our MTU-Bench covers five tool usage scenes (i.e., single-turn and single-tool, single-turn and multiple-tool, multiple-turn and single-tool, multiple-turn and multiple-tool, and out-of-distribution tasks). Besides, all evaluation metrics of our MTU-Bench are based on the prediction results and the ground truth without using any GPT or human evaluation metrics. Moreover, our MTU-Bench is collected by transforming existing high-quality datasets to simulate real-world tool usage scenarios, and we also propose an instruction dataset called MTU-Instruct data to enhance the tool-use abilities of existing LLMs. Comprehensive experimental results demonstrate the effectiveness of our MTU-Bench. Code and data will be released at https: //github.com/MTU-Bench-Team/MTU-Bench.git.

While a lot of recent research focuses on enhancing the textual reasoning capabilities of Large Language Models (LLMs) by optimizing the multi-agent framework or reasoning chains, several benchmark tasks can be solved with 100% success through direct coding, which is more scalable and avoids the computational overhead associated with textual iterating and searching. Textual reasoning has inherent limitations in solving tasks with challenges in math, logics, optimization, and searching, which is unlikely to be solved by simply scaling up the model and data size. The recently released OpenAI GPT Code Interpreter and multi-agent frameworks such as AutoGen have demonstrated remarkable proficiency of integrating code generation and execution to solve complex tasks using LLMs. However, based on our experiments on 7 existing popular methods for steering code/text generation in both single- and multi-turn settings with 14 tasks and 6 types of LLMs (including the new O1-preview), currently there is no optimal method to correctly steer LLMs to write code when needed. We discover some interesting patterns on when models use code vs. textual reasoning with the evolution to task complexity and model sizes, which even result in an astonishingly inverse scaling law. We also discover that results from LLM written code are not always better than using textual reasoning, even if the task could be solved through code. To mitigate the above issues, we propose three methods to better steer LLM code/text generation and achieve a notable improvement. The costs of token lengths and runtime are thoroughly discussed for all the methods. We believe the problem of steering LLM code/text generation is critical for future research and has much space for further improvement. Project Page, Datasets, and Codes are available at https://yongchao98.github.io/CodeSteer/.

Long-form multimodal video understanding requires integrating vision, speech, and ambient audio with coherent long-range reasoning. Existing benchmarks emphasize either temporal length or multimodal richness, but rarely both and while some incorporate open-ended questions and advanced metrics, they mostly rely on single-score accuracy, obscuring failure modes. We introduce LongShOTBench, a diagnostic benchmark with open-ended, intent-driven questions; single- and multi-turn dialogues; and tasks requiring multimodal reasoning and agentic tool use across video, audio, and speech. Each item includes a reference answer and graded rubric for interpretable, and traceable evaluation. LongShOTBench is produced via a scalable, human-validated pipeline to ensure coverage and reproducibility. All samples in our LongShOTBench are human-verified and corrected. Furthermore, we present LongShOTAgent, an agentic system that analyzes long videos via preprocessing, search, and iterative refinement. On LongShOTBench, state-of-the-art MLLMs show large gaps: Gemini-2.5-Flash achieves 52.95%, open-source models remain below 30%, and LongShOTAgent attains 44.66%. These results underscore the difficulty of real-world long-form video understanding. LongShOTBench provides a practical, reproducible foundation for evaluating and improving MLLMs. All resources are available on GitHub: https://github.com/mbzuai-oryx/longshot.

Current LLM safety benchmarks predominantly rely on binary safe or unsafe labels, lacking explanatory depth and suffering from low consistency across automated judges. To address this, we introduce DiagSafe, a diagnostic evaluation framework that moves beyond binary classification by employing prompt-level rubrics. Grounded in globally applicable AI safety policies, each rubric provides explicit criteria along two dimensions: safety, which targets concrete harms with specific exemptions to suppress false positives, and usefulness, which independently evaluates instruction fulfillment. To ensure a robust and realistic evaluation, we construct a bilingual (English-Chinese) benchmark grounded in a comprehensive 84-label safety taxonomy, consisting of two subsets: DiagBase, which covers single- and multi-turn prompts with direct harmful queries and template-based jailbreaks, and DiagAttack, which embeds harmful intents into complex instruction-following tasks to form challenging instruction attacks. Extensive experiments demonstrate that DiagSafe significantly enhances the reliability of LLM-as-a-judge evaluation, achieving 95.34% agreement with human annotations -- substantially higher than the 69%-87% achieved by the baselines -- and high inter-judge consistency (Cohen's κ = 0.864). Furthermore, applying our framework reveals that leading LLMs differ markedly in their unsafe rates, and each model also exhibits unique vulnerabilities along different dimensions. In addition, we observe that models are particularly susceptible in Chinese and multi-turn contexts. Ultimately, DiagSafe provides a reliable and explainable tool for auditing and optimizing LLM alignment.

Among the hardest tasks for humans are those found in competitive programming where problems require sophisticated algorithmic thinking, puzzle solving, and the creation of effective code. As a domain to assess language models (LMs), it has not received enough attention, though. This study presents the ICPC benchmark, which consists of 254 international collegiate programming contest (ICPC) tasks. Each problem includes official analysis, reference code, and sample, high-quality unit, and hidden tests. We are able to develop and evaluate a variety of LM inference techniques for competitive programming with these resources. With zero-shot chain-of-thought prompting, we find that o1 only achieves a 19.1\% pass@1 solve rate. With our best inference technique, which combines multi-turn self-judge with reflection and retrieval over episodic information, raises this to 42.2\%. Furthermore, we conduct a new human-in-the-loop investigation to gain a deeper understanding of the remaining difficulties. Surprisingly, we discover that o1 can solve 17 out of 18 problems that were previously unsolvable by any model or technique with just a few specific instructions. A footstep toward LMs with grounded, imaginative, and algorithmic thinking is provided by our quantitative findings and qualitative research. We open-source our code and data at https://github.com/kraritt/zolve.

Training Large Language Models ( LLMs) to generate explicit reasoning before they produce an answer has been shown to improve their performance across various tasks such as mathematics and coding. However, fine-tuning LLMs on multi-turn reasoning datasets presents a unique challenge: LLMs must generate reasoning tokens that are excluded from subsequent inputs to the LLM. This discrepancy prevents us from processing an entire conversation in a single forward pass-an optimization readily available when we fine-tune on a multi-turn non-reasoning dataset. This paper proposes a novel approach that overcomes this limitation through response token duplication and a custom attention mask that enforces appropriate visibility constraints. Our approach significantly reduces the training time and allows efficient fine-tuning on multi-turn reasoning datasets.

We present CIFLEX (Contextual Instruction Flow for Sub-task Execution), which is a novel execution system for efficient sub-task handling in multi-turn interactions with a single on-device large language model (LLM). As LLMs become increasingly capable, a single model is expected to handle diverse sub-tasks that more effectively and comprehensively support answering user requests. Naive approach reprocesses the entire conversation context when switching between main and sub-tasks (e.g., query rewriting, summarization), incurring significant computational overhead. CIFLEX mitigates this overhead by reusing the key-value (KV) cache from the main task and injecting only task-specific instructions into isolated side paths. After sub-task execution, the model rolls back to the main path via cached context, thereby avoiding redundant prefill computation. To support sub-task selection, we also develop a hierarchical classification strategy tailored for small-scale models, decomposing multi-choice decisions into binary ones. Experiments show that CIFLEX significantly reduces computational costs without degrading task performance, enabling scalable and efficient multi-task dialogue on-device.

Wearable devices such as smart glasses are transforming the way people interact with their surroundings, enabling users to seek information regarding entities in their view. Multi-Modal Retrieval-Augmented Generation (MM-RAG) plays a key role in supporting such questions, yet there is still no comprehensive benchmark for this task, especially regarding wearables scenarios. To fill this gap, we present CRAG-MM -- a Comprehensive RAG benchmark for Multi-modal Multi-turn conversations. CRAG-MM contains a diverse set of 6.5K (image, question, answer) triplets and 2K visual-based multi-turn conversations across 13 domains, including 6.2K egocentric images designed to mimic captures from wearable devices. We carefully constructed the questions to reflect real-world scenarios and challenges, including five types of image-quality issues, six question types, varying entity popularity, differing information dynamism, and different conversation turns. We design three tasks: single-source augmentation, multi-source augmentation, and multi-turn conversations -- each paired with an associated retrieval corpus and APIs for both image-KG retrieval and webpage retrieval. Our evaluation shows that straightforward RAG approaches achieve only 32% and 43% truthfulness on CRAG-MM single- and multi-turn QA, respectively, whereas state-of-the-art industry solutions have similar quality (32%/45%), underscoring ample room for improvement. The benchmark has hosted KDD Cup 2025, attracting about 1K participants and 5K submissions, with winning solutions improving baseline performance by 28%, highlighting its early impact on advancing the field.

Lead optimization in drug discovery requires efficiently navigating vast chemical space through iterative cycles to enhance molecular properties while preserving structural similarity to the original lead compound. Despite recent advances, traditional optimization methods struggle with sample efficiency-achieving good optimization performance with limited oracle evaluations. Large Language Models (LLMs) provide a promising approach through their in-context learning and instruction following capabilities, which align naturally with these iterative processes. However, existing LLM-based methods fail to leverage this strength, treating each optimization step independently. To address this, we present POLO (Preference-guided multi-turn Optimization for Lead Optimization), which enables LLMs to learn from complete optimization trajectories rather than isolated steps. At its core, POLO introduces Preference-Guided Policy Optimization (PGPO), a novel reinforcement learning algorithm that extracts learning signals at two complementary levels: trajectory-level optimization reinforces successful strategies, while turn-level preference learning provides dense comparative feedback by ranking intermediate molecules within each trajectory. Through this dual-level learning from intermediate evaluation, POLO achieves superior sample efficiency by fully exploiting each costly oracle call. Extensive experiments demonstrate that POLO achieves 84% average success rate on single-property tasks (2.3x better than baselines) and 50% on multi-property tasks using only 500 oracle evaluations, significantly advancing the state-of-the-art in sample-efficient molecular optimization.

Large language models (LLMs) have advanced significantly in code generation, yet their ability to follow complex programming instructions with layered and diverse constraints remains underexplored. Existing benchmarks often prioritize functional correctness, overlooking the nuanced requirements found in real-world development. We introduce MultiCodeIF, a comprehensive benchmark designed to evaluate instruction-following in code generation across multiple dimensions: constraint type, hierarchical levels, and iterative refinement. Built upon a structured taxonomy of 9 categories and 27 constraint types, MultiCodeIF enables granular assessment of both functional and non-functional instruction adherence. Using an automated pipeline, ConstraGen, we synthesize and evolve 2,021 code tasks sourced from 14 programming languages, supporting multi-turn evaluation through feedback-driven task variants. Empirical evaluation of six state-of-the-art LLMs uncovers substantial performance disparities. The top-performing model, Claude-3-7-Sonnet, achieves 63.0% average constraint satisfaction, while smaller models like Qwen3-1.7B fall to 44.8%. Models perform well on explicit constraints, but struggle with implicit or abstract constraints. Tasks with multiple hierarchical constraints significantly reduce model success rates, from 54.5% in single-level to just 18.8% in multi-level scenarios. However, structured feedback enables progressive improvement: average constraint satisfaction rises from 63.0% to 83.4% over four iterative refinement rounds. MultiCodeIF provides a scalable, constraint-aware, and feedback-sensitive framework to benchmark LLMs under realistic code generation scenarios, bridging the gap between synthetic evaluations and real-world instruction complexity. The full benchmark dataset, evaluation pipeline, and source code are available at https://github.com/SYSUSELab/MultiCodeIF.

Emotional intelligence (EI) in Large Language Models (LLMs) is often evaluated through static understanding tasks or single-response dialogue generation. However, emotion management is interactive: a good model should not only recognize a user's emotion, but also improve the user's emotional and relational state over several turns. We introduce EIBench, a simulator-based benchmark for interactive emotion management. EIBench contains 2,222 scenarios, with 2,009 for training and 213 for held-out testing. The scenarios are organized by a 2x2 taxonomy covering Support, Defense, Repair, and Charm, which together capture different forms of support, boundary maintenance, trust repair, and rapport building. In each scenario, an LLM simulator plays the user, updates an emotion-relation state after each turn, and maps the final state to an anchor-based score. This design makes EIBench both an evaluation benchmark and a training environment: the final state gives the outcome reward, while the per-turn state updates provide dense feedback for RL. We evaluate 15 open- and closed-source LLMs. Current models perform well on support and rapport-building scenes, but struggle with boundary maintenance under user pressure. To improve the EI ability of LLMs, we propose Centered Turn-Credit GRPO (CTC-GRPO), a GRPO extension that reuses the simulator's per-turn state updates as dense turn-level feedback while preserving the final outcome reward. CTC-GRPO improves Qwen3-8B from -22.4 to +22.4 on EIBench and also improves on out-of-distribution evaluations including SAGE (+12.4) and EQBench3 (+20.9%). Our results show that simulator-tracked user states can support both evaluation and training for multi-turn emotion management.

Large-scale Speech Language Models (SpeechLMs) have enabled voice assistants capable of understanding natural spoken queries and performing complex tasks. However, existing speech benchmarks primarily focus on isolated capabilities such as transcription, or question-answering, and do not systematically evaluate agentic scenarios encompassing multilingual and cultural understanding, as well as adversarial robustness. To address this, we introduce VoiceAgentBench, a comprehensive benchmark designed to evaluate SpeechLMs in realistic spoken agentic settings. It comprises over 5,500 synthetic spoken queries, including dialogues grounded in Indian context, covering single-tool invocations, multi-tool workflows, multi-turn interactions, and safety evaluations. The benchmark supports English, Hindi, and 5 other Indian languages, reflecting real-world linguistic and cultural diversity. We simulate speaker variability using a novel sampling algorithm that selects audios for TTS voice conversion based on its speaker embeddings, maximizing acoustic and speaker diversity. Our evaluation measures tool selection accuracy, structural consistency, and the correctness of tool invocations, including adversarial robustness. Our experiments reveal significant gaps in contextual tool orchestration tasks, Indic generalization, and adversarial robustness, exposing critical limitations of current SpeechLMs.

With the rapid development of Large Vision Language Models, the focus of Graphical User Interface (GUI) agent tasks shifts from single-screen tasks to complex screen navigation challenges. However, real-world GUI environments, such as PC software and mobile Apps, are often complex and proprietary, making it difficult to obtain the comprehensive environment information needed for agent training and evaluation. This limitation hinders systematic investigation and benchmarking of agent navigation capabilities. To address this limitation, we introduce GUI Exploration Lab, a simulation environment engine for GUI agent navigation research that enables flexible definition and composition of screens, icons, and navigation graphs, while providing full access to environment information for comprehensive agent training and evaluation. Through extensive experiments, we find that supervised fine-tuning enables effective memorization of fundamental knowledge, serving as a crucial foundation for subsequent training. Building on this, single-turn reinforcement learning further enhances generalization to unseen scenarios. Finally, multi-turn reinforcement learning encourages the development of exploration strategies through interactive trial and error, leading to further improvements in screen navigation performance. We validate our methods on both static and interactive benchmarks, demonstrating that our findings generalize effectively to real-world scenarios. These findings demonstrate the advantages of reinforcement learning approaches in GUI navigation and offer practical guidance for building more capable and generalizable GUI agents.

Large language models (LLMs) are increasingly deployed as agents, expected to decompose goals, invoke tools, and verify results in dynamic environments. Realizing these capabilities requires access to agentic data-structured interaction records that couple user intents with tool specifications, argument-grounded calls, and verifiable execution traces. However, collecting such data from human annotators is costly, time-consuming, and difficult to scale. We present a unified framework for synthesizing agentic data using only LLMs, without any human-in-the-loop supervision. This framework decomposes generation into modular pipelines that produce complete interaction records spanning task specifications, tool definitions, policy pseudocode, natural language exchanges, and execution traces. Records conform to strict syntactic and semantic constraints, ensuring machine-parseability and faithful alignment across inputs, outputs, and tool calls. Beyond single tasks, there is support for both multi-task and multi-turn agent interactions, enabling the construction of datasets that reflect the full spectrum of tool-use competencies. To ensure quality and consistency, the framework integrates constrained generation formats, JSON-schema validation, and judge-based filtering. This paper formalizes the schema for agentic records, details the prompt design principles that guide generation, and introduces scalable pipelines for high-quality synthetic data. By providing a reproducible, LLM-only alternative to manual collection, hence advancing the development of agentic LLMs capable of robust tool use.

Transcribing and understanding multi-speaker conversations requires speech recognition, speaker attribution, and timestamp localization. While speech LLMs excel at single-speaker tasks, multi-speaker scenarios remain challenging due to overlapping speech, backchannels, rapid turn-taking, and context window constraints. We propose Speaker-Reasoner, an end-to-end Speech LLM with agentic multi-turn temporal reasoning. Instead of single-pass inference, the model iteratively analyzes global audio structure, autonomously predicts temporal boundaries, and performs fine-grained segment analysis, jointly modeling speaker identity, gender, timestamps, and transcription. A speaker-aware cache further extends processing to audio exceeding the training context window. Trained with a three-stage progressive strategy, Speaker-Reasoner achieves consistent improvements over strong baselines on AliMeeting and AISHELL-4 datasets, particularly in handling overlapping speech and complex turn-taking.

Developing Large Language Models (LLMs) to cooperate and compete effectively within multi-agent systems is a critical step towards more advanced intelligence. While reinforcement learning (RL) has proven effective for enhancing reasoning in single-agent tasks, its extension to multi-turn, multi-agent scenarios remains underexplored due to the challenges of long-horizon credit assignment and agent-specific advantage estimation. To address these challenges, we introduce MARS, an end-to-end RL framework that incentivizes Multi-Agent Reasoning of LLMs through Self-play in both cooperative and competitive games. MARS features a turn-level advantage estimator that aligns learning signals with each interaction for credit assignment, and an agent-specific advantage normalization to stabilize multi-agent training. By learning with self-play across cooperative and competitive games, the MARS agent trained from Qwen3-4B develops strong strategic abilities that generalize to held-out games with up to 28.7% performance improvements. More importantly, the capability acquired through self-play generalizes beyond games, yielding consistent performance gains of multi-agent systems in reasoning benchmarks. When integrated into leading multi-agent systems, our MARS agent achieves significant performance gains of 10.0% on AIME and 12.5% on GPQA-Diamond. These results establish end-to-end RL training with self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs. Our code and models are publicly available at https://github.com/thu-nics/MARS.

Large language models (LLMs) deployed as conversational agents are increasingly expected to follow procedural logic. While factual hallucination has received extensive attention, a different failure mode has been overlooked: procedural hallucination, in which an agent produces individually fluent and factually plausible utterances while deviating from the intended conversational process. We introduce DialFork, the first benchmark to measure this failure mode on raw dialogue text. DialFork is built by forking conformant dialogues at controlled turn positions and injecting one of four hallucination types, yielding 36,278 dialogues across 24 tasks in 12 domains; 30 human annotators validate the construction. We define three evaluation tasks that characterize procedural deviation at three granularities: how far a set of dialogues deviates from conformant ones (monitoring deployed agents), how far a single dialogue does (procedural retrieval), and at which turn a dialogue first goes off-track (real-time guardrails). For the first two, we propose dialog-flow-aware metrics, FlowKL and FlowScore, that compare dialogues by how they unfold rather than their content. At the dialogue level, FlowScore significantly outperforms all baselines across seven encoders, showing that semantic similarity is no substitute for procedural similarity. We release DialFork under the MIT license at https://github.com/idiap/dialfork, hoping the community will measure conversational process alongside content.

As large language models (LLMs) become high-privilege agents in risk-sensitive settings, they introduce systemic threats beyond hallucination, where minor compliance errors can cause critical data leaks. However, existing benchmarks focus on rule-based QA, lacking agentic execution modeling, overlooking compliance drift in adversarial interactions, and relying on binary safety metrics that fail to capture behavioral degradation. To bridge these gaps, we present CNFinBench, a comprehensive benchmark spanning 29 subtasks grounded in the triad of expertise, autonomy, and integrity. It assesses domain-specific capabilities through certified regulatory corpora and professional financial tasks, reconstructs end-to-end agent workflows from requirement parsing to tool verification, and simulates multi-turn adversarial attacks that induce behavioral compliance drift. To quantify safety degradation, we introduce the Harmful Instruction Compliance Score (HICS), a multi-dimensional safety metric that integrates risk-type-specific deductions, multi-turn consistency tracking, and severity-adjusted penalty scaling based on fine-grained violation triggers. Evaluations over 22 open-/closed-source models reveal: LLMs perform well in applied tasks yet lack robust rule understanding, suffer a 15.4-point drop single modules to full execution chains, and collapse rapidly in multi-turn attacks, with average violations surging by 172.3% in Round 2. CNFinBench is available at https://cnfinbench.opencompass.org.cn and https://github.com/VertiAIBench/CNFinBench.

Large language model agents have shown strong potential in complex interactive tasks, yet their reinforcement learning (RL) is often hindered by sparse rewards, as a long multi-turn trajectory may receive only a single outcome-level signal. On-policy self-distillation (OPSD) provides dense token-level supervision from a privileged teacher, but the teacher may not be reliable at every position. Existing methods commonly rely on isolated token-level discrepancies, which can be sensitive to noise, or assign a shared step-level weight that may overlook positional variation. We propose Persistent Consistency Self-Distillation (PCSD), which derives token-level distillation weights from the local persistence of teacher-favoring signals. PCSD combines adaptive windows with exponentially decayed aggregation to capture persistent relative teacher support, applies trend-aware modulation to attenuate locally declining support, and produces continuous weights through sigmoid gating. The resulting objective is jointly optimized with GRPO, combining dense teacher guidance with sparse environmental feedback. Without inference-time skills, PCSD achieves the best ALFWorld Overall results among all baselines on both backbones, exceeding GRPO by 15.6 and 13.3 points and SDAR by 6.2 and 5.5 points, while remaining competitive on WebShop and gaining 15.8 points over GRPO on unseen ALFWorld split.

Existing Task-Oriented Dialogue (TOD) systems primarily focus on single-session dialogues, limiting their effectiveness in long-term memory augmentation. To address this challenge, we introduce a MS-TOD dataset, the first multi-session TOD dataset designed to retain long-term memory across sessions, enabling fewer turns and more efficient task completion. This defines a new benchmark task for evaluating long-term memory in multi-session TOD. Based on this new dataset, we propose a Memory-Active Policy (MAP) that improves multi-session dialogue efficiency through a two-stage approach. 1) Memory-Guided Dialogue Planning retrieves intent-aligned history, identifies key QA units via a memory judger, refines them by removing redundant questions, and generates responses based on the reconstructed memory. 2) Proactive Response Strategy detects and correct errors or omissions, ensuring efficient and accurate task completion. We evaluate MAP on MS-TOD dataset, focusing on response quality and effectiveness of the proactive strategy. Experiments on MS-TOD demonstrate that MAP significantly improves task success and turn efficiency in multi-session scenarios, while maintaining competitive performance on conventional single-session tasks.

Humans can effortlessly perceive spatial layouts, form cognitive representations, reason about spatial relations, and translate such reasoning into actions in everyday 3D environments. Although recent vision-language models (VLMs) have shown promising performance on observation-conditioned spatial perception and reasoning tasks, it remains unclear whether they can build coherent spatial understanding, act upon it, and refine their actions through multi-turn feedback. To study this problem, we introduce SpatialAct, a simulator-grounded benchmark for probing action-conditioned spatial reasoning in 3D scenes. Starting from the most challenging setting, Multi-turn Interactive Refinement, we further design its decomposed counterpart, Single-step Error Detection and Fix, together with five fundamental spatial ability tasks to diagnose the underlying causes of model failures. Experiments reveal a clear reasoning-to-action gap: current VLMs can perform well on isolated spatial reasoning tasks, but struggle to maintain coherent spatial beliefs and produce reliable actions during multi-turn feedback, substantially underperforming humans. These results suggest that current VLM agents still lack robust spatial state tracking under action-induced environment changes, even when low-level control is abstracted away.

Real-world agents operate over long and evolving horizons, where information is repeatedly updated and may interfere across memories, requiring accurate recall and aggregated reasoning over multiple pieces of information. However, existing benchmarks focus on static, independent recall and fail to capture these dynamic interactions between evolving memories. In this paper, we study how current memory-augmented agents perform in realistic, interference-heavy, long-horizon settings across diverse domains and question types. We introduce MINTEval (Long-Horizon Memory under INTerference Evaluation), a benchmark featuring (1) long, highly interconnected contexts with frequently updated information that induces substantial interference, (2) diverse domains (state tracking, multi-turn dialogue, Wikipedia revisions, and GitHub commits), enabling evaluation of domain generalization, and (3) diverse question types that assess robustness to interference, including (i) single-target recall tasks requiring retrieval of a specific target from long contexts, and (ii) multi-target aggregation tasks requiring reasoning over multiple relevant pieces of information. Overall, MINTEval has 15.6k question-answering pairs over long-horizon contexts averaging 138.8k tokens and extending up to 1.8M tokens per instance. We evaluate 7 representative systems, including vanilla long-context LLMs, RAG, and memory-augmented agent frameworks. Across all systems, we observe consistently low performance (avg. 27.9% accuracy), especially on questions requiring aggregated reasoning over multiple pieces of evidence. Our analysis shows that performance is primarily limited by retrieval and memory construction. Furthermore, current memory systems struggle to recall and reason over earlier facts that are revised or interfered with by subsequent context, with accuracy degrading as the number of intervening updates increases.

Recent work has shown that reinforcement learning from execution feedback can substantially improve text-to-SQL performance, often enabling smaller models to match or exceed much larger systems. However, most existing approaches treat SQL generation as a single-turn task, limiting the model's ability to recover from errors through iterative refinement. We present ReToolSQL, a two-stage training framework for text-to-SQL that combines (i) a supervised warm-start on rejection-sampled reasoning traces with (ii) agentic reinforcement fine-tuning (RFT) over multi-turn tool-use trajectories. The key insight is that the two stages act on complementary axes, the supervised fine-tuning (SFT) on verified privileged-teacher traces expands the set of solvable questions (raising pass@k coverage on the hardest cases), while RFT converts that expanded capability into higher single-pass accuracy by teaching the model when to verify, what evidence to retrieve, and how to repair faulty SQL from execution feedback. Applied to Gemma 4 instruction-tuned (31B), RFT alone achieves 73.66% execution accuracy (EX) on the BIRD-SQL development benchmark (74.12% EX with self-consistency). Initializing RFT from the SFT checkpoint (SFTtoRFT) yields our strongest model at 74.32% EX single-pass and 74.77% EX with self-consistency. At the time of writing, this ranked first on the BIRD single-model development-set leaderboard. The approach uses composite rewards anchored on execution correctness, requires no human annotation beyond the benchmark itself, and operates within a single dense 31B model, showing that a properly designed SFTtoRFT pipeline over tool-use trajectories is a practical path toward robust enterprise-grade text-to-SQL.

Large-language-model (LLM)-based AI agents have recently showcased impressive versatility by employing dynamic reasoning, an adaptive, multi-step process that coordinates with external tools. This shift from static, single-turn inference to agentic, multi-turn workflows broadens task generalization and behavioral flexibility, but it also introduces serious concerns about system-level cost, efficiency, and sustainability. This paper presents the first comprehensive system-level analysis of AI agents, quantifying their resource usage, latency behavior, energy consumption, and datacenter-wide power consumption demands across diverse agent designs and test-time scaling strategies. We further characterize how AI agent design choices, such as few-shot prompting, reflection depth, and parallel reasoning, impact accuracy-cost tradeoffs. Our findings reveal that while agents improve accuracy with increased compute, they suffer from rapidly diminishing returns, widening latency variance, and unsustainable infrastructure costs. Through detailed evaluation of representative agents, we highlight the profound computational demands introduced by AI agent workflows, uncovering a looming sustainability crisis. These results call for a paradigm shift in agent design toward compute-efficient reasoning, balancing performance with deployability under real-world constraints.

MLLM-based GUI agents have demonstrated strong capabilities in complex user interface interaction tasks. However, long-horizon scenarios remain challenging, as these agents are burdened with tasks beyond their intrinsic capabilities, suffering from memory degradation, progress confusion, and math hallucination. To address these challenges, we present UI-Copilot, a collaborative framework where the GUI agent focuses on task execution while a lightweight copilot provides on-demand assistance for memory retrieval and numerical computation. We introduce memory decoupling to separate persistent observations from transient execution context, and train the policy agent to selectively invoke the copilot as Retriever or Calculator based on task demands. To enable effective tool invocation learning, we propose Tool-Integrated Policy Optimization (TIPO), which separately optimizes tool selection through single-turn prediction and task execution through on-policy multi-turn rollouts. Experimental results show that UI-Copilot-7B achieves state-of-the-art performance on challenging MemGUI-Bench, outperforming strong 7B-scale GUI agents such as GUI-Owl-7B and UI-TARS-1.5-7B. Moreover, UI-Copilot-7B delivers a 17.1% absolute improvement on AndroidWorld over the base Qwen model, highlighting UI-Copilot's strong generalization to real-world GUI tasks.

Emotional Support Conversation (ESC) aims to assist individuals experiencing distress by generating empathetic and supportive dialogue. While prior work typically assumes that each supporter turn corresponds to a single strategy, real-world supportive communication often involves multiple strategies within a single utterance. In this paper, we revisit the ESC task by formulating it as multi-strategy utterance generation, where each utterance may contain one or more strategy-response pairs. We propose two generation methods: All-in-One, which predicts all strategy-response pairs in a single decoding step, and One-by-One, which iteratively generates strategy-response pairs until completion. Both methods are further enhanced with cognitive reasoning guided by reinforcement learning to improve strategy selection and response composition. We evaluate our models on the ESConv dataset under both utterance-level and dialogue-level settings. Experimental results show that our methods effectively model multi-strategy utterances and lead to improved supportive quality and dialogue success. To our knowledge, this work provides the first systematic empirical evidence that allowing multiple support strategies within a single utterance is both feasible and beneficial for emotional support conversations. All code and data will be publicly available at https://github.com/aliyun/qwen-dianjin.

Creating a system that can have meaningful conversations with humans to help accomplish tasks is one of the ultimate goals of Artificial Intelligence (AI). It has defined the meaning of AI since the beginning. A lot has been accomplished in this area recently, with voice assistant products entering our daily lives and chat bot systems becoming commonplace in customer service. At first glance there seems to be no shortage of options for dialogue systems. However, the frequently deployed dialogue systems today seem to all struggle with a critical weakness - they are hard to build and harder to maintain. At the core of the struggle is the need to script every single turn of interactions between the bot and the human user. This makes the dialogue systems more difficult to maintain as the tasks become more complex and more tasks are added to the system. In this paper, we propose Converse, a flexible tree-based modular task-oriented dialogue system. Converse uses an and-or tree structure to represent tasks and offers powerful multi-task dialogue management. Converse supports task dependency and task switching, which are unique features compared to other open-source dialogue frameworks. At the same time, Converse aims to make the bot building process easy and simple, for both professional and non-professional software developers. The code is available at https://github.com/salesforce/Converse.

Recent evaluations of Large Language Models (LLMs) have centered around testing their zero-shot/few-shot capabilities for basic natural language tasks and their ability to translate instructions into tool APIs. However, the evaluation of LLMs utilizing complex tools to finish multi-turn, multi-modal instructions in a complex multi-modal environment has not been investigated. To address this gap, we introduce the PowerPoint Task Completion (PPTC) benchmark to assess LLMs' ability to create and edit PPT files based on user instructions. It contains 279 multi-turn sessions covering diverse topics and hundreds of instructions involving multi-modal operations. We also propose the PPTX-Match Evaluation System that evaluates if LLMs finish the instruction based on the prediction file rather than the label API sequence, thus it supports various LLM-generated API sequences. We measure 3 closed LLMs and 6 open-source LLMs. The results show that GPT-4 outperforms other LLMs with 75.1\% accuracy in single-turn dialogue testing but faces challenges in completing entire sessions, achieving just 6\% session accuracy. We find three main error causes in our benchmark: error accumulation in the multi-turn session, long PPT template processing, and multi-modality perception. These pose great challenges for future LLM and agent systems. We release the data, code, and evaluation system of PPTC at https://github.com/gydpku/PPTC.

Major open-source serving stacks reuse the key-value tensors of a shared prefix by default. We measure what this prefix reuse costs in reproducibility. Holding model, decoding, seed and request order fixed at batch size one, we ran an eighty-episode multi-turn agentic tool-use workload with prefix reuse on and off across two engines and four weight formats. Without prefix reuse, repeated runs reproduced every token and log-probability in all ten configurations. With it, cache-hit and recompute paths diverged on both engines. On llama.cpp with GGUF weights this path divergence rises with quantization: for Qwen2.5-7B over ten independent server histories, it reached 37.4 percent of episodes at 16-bit and 82.5 percent at four-bit. Near-tie decisions became only about 1.2 times as common, while the cache-path change in top-two margins grew 4.9-fold from 16-bit to three-bit. Serving four-bit weights through 16-bit kernels left divergence not detectably different from 16-bit, placing the amplification mainly in the quantized execution path. At Q4_K_M, run-to-run divergence came mainly from llama.cpp's server-level prompt cache: under the canonical episode order it moved divergence by 37.5 points, and at its default, divergence spanned 20.0 to 28.7 percent under five random orders. With cache state restored, each path reproduced exactly while the two still differed. In the tested stacks, cached serving is irreproducible because the cache state is absent from the request and not reset by default. A single-turn bridge shows path divergence reaching task outcomes with no detectable net change in accuracy. We release the harness, logs, and analysis pipeline.

Large language models handle single-turn generation well, but multi-turn interactions still require the model to reconstruct user intent and task state from an expanding token history because internal representations do not persist across turns. This token-first paradigm leads to drift, inconsistent reasoning modes, and growing prompts as conversations deepen. We propose CORE, a concept-first interaction layer that improves multi-turn stability without modifying model weights. CORE combines a small library of universal cognitive operators with a persistent Local Concept - a compact semantic state capturing the task, constraints, preferences, and intermediate results. Each model call receives only this concept state, the user's latest instruction, and the selected operator, eliminating the need to replay full history. A preliminary prototype simulating CORE's behavior shows about 42% reduction in cumulative prompt tokens, though this number reflects prototype conditions and should not be interpreted as a real-world performance estimate. CORE offers a model-agnostic mechanism that separates conceptual reasoning from language generation, suggesting a scalable direction for more stable multi-turn systems.

Autoregressive (AR) vision-language models (VLMs) have long dominated multimodal understanding, reasoning, and graphical user interface (GUI) grounding. Recently, discrete diffusion vision-language models (DVLMs) have shown strong performance in multimodal reasoning, offering bidirectional attention, parallel token generation, and iterative refinement. However, their potential for GUI grounding remains unexplored. In this work, we evaluate whether discrete DVLMs can serve as a viable alternative to AR models for GUI grounding. We adapt LLaDA-V for single-turn action and bounding-box prediction, framing the task as text generation from multimodal input. To better capture the hierarchical structure of bounding-box geometry, we propose a hybrid masking schedule that combines linear and deterministic masking, improving grounding accuracy by up to 6.1 points in Step Success Rate (SSR) over the GUI-adapted LLaDA-V trained with linear masking. Evaluations on four datasets spanning web, desktop, and mobile interfaces show that the adapted diffusion model with hybrid masking consistently outperforms the linear-masked variant and performs competitively with autoregressive counterparts despite limited pretraining. Systematic ablations reveal that increasing diffusion steps, generation length, and block length improves accuracy but also increases latency, with accuracy plateauing beyond a certain number of diffusion steps. Expanding the training data with diverse GUI domains further reduces latency by about 1.3 seconds and improves grounding accuracy by an average of 20 points across benchmarks. These results demonstrate that discrete DVLMs are a promising modeling framework for GUI grounding and represent an important step toward diffusion-based GUI agents.

As large language models (LLMs) evolve into autonomous agents that execute long-horizon workflows, invoking a high-capability model at every step becomes economically unsustainable. While model routing is effective for single-turn queries, agentic routing is a sequential, path-dependent problem: early mistakes compound, feedback is often at the end of the episode, and deployments often demand strict per-task spending limits. We propose Budget-Aware Agentic Routing, which selects between a cheap and an expensive model at each step to optimize the cost--success frontier and to operate under strict per-task budgets. We propose Boundary-Guided Training, which leverages two boundary policies (always-small vs.\ always-large) to build a difficulty taxonomy and to anchor learning under sparse rewards. Our approach warms start with boundary-guided SFT data synthesis via stratified sampling of cost-efficient trajectories, then applies Boundary-Guided Policy Optimization (BoPO), combining boundary-relative rewards with a reference-guided advantage to avoid degenerate cheap-failure solutions. Experiment results show that our method improves the efficiency frontier, matching strong routing baselines at substantially lower cost while demonstrating generalization to strict inference-time budget constraints. Overall, our work establishes a foundational framework for agentic routing, shifting the paradigm from static model selection to dynamic, budget-aware sequential decision-making.

World models (WMs) are intended to serve as internal simulators of the real world that enable agents to understand, anticipate, and act upon complex environments. Existing WM benchmarks remain narrowly focused on next-state prediction and visual fidelity, overlooking the richer simulation capabilities required for intelligent behavior. To address this gap, we introduce WR-Arena, a comprehensive benchmark for evaluating WMs along three fundamental dimensions of next world simulation: (i) Action Simulation Fidelity, the ability to interpret and follow semantically meaningful, multi-step instructions and generate diverse counterfactual rollouts; (ii) Long-horizon Forecast, the ability to sustain accurate, coherent, and physically plausible simulations across extended interactions; and (iii) Simulative Reasoning and Planning, the ability to support goal-directed reasoning by simulating, comparing, and selecting among alternative futures in both structured and open-ended environments. We build a task taxonomy and curate diverse datasets designed to probe these capabilities, moving beyond single-turn and perceptual evaluations. Through extensive experiments with state-of-the-art WMs, our results expose a substantial gap between current models and human-level hypothetical reasoning, and establish WR-Arena as both a diagnostic tool and a guideline for advancing next-generation world models capable of robust understanding, forecasting, and purposeful action. The code is available at https://github.com/MBZUAI-IFM/WR-Arena.

Current instruction data synthesis methods primarily focus on single-turn instructions and often neglect cross-turn coherence, resulting in context drift and reduced task completion rates in extended conversations. To address this limitation, we propose Skeleton-Guided Multi-Turn Dialogue Generation, a framework that constrains multi-turn instruction synthesis by explicitly modeling human conversational intent. It operates in two stages: (1) Intent Modeling, which captures the global structure of human dialogues by assigning each conversation to one of nine well-defined intent trajectories, ensuring a coherent and goal-oriented information flow; and (2) Skeleton Generation, which constructs a structurally grounded sequence of user queries aligned with the modeled intent, thereby serving as a scaffold that constrains and guides the downstream instruction synthesis process. Based on this process, we construct ConsistentChat, a multi-turn instruction dataset with approximately 15,000 multi-turn conversations and 224,392 utterances. Experiments on the Light, Topdial, and MT-Eval benchmarks show that models fine-tuned on ConsistentChat achieve a 20-30% improvement in chat consistency and up to a 15% increase in task success rate, significantly outperforming models trained on existing single-turn and multi-turn instruction datasets.

The advancement of large language model (LLM) based agents has shifted AI evaluation from single-turn response assessment to multi-step task completion in interactive environments. We present an empirical study evaluating frontier AI models on 150 workplace tasks within a realistic e-commerce RL environment from Surge. Our analysis reveals an empirically-derived hierarchy of agentic capabilities that models must master for real-world deployment: (1) tool use, (2) planning and goal formation, (3) adaptability, (4) groundedness, and (5) common-sense reasoning. Even the best-performing models fail approximately 40\% of the tasks, with failures clustering predictably along this hierarchy. Weaker models struggle with fundamental tool use and planning, whereas stronger models primarily fail on tasks requiring contextual inference beyond explicit instructions. We introduce a task-centric design methodology for RL environments that emphasizes diversity and domain expert contributions, provide detailed failure analysis, and discuss implications for agent development. Our findings suggest that while current frontier models can demonstrate coherent multi-step behavior, substantial capability gaps remain before achieving human-level task completion in realistic workplace settings.

Systems for language-guided human-robot interaction must satisfy two key desiderata for broad adoption: adaptivity and learning efficiency. Unfortunately, existing instruction-following agents cannot adapt, lacking the ability to incorporate online natural language supervision, and even if they could, require hundreds of demonstrations to learn even simple policies. In this work, we address these problems by presenting Language-Informed Latent Actions with Corrections (LILAC), a framework for incorporating and adapting to natural language corrections - "to the right," or "no, towards the book" - online, during execution. We explore rich manipulation domains within a shared autonomy paradigm. Instead of discrete turn-taking between a human and robot, LILAC splits agency between the human and robot: language is an input to a learned model that produces a meaningful, low-dimensional control space that the human can use to guide the robot. Each real-time correction refines the human's control space, enabling precise, extended behaviors - with the added benefit of requiring only a handful of demonstrations to learn. We evaluate our approach via a user study where users work with a Franka Emika Panda manipulator to complete complex manipulation tasks. Compared to existing learned baselines covering both open-loop instruction following and single-turn shared autonomy, we show that our corrections-aware approach obtains higher task completion rates, and is subjectively preferred by users because of its reliability, precision, and ease of use.

A core challenge for autonomous LLM agents in collaborative settings is balancing robust privacy understanding and preservation alongside task efficacy. Existing privacy benchmarks only focus on simplistic, single-turn interactions where private information can be trivially omitted without affecting task outcomes. In this paper, we introduce MAGPIE (Multi-AGent contextual PrIvacy Evaluation), a novel benchmark of 200 high-stakes tasks designed to evaluate privacy understanding and preservation in multi-agent collaborative, non-adversarial scenarios. MAGPIE integrates private information as essential for task resolution, forcing agents to balance effective collaboration with strategic information control. Our evaluation reveals that state-of-the-art agents, including GPT-5 and Gemini 2.5-Pro, exhibit significant privacy leakage, with Gemini 2.5-Pro leaking up to 50.7% and GPT-5 up to 35.1% of the sensitive information even when explicitly instructed not to. Moreover, these agents struggle to achieve consensus or task completion and often resort to undesirable behaviors such as manipulation and power-seeking (e.g., Gemini 2.5-Pro demonstrating manipulation in 38.2% of the cases). These findings underscore that current LLM agents lack robust privacy understanding and are not yet adequately aligned to simultaneously preserve privacy and maintain effective collaboration in complex environments.

A significant use case of instruction-finetuned Large Language Models (LLMs) is to solve question-answering tasks interactively. In this setting, an LLM agent is tasked with making a prediction by sequentially querying relevant information from the user, as opposed to a single-turn conversation. This paper explores sequential querying strategies that aim to minimize the expected number of queries. One such strategy is Information Pursuit (IP), a greedy algorithm that at each iteration selects the query that maximizes information gain or equivalently minimizes uncertainty. However, obtaining accurate estimates of mutual information or conditional entropy for LLMs is very difficult in practice due to over- or under-confident LLM probabilities, which leads to suboptimal query selection and predictive performance. To better estimate the uncertainty at each iteration, we propose Conformal Information Pursuit (C-IP), an alternative approach to sequential information gain based on conformal prediction sets. More specifically, C-IP leverages a relationship between prediction sets and conditional entropy at each iteration to estimate uncertainty based on the average size of conformal prediction sets. In contrast to conditional entropy, we find that conformal prediction sets are a distribution-free and robust method of measuring uncertainty. Experiments with 20 Questions show that C-IP obtains better predictive performance and shorter query-answer chains compared to previous approaches to IP and uncertainty-based chain-of-thought methods. Furthermore, extending to an interactive medical setting between a doctor and a patient on the MediQ dataset, C-IP achieves competitive performance with direct single-turn prediction while offering greater interpretability.

Recent Large Language Models (LLMs) have demonstrated remarkable proficiency in code generation. However, their ability to create complex visualizations for scaled and structured data remains largely unevaluated and underdeveloped. To address this gap, we introduce PlotCraft, a new benchmark featuring 1k challenging visualization tasks that cover a wide range of topics, such as finance, scientific research, and sociology. The benchmark is structured around seven high-level visualization tasks and encompasses 48 distinct chart types. Crucially, it is the first to systematically evaluate both single-turn generation and multi-turn refinement across a diverse spectrum of task complexities. Our comprehensive evaluation of 23 leading LLMs on PlotCraft reveals obvious performance deficiencies in handling sophisticated visualization tasks. To bridge this performance gap, we develope SynthVis-30K, a large-scale, high-quality dataset of complex visualization code synthesized via a collaborative agent framework. Building upon this dataset, we develope PlotCraftor, a novel code generation model that achieves strong capabilities in complex data visualization with a remarkably small size. Across VisEval, PandasPlotBench, and our proposed PlotCraft, PlotCraftor shows performance comparable to that of leading proprietary approaches. Especially, on hard task, Our model achieves over 50% performance improvement. We will release the benchmark, dataset, and code at https://github.com/Speakn0w/PlotCraft-Benchmark.

In recent years, unified multimodal models (UMMs) have emerged to support both understanding and generation within a single framework. Mastering dynamic, multi-turn interleaved image-text dialogues is a crucial task for UMMs in real-world applications. However, existing benchmarks fail to evaluate this important task, as they are often limited to single-turn or static settings, and typically overlook exposure bias in multi-turn interactions. To bridge this gap, we propose IMUG-Bench, a comprehensive benchmark for multi-turn interleaved image-text dialogue of UMMs that jointly evaluates their understanding and generation capabilities. Our IMUG-Bench comprises three classes: Static Spatial, Temporal Causal, and Hybrid, covering 3,113 samples and 12,034 interaction turns. It also includes dynamic understanding questions, thereby supporting evaluation that better reflects real-world multi-turn interaction scenarios. Large-scale experiments on IMUG-Bench systematically evaluate mainstream open-source and closed-source UMMs, revealing their capability boundaries and failure modes, and uncovering pronounced exposure bias on the generation side in multi-turn interactions. We further explore several test-time scaling strategies, including Chain-of-Thought, Self-Verification, and Best-of-N Sampling, which effectively improve generation accuracy and mitigate exposure bias in generation tasks. These findings provide insights into enhancing the robustness and multi-turn interaction capability of future UMMs.

Commercial design platforms increasingly edit documents through large language model (LLM) agents, but two practical problems block reliable deployment: legacy document formats expose only flat, absolutely positioned elements, so agents must recompute coordinates and routinely break layouts; and design has no unique ground truth, so diff-against-reference metrics penalize valid-but-different outputs. We present ACE, an agentic canvas editor over a hierarchical scene-graph with a presentation-specialized action space (98 tools), paired with CARE, a content-aware router that feeds the agent only the relevant slice of each deck (avg.\ sim89\% input-token reduction), and a self-correction loop driven by a ground-truth-free instruction-following (IF) judge whose natural-language critique is fed back as the next-turn instruction. With a fixed backbone, a scene-graph editor in a single turn already matches a same-backbone agentic HTML pipeline that iterates internally; adding self-correction lifts ACE significantly above it on instruction following (IF 4.23 vs.\ 3.81 on the full 94-task benchmark, paired p{=}.010, replicated by an out-of-loop judge) at 1.75times the speed and sim44\% lower cost. VQ means are statistically indistinguishable, but 26 blind raters prefer ACE overall (58.7\% decisive win-rate) and prefer the self-corrected output 81\% of the time; the ranking is invariant across three judge families, and out-of-loop judges retain two-thirds of the self-correction gain, bounding circularity. 66\% of cases halt after one pass, and a strict-peak rollback removes every observed regression.

To create rich visualizations, data analysts often need to iterate back and forth among data processing and chart specification to achieve their goals. To achieve this, analysts need not only proficiency in data transformation and visualization tools but also efforts to manage the branching history consisting of many different versions of data and charts. Recent LLM-powered AI systems have greatly improved visualization authoring experiences, for example by mitigating manual data transformation barriers via LLMs' code generation ability. However, these systems do not work well for iterative visualization authoring, because they often require analysts to provide, in a single turn, a text-only prompt that fully describes the complex visualization task to be performed, which is unrealistic to both users and models in many cases. In this paper, we present Data Formulator 2, an LLM-powered visualization system to address these challenges. With Data Formulator 2, users describe their visualization intent with blended UI and natural language inputs, and data transformation are delegated to AI. To support iteration, Data Formulator 2 lets users navigate their iteration history and reuse previous designs towards new ones so that they don't need to start from scratch every time. In a user study with eight participants, we observed that Data Formulator 2 allows participants to develop their own iteration strategies to complete challenging data exploration sessions.

Automatic speech recognition (ASR) is a core component of human--computer interaction and an increasingly important front-end for LLM-based assistants and agents. However, most current ASR systems still follow a single-pass paradigm, which is poorly aligned with human communication, where misunderstandings are resolved through iterative clarification and refinement. This mismatch makes it difficult to correct meaning-critical errors once they occur. Meanwhile, token-level metrics such as WER or CER cannot adequately reflect such a problem. To address these limitations, we formulate Interactive ASR as a multi-turn refinement task and propose Agentic ASR, a closed-loop framework that combines a single-pass ASR front-end with semantic correction, intent routing, and reasoning-based editing. We further introduce the Sentence-level Semantic Error Rate (S^2ER), an LLM-based semantic evaluation metric, together with an Interactive Simulation System for scalable and reproducible benchmarking. Experiments on multilingual, named-entity-intensive, and code-switching benchmarks show that iterative interaction consistently reduces semantic errors, with much larger gains in S^2ER than in conventional token-level metrics. Human--AI alignment and ablation studies further validate the reliability of the semantic judge and the robustness of the proposed framework. The code is available at: https://interactiveasr.github.io/ and the live demo is available at https://i-asr.sjtuxlance.com/

Long-horizon terminal tasks are used to evaluate and train agents. Human-curated datasets are scarce, so many synthesis methods have emerged. They usually measure the length of a task by the turns a model needs to pass it, and its difficulty by labels the LLM sets during generation. We find, however, that the turn count largely reflects the habits of the model, not only the demands of the task, and that the labels barely track how often models pass. We describe a terminal task by its work, the number of operations it requires, and its depth, the number of turns that must happen in sequence because the commands of each turn can only be chosen after reading the output of the previous one. Estimating depth as the smallest turn cap, stated to the agent, under which the task passes, we audit 734 tasks from six existing datasets. Synthesis has mostly added work, not depth: 94% of the solved synthetic tasks pass within three turns, yet depth, unlike the turn count, predicts difficulty beyond work. We propose DEPTH, in which a generation agent writes each task one segment at a time in a real software environment, carrying out each segment before writing it and starting the next from the state it leaves. Our 105 tasks exceed every audited dataset in depth and work, and none of six current models passes half of them. Fine-tuning Qwen3-8B on runs from these tasks improves it on two terminal benchmarks.

As LLM agents transition from short, static problem solving to executing complex, long-horizon tasks in dynamic environments, the ability to handle user interruptions, such as adding requirement or revising goals, during mid-task execution is becoming a core requirement for realistic deployment. However, existing benchmarks largely assume uninterrupted agent behavior or study interruptions only in short, unconstrained language tasks. In this paper, we present the first systematic study of interruptible agents in long-horizon, environmentally grounded web navigation tasks, where actions induce persistent state changes. We formalize three realistic interruption types, including addition, revision, and retraction, and introduce InterruptBench, a benchmark derived from WebArena-Lite that synthesizes high-quality interruption scenarios under strict semantic constraints. Using a unified interruption simulation framework, we evaluate six strong LLM backbones across single- and multi-turn interruption settings, analyzing both their effectiveness in adapting to updated intents and their efficiency in recovering from mid-task changes. Our results show that handling user interruptions effectively and efficiently during long-horizon agentic tasks remains challenging for powerful large-scale LLMs. Code and dataset are available at https://github.com/HenryPengZou/InterruptBench.

Large Language Models (LLMs) with API-calling capabilities enabled building effective Language Agents (LA), while also revolutionizing the conventional task-oriented dialogue (TOD) paradigm. However, current approaches face a critical dilemma: TOD systems are often trained on a limited set of target APIs, requiring new data to maintain their quality when interfacing with new services, while LAs are not trained to maintain user intent over multi-turn conversations. Because both robust multi-turn management and advanced function calling are crucial for effective conversational agents, we evaluate these skills on three popular benchmarks: MultiWOZ 2.4 (TOD), BFCL V3 (LA), and API-Bank (LA), and our analyses reveal that specialized approaches excel in one domain but underperform in the other. To bridge this chasm, we introduce CALM (Conversational Agentic Language Model), a unified approach that integrates both conversational and agentic capabilities. We created CALM-IT, a carefully constructed multi-task dataset that interleave multi-turn ReAct reasoning with complex API usage. Using CALM-IT, we train three models CALM 8B, CALM 70B, and CALM 405B, which outperform top domain-specific models, including GPT-4o, across all three benchmarks.

We present the AILS-NTUA system for SemEval-2026 Task 8 (MTRAGEval), addressing all three subtasks of multi-turn retrieval-augmented generation: passage retrieval (A), reference-grounded response generation (B), and end-to-end RAG (C). Our unified architecture is built on two principles: (i) a query-diversity-over-retriever-diversity strategy, where five complementary LLM-based query reformulations are issued to a single corpus-aligned sparse retriever and fused via variance-aware nested Reciprocal Rank Fusion; and (ii) a multistage generation pipeline that decomposes grounded generation into evidence span extraction, dual-candidate drafting, and calibrated multi-judge selection. Our system ranks 1st in Task A (nDCG@5: 0.5776, +20.5% over the strongest baseline) and 2nd in Task B (HM: 0.7698). Empirical analysis shows that query diversity over a well-aligned retriever outperforms heterogeneous retriever ensembling, and that answerability calibration-rather than retrieval coverage-is the primary bottleneck in end-to-end performance.

Multi-turn long-horizon planning is critical for foundation model agents, yet how to fundamentally improve it remains unclear. Existing models are trained on uncontrollable and opaque Internet data, making it difficult to identify how planning ability is acquired, shaped, and integrated. To address this challenge, we introduce a unified and controlled multi-turn environment that enables precise control. It allows systematically study long-horizon planning across three stages. (1) Planning ability acquisition during pre-training. We study data format, distribution, and quality. Explicit world model construction through CoT state transition modeling yields stronger long-horizon generalization. Atomic skills alone are insufficient for compositional generalization, whereas a litte long-horizon data works. Moreover, suboptimal trajectories severely impair performance because errors amplify over long horizons. (2) Planning ability shaping via GRPO and OPD post-training. Through mutual information, we distinguish general planning patterns from task-specific planning knowledge. For planning patterns, we identify three application regions of post-training: unnecessary, effective, and unsupported. OPD has a broader effective region than GRPO under low-quality and long-horizon settings, as it provides more consistent update directions. For planning knowledge, distilling unseen procedures from a teacher with different knowledge may impair student's prior world modeling without fully establishing new knowledge. (3) Planning ability integration through MOPD post-training. We show that multi-teacher on-policy distillation (MOPD) integrates capabilities by converging to shared planning-pattern across environments. Compatible patterns enable cross-environment generalization, partially shared patterns support continual learning, while completely conflicting patterns cause severe interference.

We present our winning system for Task~B (generation with reference passages) in SemEval-2026 Task~8: MTRAGEval. Our method is a heterogeneous ensemble of seven LLMs with two prompting variants, where a GPT-4o-mini judge selects the best candidate per instance. We ranked 1st out of 26 teams, achieving a conditioned harmonic mean of 0.7827 and outperforming the strongest baseline (gpt-oss-120b, 0.6390). Ablations show that diversity in model families, scales, and prompting strategies is essential, with the ensemble consistently beating any single model. We also introduce Meno-Lite-0.1, a 7B domain-adapted model with a strong cost--performance trade-off, and analyse MTRAGEval, highlighting annotation limitations and directions for improvement. Our code is publicly available: https://github.com/RaguTeam/ragu_mtrag_semeval

Neuroimage processing tasks like segmentation, reconstruction, and registration are central to the study of neuroscience. Robust deep learning strategies and architectures used to solve these tasks are often similar. Yet, when presented with a new task or a dataset with different visual characteristics, practitioners most often need to train a new model, or fine-tune an existing one. This is a time-consuming process that poses a substantial barrier for the thousands of neuroscientists and clinical researchers who often lack the resources or machine-learning expertise to train deep learning models. In practice, this leads to a lack of adoption of deep learning, and neuroscience tools being dominated by classical frameworks. We introduce Neuralizer, a single model that generalizes to previously unseen neuroimaging tasks and modalities without the need for re-training or fine-tuning. Tasks do not have to be known a priori, and generalization happens in a single forward pass during inference. The model can solve processing tasks across multiple image modalities, acquisition methods, and datasets, and generalize to tasks and modalities it has not been trained on. Our experiments on coronal slices show that when few annotated subjects are available, our multi-task network outperforms task-specific baselines without training on the task.