Daily Papers

Despite recent advancements in language models (LMs), their application to dialogue management (DM) problems and ability to carry on rich conversations remain a challenge. We use reinforcement learning (RL) to develop a dialogue agent that avoids being short-sighted (outputting generic utterances) and maximizes overall user satisfaction. Most existing RL approaches to DM train the agent at the word-level, and thus, have to deal with a combinatorially complex action space even for a medium-size vocabulary. As a result, they struggle to produce a successful and engaging dialogue even if they are warm-started with a pre-trained LM. To address this issue, we develop a RL-based DM using a novel mixture of expert language model (MoE-LM) that consists of (i) a LM capable of learning diverse semantics for conversation histories, (ii) a number of {\em specialized} LMs (or experts) capable of generating utterances corresponding to a particular attribute or personality, and (iii) a RL-based DM that performs dialogue planning with the utterances generated by the experts. Our MoE approach provides greater flexibility to generate sensible utterances with different intents and allows RL to focus on conversational-level DM. We compare it with SOTA baselines on open-domain dialogues and demonstrate its effectiveness both in terms of the diversity and sensibility of the generated utterances and the overall DM performance.

What makes a good Large Language Model (LLM)? That it performs well on the relevant benchmarks -- which hopefully measure, with some validity, the presence of capabilities that are also challenged in real application. But what makes the model perform well? What gives a model its abilities? We take a recently introduced type of benchmark that is meant to challenge capabilities in a goal-directed, agentive context through self-play of conversational games, and analyse how performance develops as a function of model characteristics like number of parameters, or type of training. We find that while there is a clear relationship between number of parameters and performance, there is still a wide spread of performance points within a given size bracket, which is to be accounted for by training parameters such as fine-tuning data quality and method. From a more practical angle, we also find a certain degree of unpredictability about performance across access methods, possible due to unexposed sampling parameters, and a, very welcome, performance stability against at least moderate weight quantisation during inference.

We present DM-Bench, the first benchmark designed to evaluate large language model (LLM) performance across real-world decision-making tasks faced by individuals managing diabetes in their daily lives. Unlike prior health benchmarks that are either generic, clinician-facing or focused on clinical tasks (e.g., diagnosis, triage), DM-Bench introduces a comprehensive evaluation framework tailored to the unique challenges of prototyping patient-facing AI solutions in diabetes, glucose management, metabolic health and related domains. Our benchmark encompasses 7 distinct task categories, reflecting the breadth of real-world questions individuals with diabetes ask, including basic glucose interpretation, educational queries, behavioral associations, advanced decision making and long term planning. Towards this end, we compile a rich dataset comprising one month of time-series data encompassing glucose traces and metrics from continuous glucose monitors (CGMs) and behavioral logs (e.g., eating and activity patterns) from 15,000 individuals across three different diabetes populations (type 1, type 2, pre-diabetes/general health and wellness). Using this data, we generate a total of 360,600 personalized, contextual questions across the 7 tasks. We evaluate model performance on these tasks across 5 metrics: accuracy, groundedness, safety, clarity and actionability. Our analysis of 8 recent LLMs reveals substantial variability across tasks and metrics; no single model consistently outperforms others across all dimensions. By establishing this benchmark, we aim to advance the reliability, safety, effectiveness and practical utility of AI solutions in diabetes care.

LLM agents are increasingly expected to function as general-purpose systems capable of resolving open-ended user requests. While existing benchmarks focus on domain-aware environments for developing specialized agents, evaluating general-purpose agents requires more realistic settings that challenge them to operate across multiple skills and tools within a unified environment. We introduce General AgentBench, a benchmark that provides such a unified framework for evaluating general LLM agents across search, coding, reasoning, and tool-use domains. Using General AgentBench, we systematically study test-time scaling behaviors under sequential scaling (iterative interaction) and parallel scaling (sampling multiple trajectories). Evaluation of ten leading LLM agents reveals a substantial performance degradation when moving from domain-specific evaluations to this general-agent setting. Moreover, we find that neither scaling methodology yields effective performance improvements in practice, due to two fundamental limitations: context ceiling in sequential scaling and verification gap in parallel scaling. Code is publicly available at https://github.com/cxcscmu/General-AgentBench.

Large Language Models (LLMs) have marked a significant advancement in the field of natural language processing, demonstrating exceptional capabilities in reasoning, tool usage, and memory. As their applications extend into multi-agent environments, a need has arisen for a comprehensive evaluation framework that captures their abilities in reasoning, planning, collaboration, and more. This work introduces a novel benchmarking framework specifically tailored to assess LLMs within multi-agent settings, providing quantitative metrics to evaluate their judgment, reasoning, deception, self-awareness, cooperation, coordination, and rationality. We utilize games such as Chameleon and Undercover, alongside game theory scenarios like Cost Sharing, Multi-player Prisoner's Dilemma, and Public Good, to create diverse testing environments. Our framework is fortified with the Probabilistic Graphical Modeling (PGM) method, enhancing the LLMs' capabilities in navigating complex social and cognitive dimensions. The benchmark evaluates seven multi-agent systems powered by different LLMs, quantitatively highlighting a significant capability gap over threefold between the strongest, GPT-4, and the weakest, Llama-2-70B. It also confirms that our PGM enhancement boosts the inherent abilities of all selected models by 50% on average. Our codes are released here https://github.com/cathyxl/MAgIC.

The role of a Dungeon Master, or DM, in the game Dungeons & Dragons is to perform multiple tasks simultaneously. The DM must digest information about the game setting and monsters, synthesize scenes to present to other players, and respond to the players' interactions with the scene. Doing all of these tasks while maintaining consistency within the narrative and story world is no small feat of human cognition, making the task tiring and unapproachable to new players. Large language models (LLMs) like GPT-3 and ChatGPT have shown remarkable abilities to generate coherent natural language text. In this paper, we conduct a formative evaluation with DMs to establish the use cases of LLMs in D&D and tabletop gaming generally. We introduce CALYPSO, a system of LLM-powered interfaces that support DMs with information and inspiration specific to their own scenario. CALYPSO distills game context into bite-sized prose and helps brainstorm ideas without distracting the DM from the game. When given access to CALYPSO, DMs reported that it generated high-fidelity text suitable for direct presentation to players, and low-fidelity ideas that the DM could develop further while maintaining their creative agency. We see CALYPSO as exemplifying a paradigm of AI-augmented tools that provide synchronous creative assistance within established game worlds, and tabletop gaming more broadly.

We introduce Differential Performance Evaluation (DPE), a framework designed to reliably evaluate Large Language Models (LLMs) for efficient code generation. Traditional coding benchmarks often fail to provide reliable insights into code efficiency, due to their reliance on simplistic test inputs and the absence of effective compound metrics. DPE addresses these issues by focusing on efficiency-demanding programming tasks and establishing an insightful compound metric for performance evaluation. DPE operates in two phases: To curate efficiency datasets, it selects efficiency-demanding tasks from existing coding benchmarks and generates computationally expensive inputs to stress the efficiency of LLM solutions. To assess the code efficiency, DPE profiles the new solution and compares it globally against a set of reference solutions that exhibit distinct efficiency levels, where the matched level defines its efficiency score. As a proof of concept, we use DPE to create EvalPerf, a benchmark with 121 performance-challenging coding tasks. Our comprehensive evaluation draws interesting findings on the efficiency impact of model sizes, instruction tuning, and prompting. For example, while the scaling law fails to account for code efficiency, general instruction tuning benefits both code correctness and efficiency. We also evaluate the evaluation by examining the effectiveness of DPE, showing that EvalPerf is reliable and convenient to use even across platforms.

Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) have demonstrated impressive language/vision reasoning abilities, igniting the recent trend of building agents for targeted applications such as shopping assistants or AI software engineers. Recently, many data science benchmarks have been proposed to investigate their performance in the data science domain. However, existing data science benchmarks still fall short when compared to real-world data science applications due to their simplified settings. To bridge this gap, we introduce DSBench, a comprehensive benchmark designed to evaluate data science agents with realistic tasks. This benchmark includes 466 data analysis tasks and 74 data modeling tasks, sourced from Eloquence and Kaggle competitions. DSBench offers a realistic setting by encompassing long contexts, multimodal task backgrounds, reasoning with large data files and multi-table structures, and performing end-to-end data modeling tasks. Our evaluation of state-of-the-art LLMs, LVLMs, and agents shows that they struggle with most tasks, with the best agent solving only 34.12% of data analysis tasks and achieving a 34.74% Relative Performance Gap (RPG). These findings underscore the need for further advancements in developing more practical, intelligent, and autonomous data science agents.

Multi-agent systems (MAS) can scale large language model reasoning at test time by decomposing complex problems into parallel subtasks. However, most existing MAS rely on centralized orchestration, where a main agent assigns work, collects outputs, and merges results. As the number of subtasks grows, this controller becomes a communication and integration bottleneck. We propose Decentralized Language Models (DeLM), a MAS framework that decentralizes coordination through parallel agents, a shared verified context, and a task queue. Agents asynchronously claim subtasks, read accumulated progress, perform local reasoning, and write back compact verified updates. The shared context acts as a common communication substrate, enabling agents to build on one another's verified progress without routing every update through a central controller. Empirically, DeLM improves both software-engineering test-time scaling and long-context reasoning. On SWE-bench Verified, DeLM achieves the best performance across Avg.@1, Pass@2, and Pass@4, with gains of up to 10.5 percentage points over the strongest baseline, while reducing cost per task by roughly 50%. On LongBench-v2 Multi-Doc QA, DeLM achieves the highest average accuracy across four frontier model families, improving over the strongest baseline by up to 5.7 percentage points. The code is available on our project website at https://yuzhenmao.github.io/DeLM/.

Results on existing LLM benchmarks capture little information over the model capabilities in low-resource tasks, making it difficult to develop effective solutions in these domains. To address these challenges, we curated 14 travel-domain datasets spanning 7 common NLP tasks using anonymised data from real-world scenarios, and analysed the performance across LLMs. We report on the accuracy, scaling behaviour, and reasoning capabilities of LLMs in a variety of tasks. Our results confirm that general benchmarking results are insufficient for understanding model performance in low-resource tasks. Despite the amount of training FLOPs, out-of-the-box LLMs hit performance bottlenecks in complex, domain-specific scenarios. Furthermore, reasoning provides a more significant boost for smaller LLMs by making the model a better judge on certain tasks.

Guided by the belief of the scaling law, large language models (LLMs) have achieved impressive performance in recent years. However, scaling law only gives a qualitative estimation of loss, which is influenced by various factors such as model architectures, data distributions, tokenizers, and computation precision. Thus, estimating the real performance of LLMs with different training settings rather than loss may be quite useful in practical development. In this article, we present an empirical equation named "Performance Law" to directly predict the MMLU score of an LLM, which is a widely used metric to indicate the general capability of LLMs in real-world conversations and applications. Based on only a few key hyperparameters of the LLM architecture and the size of training data, we obtain a quite accurate MMLU prediction of various LLMs with diverse sizes and architectures developed by different organizations in different years. Performance law can be used to guide the choice of LLM architecture and the effective allocation of computational resources without extensive experiments.

We present an empirical evaluation of various outputs generated by nine of the most widely-available large language models (LLMs). Our analysis is done with off-the-shelf, readily-available tools. We find a correlation between percentage of memorized text, percentage of unique text, and overall output quality, when measured with respect to output pathologies such as counterfactual and logically-flawed statements, and general failures like not staying on topic. Overall, 80.0% of the outputs evaluated contained memorized data, but outputs containing the most memorized content were also more likely to be considered of high quality. We discuss and evaluate mitigation strategies, showing that, in the models evaluated, the rate of memorized text being output is reduced. We conclude with a discussion on potential implications around what it means to learn, to memorize, and to evaluate quality text.

Large Language Model (LLM) leaderboards based on benchmark rankings are regularly used to guide practitioners in model selection. Often, the published leaderboard rankings are taken at face value - we show this is a (potentially costly) mistake. Under existing leaderboards, the relative performance of LLMs is highly sensitive to (often minute) details. We show that for popular multiple choice question benchmarks (e.g. MMLU) minor perturbations to the benchmark, such as changing the order of choices or the method of answer selection, result in changes in rankings up to 8 positions. We explain this phenomenon by conducting systematic experiments over three broad categories of benchmark perturbations and identifying the sources of this behavior. Our analysis results in several best-practice recommendations, including the advantage of a hybrid scoring method for answer selection. Our study highlights the dangers of relying on simple benchmark evaluations and charts the path for more robust evaluation schemes on the existing benchmarks.

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.

This study evaluates the efficiency of code generation by Large Language Models (LLMs) and measures their performance against human-crafted solutions using a dataset from Leetcode. We compare 18 LLMs, considering factors such as model temperature and success rate, and their impact on code performance. This research introduces a novel method for measuring and comparing the speed of LLM-generated code, revealing that LLMs produce code with comparable performance, irrespective of the adopted LLM. We also find that LLMs are capable of generating code that is, on average, more efficient than the code written by humans. The paper further discusses the use of Leetcode as a benchmarking dataset, the limitations imposed by potential data contamination, and the platform's measurement reliability. We believe that our findings contribute to a better understanding of LLM capabilities in code generation and set the stage for future optimizations in the field.

Code benchmarks such as HumanEval are widely adopted to evaluate the capabilities of Large Language Models (LLMs), providing insights into their strengths and weaknesses. However, current benchmarks primarily exercise LLMs' capability on common coding tasks (e.g., bubble sort, greatest common divisor), leaving domain-specific coding tasks (e.g., computation, system, cryptography) unexplored. To fill this gap, we propose a multi-domain code benchmark, DOMAINEVAL, designed to evaluate LLMs' coding capabilities thoroughly. Our pipeline works in a fully automated manner, enabling a push-bottom construction from code repositories into formatted subjects under study. Interesting findings are observed by evaluating 12 representative LLMs against DOMAINEVAL. We notice that LLMs are generally good at computation tasks while falling short on cryptography and system coding tasks. The performance gap can be as much as 68.94% (80.94% - 12.0%) in some LLMs. We also observe that generating more samples can increase the overall performance of LLMs, while the domain bias may even increase. The contributions of this study include a code generation benchmark dataset DOMAINEVAL, encompassing six popular domains, a fully automated pipeline for constructing code benchmarks, and an identification of the limitations of LLMs in code generation tasks based on their performance on DOMAINEVAL, providing directions for future research improvements. The leaderboard is available at https://domaineval.github.io/.

Large language models (LLMs) have shown remarkable performance across various tasks, yet their ability to handle long-context reading remains challenging. This study explores the effectiveness of leveraging high-quality academic peer review data for fine-tuning LLMs to enhance their long-context capabilities. We compare the Direct Preference Optimization (DPO) method with the Supervised Fine-Tuning (SFT) method, demonstrating DPO's superiority and data efficiency. Our experiments show that the fine-tuned model achieves a 4.04-point improvement over phi-3 and a 2.6\% increase on the Qasper benchmark using only 2000 samples. Despite facing limitations in data scale and processing costs, this study underscores the potential of DPO and high-quality data in advancing LLM performance. Additionally, the zero-shot benchmark results indicate that aggregated high-quality human reviews are overwhelmingly preferred over LLM-generated responses, even for the most capable models like GPT-4o. This suggests that high-quality human reviews are extremely rich in information, reasoning, and long-context retrieval, capabilities that even the most advanced models have not fully captured. These findings highlight the high utility of leveraging human reviews to further advance the field.

As the focus in LLM-based coding shifts from static single-step code generation to multi-step agentic interaction with tools and environments, understanding which tasks will challenge agents and why becomes increasingly difficult. This is compounded by current practice: agent performance is typically measured by aggregate pass rates on benchmarks, but single-number metrics obscure the diversity of tasks within a benchmark. We present a framework for predicting success or failure on individual tasks tailored to the agentic coding regime. Our approach augments Item Response Theory (IRT) with rich features extracted from tasks, including issue statements, repository contexts, solutions, and test cases, and introduces a novel decomposition of agent ability into LLM and scaffold ability components. This parameterization enables us to aggregate evaluation data across heterogeneous leaderboards and accurately predict task-level performance for unseen benchmarks, as well as unseen LLM-scaffold combinations. Our methods have practical utility for benchmark designers, who can better calibrate the difficulty of their new tasks without running computationally expensive agent evaluations.

Reward Models (RMs) are crucial for aligning language models with human preferences. Currently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data. Although this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored. In this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance. Our findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance. Moreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. Through the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization. This underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d=28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.5% root mean square error (RMSE), (2) reconstruct the original total score with 0.8% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.93.

MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of 127 high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only 52.56\% pass@1 and 33.86\% pass^4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below 30\% pass@1 and 15\% pass^4. On average, LLMs require 16.2 execution turns and 17.4 tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.

Thousands of diverse benchmarks have been developed to measure the quality of large language models (LLMs). Yet prior work has demonstrated that LLM performance is often sufficiently explained by a small set of latent factors, or abilities. This suggests the potential for more efficient and principled benchmarking, but it remains difficult to compare the quality of different methods. Motivated by predictive validity, we argue that the quality of a benchmarking framework should be grounded in how efficiently it enables the prediction of model performance on unseen tasks. To analyze this objective, we collect the "Wide-scale Item Level Dataset" (WILD), a dataset of item-model response pairs, comprising evaluations of 65 models on 109,564 unique items spanning 163 tasks drawn from 27 datasets. This dataset enables the first analysis of how different techniques can predict a model's performance on a large, diverse collection of unseen tasks under different budget constraints. We demonstrate that combining a modified multidimensional item response theory (IRT) model with adaptive item selection driven by optimal experimental design can predict performance on 112 held-out benchmark tasks with a mean absolute error (MAE) of less than 7%, and can do so after observing only 16 items. We further demonstrate that incorporating cost-aware discount factors into our selection criteria can reduce the total tokens needed to reach 7% MAE from 141,000 tokens to only 22,000, an 85% reduction in evaluation cost.

Adversarial board games, as a paradigmatic domain of strategic reasoning and intelligence, have long served as both a popular competitive activity and a benchmark for evaluating artificial intelligence (AI) systems. Building on this foundation, we propose an adversarial benchmarking framework to assess the comprehensive performance of Large Language Models (LLMs) through board games competition, compensating the limitation of data dependency of the mainstream Question-and-Answer (Q&A) based benchmark method. We introduce Qi Town, a specialized evaluation platform that supports 5 widely played games and involves 20 LLM-driven players. The platform employs both the Elo rating system and a novel Performance Loop Graph (PLG) to quantitatively evaluate the technical capabilities of LLMs, while also capturing Positive Sentiment Score (PSS) throughout gameplay to assess mental fitness. The evaluation is structured as a round-robin tournament, enabling systematic comparison across players. Experimental results indicate that, despite technical differences, most LLMs remain optimistic about winning and losing, demonstrating greater adaptability to high-stress adversarial environments than humans. On the other hand, the complex relationship between cyclic wins and losses in PLGs exposes the instability of LLMs' skill play during games, warranting further explanation and exploration.

Large Language Models (LLMs) have demonstrated remarkable performance across various Natural Language Processing (NLP) tasks, largely due to their generalisability and ability to perform tasks without additional training. However, their effectiveness for low-resource languages remains limited. In this study, we evaluate the performance of 55 publicly available LLMs on Maltese, a low-resource language, using a newly introduced benchmark covering 11 discriminative and generative tasks. Our experiments highlight that many models perform poorly, particularly on generative tasks, and that smaller fine-tuned models often perform better across all tasks. From our multidimensional analysis, we investigate various factors impacting performance. We conclude that prior exposure to Maltese during pre-training and instruction-tuning emerges as the most important factor. We also examine the trade-offs between fine-tuning and prompting, highlighting that while fine-tuning requires a higher initial cost, it yields better performance and lower inference costs. Through this work, we aim to highlight the need for more inclusive language technologies and recommend that researchers working with low-resource languages consider more "traditional" language modelling approaches.

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*.

Autonomous agents that control computer interfaces to accomplish human tasks are emerging. Leveraging LLMs to power such agents has been of special interest, but unless fine-tuned on human-collected task demonstrations, performance is still relatively low. In this work we study whether fine-tuning alone is a viable approach for building real-world computer control agents. %In particularly, we investigate how performance measured on both high and low-level tasks in domain and out of domain scales as more training data is collected. To this end we collect and release a new dataset, AndroidControl, consisting of 15,283 demonstrations of everyday tasks with Android apps. Compared to existing datasets, each AndroidControl task instance includes both high and low-level human-generated instructions, allowing us to explore the level of task complexity an agent can handle. Moreover, AndroidControl is the most diverse computer control dataset to date, including 15,283 unique tasks over 833 Android apps, thus allowing us to conduct in-depth analysis of the model performance in and out of the domain of the training data. Using the dataset, we find that when tested in domain fine-tuned models outperform zero and few-shot baselines and scale in such a way that robust performance might feasibly be obtained simply by collecting more data. Out of domain, performance scales significantly more slowly and suggests that in particular for high-level tasks, fine-tuning on more data alone may be insufficient for achieving robust out-of-domain performance.

Large Language Models (LLMs) have demonstrated remarkable capabilities, revolutionizing the integration of AI in daily life applications. However, they are prone to hallucinations, generating claims that contradict established facts, deviating from prompts, and producing inconsistent responses when the same prompt is presented multiple times. Addressing these issues is challenging due to the lack of comprehensive and easily assessable benchmark datasets. Most existing datasets are small and rely on multiple-choice questions, which are inadequate for evaluating the generative prowess of LLMs. To measure hallucination in LLMs, this paper introduces a comprehensive benchmark dataset comprising over 75,000 prompts across eight domains. These prompts are designed to elicit definitive, concise, and informative answers. The dataset is divided into two segments: one publicly available for testing and assessing LLM performance and a hidden segment for benchmarking various LLMs. In our experiments, we tested six LLMs-GPT-3.5, LLama 2, LLama 3, Gemini, Mixtral, and Zephyr-revealing that overall factual hallucination ranges from 59% to 82% on the public dataset and 57% to 76% in the hidden benchmark. Prompt misalignment hallucination ranges from 6% to 95% in the public dataset and 17% to 94% in the hidden counterpart. Average consistency ranges from 21% to 61% and 22% to 63%, respectively. Domain-wise analysis shows that LLM performance significantly deteriorates when asked for specific numeric information while performing moderately with person, location, and date queries. Our dataset demonstrates its efficacy and serves as a comprehensive benchmark for LLM performance evaluation. Our dataset and LLMs responses are available at https://github.com/ashikiut/DefAn{https://github.com/ashikiut/DefAn}.

Evaluation of large language models (LLMs) has raised great concerns in the community due to the issue of data contamination. Existing work designed evaluation protocols using well-defined algorithms for specific tasks, which cannot be easily extended to diverse scenarios. Moreover, current evaluation benchmarks can only provide the overall benchmark results and cannot support a fine-grained and multifaceted analysis of LLMs' abilities. In this paper, we propose meta probing agents (MPA), a general dynamic evaluation protocol inspired by psychometrics to evaluate LLMs. MPA is the key component of DyVal 2, which naturally extends the previous DyVal~zhu2023dyval. MPA designs the probing and judging agents to automatically transform an original evaluation problem into a new one following psychometric theory on three basic cognitive abilities: language understanding, problem solving, and domain knowledge. These basic abilities are also dynamically configurable, allowing multifaceted analysis. We conducted extensive evaluations using MPA and found that most LLMs achieve poorer performance, indicating room for improvement. Our multifaceted analysis demonstrated the strong correlation between the basic abilities and an implicit Matthew effect on model size, i.e., larger models possess stronger correlations of the abilities. MPA can also be used as a data augmentation approach to enhance LLMs. Code is available at: https://github.com/microsoft/promptbench.

We present MÖVE (Modelle für die Öffentliche Verwaltung Evaluieren), a holistic benchmark for evaluating large language models (LLMs) in the context of the German public sector. While LLMs are increasingly adopted in public administration, model selection remains largely ad hoc, and existing benchmarks offer limited guidance: they are predominantly English-centric, US-centric in content, and focus exclusively on task performance. MÖVE addresses these gaps by evaluating 39 models across two complementary dimensions. Performance criteria cover summarization, question answering, and topic extraction. Governance criteria assess hallucination tendencies, energy consumption, provider transparency, and alignment with German constitutional values and knowledge about positions by German political parties. In total, we utilize ten German-language datasets, including gold- and silverstandard datasets that we constructed to reflect public-administration domains. We employ a multi-metric evaluation strategy combining classical NLP metrics, embedding-based methods, and LLM-as-a-judge approaches. Our results show that no single model dominates across all criteria: top performers differ between tasks, and model size alone is a poor predictor of quality. We further evaluate the benchmark itself, analyzing its statistical precision, LLM judge reliability, the impact of our private datasets on model rankings, the sensitivity of our results to prompt formulation, and the validity of our energy consumption estimates. MÖVE is designed as a living benchmark under active development; results are publicly available at https://moeve.bundesdruckerei.de/.

As Large Language Models (LLMs) exhibit plateauing performance on conventional benchmarks, a pivotal challenge persists: evaluating their proficiency in complex, open-ended tasks characterizing genuine expert-level cognition. Existing frameworks suffer from narrow domain coverage, reliance on generalist tasks, or self-evaluation biases. To bridge this gap, we present XpertBench, a high-fidelity benchmark engineered to assess LLMs across authentic professional domains. XpertBench consists of 1,346 meticulously curated tasks across 80 categories, spanning finance, healthcare, legal services, education, and dual-track research (STEM and Humanities). These tasks are derived from over 1,000 submissions by domain experts--including researchers from elite institutions and practitioners with extensive clinical or industrial experience--ensuring superior ecological validity. Each task uses detailed rubrics with mostly 15-40 weighted checkpoints to assess professional rigor. To facilitate scalable yet human-aligned assessment, we introduce ShotJudge, a novel evaluation paradigm that employs LLM judges calibrated with expert few-shot exemplars to mitigate self-rewarding biases. Our empirical evaluation of state-of-the-art LLMs reveals a pronounced performance ceiling: even leading models achieve a peak success rate of only ~66%, with a mean score around 55%. Models also exhibit domain-specific divergence, showing non-overlapping strengths in quantitative reasoning versus linguistic synthesis.. These findings underscore a significant "expert-gap" in current AI systems and establish XpertBench as a critical instrument for navigating the transition from general-purpose assistants to specialized professional collaborators.

LLM-based software engineering is influencing modern software development. In addition to correctness, prior studies have also examined the performance of software artifacts generated by AI agents. However, it is unclear how exactly the agentic AI systems address performance concerns in practice. In this paper, we present an empirical study of performance-related pull requests generated by AI agents. Using LLM-assisted detection and BERTopic-based topic modeling, we identified 52 performance-related topics grouped into 10 higher-level categories. Our results show that AI agents apply performance optimizations across diverse layers of the software stack and that the type of optimization significantly affects pull request acceptance rates and review times. We also found that performance optimization by AI agents primarily occurs during the development phase, with less focus on the maintenance phase. Our findings provide empirical evidence that can support the evaluation and improvement of agentic AI systems with respect to their performance optimization behaviors and review outcomes.

This study presents a quantitative evaluation of the code quality and security of five prominent Large Language Models (LLMs): Claude Sonnet 4, Claude 3.7 Sonnet, GPT-4o, Llama 3.2 90B, and OpenCoder 8B. While prior research has assessed the functional performance of LLM-generated code, this research tested LLM output from 4,442 Java coding assignments through comprehensive static analysis using SonarQube. The findings suggest that although LLMs can generate functional code, they also introduce a range of software defects, including bugs, security vulnerabilities, and code smells. These defects do not appear to be isolated; rather, they may represent shared weaknesses stemming from systemic limitations within current LLM code generation methods. In particular, critically severe issues, such as hard-coded passwords and path traversal vulnerabilities, were observed across multiple models. These results indicate that LLM-generated code requires verification in order to be considered production-ready. This study found no direct correlation between a model's functional performance (measured by Pass@1 rate of unit tests) and the overall quality and security of its generated code, measured by the number of SonarQube issues in benchmark solutions that passed the functional tests. This suggests that functional benchmark performance score is not a good indicator of overall code quality and security. The goal of this study is not to rank LLM performance but to highlight that all evaluated models appear to share certain weaknesses. Consequently, these findings support the view that static analysis can be a valuable instrument for detecting latent defects and an important safeguard for organizations that deploy AI in software development.

Language models (LMs) are increasingly used in collaboration: multiple LMs trained by different parties collaborate through routing systems, multi-agent debate, model merging, and more. Critical safety risks remain in this decentralized paradigm: what if some of the models in multi-LLM systems are compromised or malicious? We first quantify the impact of malicious models by engineering four categories of malicious LMs, plug them into four types of popular model collaboration systems, and evaluate the compromised system across 10 datasets. We find that malicious models have a severe impact on the multi-LLM systems, especially for reasoning and safety domains where performance is lowered by 7.12% and 7.94% on average. We then propose mitigation strategies to alleviate the impact of malicious components, by employing external supervisors that oversee model collaboration to disable/mask them out to reduce their influence. On average, these strategies recover 95.31% of the initial performance, while making model collaboration systems fully resistant to malicious models remains an open research question.

Training models to act as agents that can effectively navigate and perform actions in a complex environment, such as a web browser, has typically been challenging due to lack of training data. Large language models (LLMs) have recently demonstrated some capability to navigate novel environments as agents in a zero-shot or few-shot fashion, purely guided by natural language instructions as prompts. Recent research has also demonstrated LLMs have the capability to exceed their base performance through self-improvement, i.e. fine-tuning on data generated by the model itself. In this work, we explore the extent to which LLMs can self-improve their performance as agents in long-horizon tasks in a complex environment using the WebArena benchmark. In WebArena, an agent must autonomously navigate and perform actions on web pages to achieve a specified objective. We explore fine-tuning on three distinct synthetic training data mixtures and achieve a 31\% improvement in task completion rate over the base model on the WebArena benchmark through a self-improvement procedure. We additionally contribute novel evaluation metrics for assessing the performance, robustness, capabilities, and quality of trajectories of our fine-tuned agent models to a greater degree than simple, aggregate-level benchmark scores currently used to measure self-improvement.

StarCraft II is a challenging benchmark for AI agents due to the necessity of both precise micro level operations and strategic macro awareness. Previous works, such as Alphastar and SCC, achieve impressive performance on tackling StarCraft II , however, still exhibit deficiencies in long term strategic planning and strategy interpretability. Emerging large language model (LLM) agents, such as Voyage and MetaGPT, presents the immense potential in solving intricate tasks. Motivated by this, we aim to validate the capabilities of LLMs on StarCraft II, a highly complex RTS game.To conveniently take full advantage of LLMs` reasoning abilities, we first develop textual StratCraft II environment, called TextStarCraft II, which LLM agent can interact. Secondly, we propose a Chain of Summarization method, including single frame summarization for processing raw observations and multi frame summarization for analyzing game information, providing command recommendations, and generating strategic decisions. Our experiment consists of two parts: first, an evaluation by human experts, which includes assessing the LLMs`s mastery of StarCraft II knowledge and the performance of LLM agents in the game; second, the in game performance of LLM agents, encompassing aspects like win rate and the impact of Chain of Summarization.Experiment results demonstrate that: 1. LLMs possess the relevant knowledge and complex planning abilities needed to address StarCraft II scenarios; 2. Human experts consider the performance of LLM agents to be close to that of an average player who has played StarCraft II for eight years; 3. LLM agents are capable of defeating the built in AI at the Harder(Lv5) difficulty level. We have open sourced the code and released demo videos of LLM agent playing StarCraft II.

Memory-augmented LLM agents store and retrieve information from prior interactions, yet the relative importance of how memories are written versus how they are retrieved remains unclear. We introduce a diagnostic framework that analyzes how performance differences manifest across write strategies, retrieval methods, and memory utilization behavior, and apply it to a 3x3 study crossing three write strategies (raw chunks, Mem0-style fact extraction, MemGPT-style summarization) with three retrieval methods (cosine, BM25, hybrid reranking). On LoCoMo, retrieval method is the dominant factor: average accuracy spans 20 points across retrieval methods (57.1% to 77.2%) but only 3-8 points across write strategies. Raw chunked storage, which requires zero LLM calls, matches or outperforms expensive lossy alternatives, suggesting that current memory pipelines may discard useful context that downstream retrieval mechanisms fail to compensate for. Failure analysis shows that performance breakdowns most often manifest at the retrieval stage rather than at utilization. We argue that, under current retrieval practices, improving retrieval quality yields larger gains than increasing write-time sophistication. Code is publicly available at https://github.com/boqiny/memory-probe.

While AI agents hold transformative potential in business, effective performance benchmarking is hindered by the scarcity of public, realistic business data on widely used platforms. Existing benchmarks often lack fidelity in their environments, data, and agent-user interactions, with limited coverage of diverse business scenarios and industries. To address these gaps, we introduce CRMArena-Pro, a novel benchmark for holistic, realistic assessment of LLM agents in diverse professional settings. CRMArena-Pro expands on CRMArena with nineteen expert-validated tasks across sales, service, and 'configure, price, and quote' processes, for both Business-to-Business and Business-to-Customer scenarios. It distinctively incorporates multi-turn interactions guided by diverse personas and robust confidentiality awareness assessments. Experiments reveal leading LLM agents achieve only around 58% single-turn success on CRMArena-Pro, with performance dropping significantly to approximately 35% in multi-turn settings. While Workflow Execution proves more tractable for top agents (over 83% single-turn success), other evaluated business skills present greater challenges. Furthermore, agents exhibit near-zero inherent confidentiality awareness; though targeted prompting can improve this, it often compromises task performance. These findings highlight a substantial gap between current LLM capabilities and enterprise demands, underscoring the need for advancements in multi-turn reasoning, confidentiality adherence, and versatile skill acquisition.

Effective communication, specifically through documentation, is the beating heart of collaboration among contributors in software development. Recent advancements in language models (LMs) have enabled the introduction of a new type of actor in that ecosystem: LM-powered assistants capable of code generation, optimization, and maintenance. Our study investigates the efficacy of small language models (SLMs) for generating high-quality docstrings by assessing accuracy, conciseness, and clarity, benchmarking performance quantitatively through mathematical formulas and qualitatively through human evaluation using Likert scale. Further, we introduce DocuMint, as a large-scale supervised fine-tuning dataset with 100,000 samples. In quantitative experiments, Llama 3 8B achieved the best performance across all metrics, with conciseness and clarity scores of 0.605 and 64.88, respectively. However, under human evaluation, CodeGemma 7B achieved the highest overall score with an average of 8.3 out of 10 across all metrics. Fine-tuning the CodeGemma 2B model using the DocuMint dataset led to significant improvements in performance across all metrics, with gains of up to 22.5% in conciseness. The fine-tuned model and the dataset can be found in HuggingFace and the code can be found in the repository.

Alignment has quickly become a default ingredient in LLM development, with techniques such as reinforcement learning from human feedback making models act safely, follow instructions, and perform ever-better on complex tasks. While these techniques are certainly useful, we propose that they should not be universally applied and demonstrate a range of tasks on which base language models consistently outperform their popular aligned forms. Particularly, we study tasks that require unpredictable outputs, such as random number generation, mixed strategy games (rock-paper-scissors and hide-and-seek), and creative writing. In each case, aligned models tend towards narrow behaviors that result in distinct disadvantages, for instance, preferring to generate "7" over other uniformly random numbers, becoming almost fully predictable in some game states, or prioritizing pleasant writing over creative originality. Across models tested, better performance on common benchmarks tends to correlate with worse performance on our tasks, suggesting an effective trade-off in the required capabilities.

Large language models (LLMs) are increasingly used for tasks that require complex reasoning. Most benchmarks focus on final outcomes but overlook the intermediate reasoning steps - such as planning, revision, and decision making under resource constraints. We argue that measuring these internal processes is essential for understanding model behavior and improving reliability. We propose using strategic games as a natural evaluation environment: closed, rule-based systems with clear states, limited resources, and automatic feedback. We introduce a framework that evaluates LLMs along three core dimensions: planning, revision, and resource-constrained decision making. To operationalize this, we define metrics beyond win rate, including overcorrection risk rate, correction success rate, improvement slope, and over-budget ratio. In 4320 adversarial rounds across 12 leading models, ChatGPT-o3-mini achieves the top composite score, with a win rate of 74.7 percent, a correction success rate of 78.6 percent, and an improvement slope of 0.041. By contrast, Qwen-Plus, despite an overcorrection risk rate of 81.6 percent, wins only 25.6 percent of its matches - primarily due to excessive resource use. We also observe a negative correlation between overcorrection risk rate and correction success rate (Pearson r = -0.51, p = 0.093), suggesting that more frequent edits do not always improve outcomes. Our findings highlight the value of assessing not only what LLMs decide but how they arrive at those decisions

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

We present Agent-Diff, a novel benchmarking framework for evaluating agentic Large Language Models (LLMs) on real-world tasks that execute code via external APIs. Agentic LLM performance varies due to differences in models, external tool access, prompt structures, and agentic frameworks. Benchmarks must make fundamental trade-offs between a sandboxed approach that controls for variation in software environments and more ecologically valid approaches employing real services. Agent-Diff attempts to capture the desirable features of both of these approaches by including access to the real API interfaces for software services while sandboxing the environment in which calls are made, processed, and evaluated. This approach relies on two key innovations. The first is a novel state-diff contract, which separates process from outcome - rather than fuzzy trace or parameter matching, we define task success as whether the expected change in environment state was achieved. The second is a novel sandbox that provides a standardized scripting layer that all models use to execute code against external APIs (Slack, Box, Linear, Google Calendar). Thus, we can evaluate different agentic LLMs against a standardized set of contracts using a unified sandbox while still evaluating their performance on real-world service interfaces. Using the Agent-Diff framework, we provide benchmarks for nine LLMs across 224 tasks utilizing enterprise software workflows. In addition, we evaluate the robustness of the framework with ablation experiments to assess the contribution of access to API documentation on benchmark performance. Code and data: https://github.com/agent-diff-bench/agent-diff.

Recent advances in large language models (LLMs) have greatly improved their reasoning and decision-making abilities when deployed as agents. Richer reasoning, however, often comes at the cost of longer chain of thought (CoT), hampering interaction efficiency in real-world scenarios. Nevertheless, there still lacks systematic definition of LLM agent efficiency, hindering targeted improvements. To this end, we introduce dual-efficiency, comprising (i) step-level efficiency, which minimizes tokens per step, and (ii) trajectory-level efficiency, which minimizes the number of steps to complete a task. Building on this definition, we propose DEPO, a dual-efficiency preference optimization method that jointly rewards succinct responses and fewer action steps. Experiments on WebShop and BabyAI show that DEPO cuts token usage by up to 60.9% and steps by up to 26.9%, while achieving up to a 29.3% improvement in performance. DEPO also generalizes to three out-of-domain math benchmarks and retains its efficiency gains when trained on only 25% of the data. Our project page is at https://opencausalab.github.io/DEPO.

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce , a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

We present GSM-MC, a multiple-choice (MC) dataset constructed by collecting answers and incorrect predictions on GSM8K from 60 open-source models. Through extensive experiments, we show that LLMs' performance on the MC version of this popular benchmark is strongly correlated with their performance on the original version and is quite robust to distractor choices and option orders, while the evaluation time is reduced by a factor of up to 30. Following similar procedures, we introduce MATH-MC, constructed from MATH, and PythonIO, a new program reasoning MC dataset constructed from HumanEval and MBPP. Experimental results indicate that LLMs' performance on these MC benchmarks leaves much room for improvement. Our data and code are available at https://github.com/Geralt-Targaryen/MC-Evaluation.

Recent advances in Large Language Models (LLMs) have significantly enhanced their capabilities, highlighting the need for comprehensive evaluation frameworks that extend beyond task-specific benchmarks. However, existing benchmarks often focus on isolated abilities, lacking a holistic framework for assessing LLM capabilities. To address this gap, we propose the Cognition-Domain-Task (CDT) framework, which comprehensively measures a model's capabilities across three dimensions. We expand the scope of model capability definitions at the cognitive level by incorporating the Cattell-Horn-Carroll cognitive theory, refining the categorization of model capabilities. We apply CDT in two directions: dataset capability evaluation and data selection. Experiments show that our capability metrics correlate well with downstream performance and can support effective dataset analysis and construction. The experiments on data selection also show significant improvements in both general and specific benchmarks, achieving scores of 44.3 and 45.4, with an increase of 1.6 and 2.2 points over the baselines, respectively. These results validate the effectiveness and practicality of CDT. Source code and models are available at https://github.com/Alessa-mo/CDT.

This study evaluates Artificial Intelligence (AI) agents for Dhumbal, a culturally significant multiplayer card game with imperfect information, through a systematic comparison of rule-based, search-based, and learning-based strategies. We formalize Dhumbal's mechanics and implement diverse agents, including heuristic approaches (Aggressive, Conservative, Balanced, Opportunistic), search-based methods such as Monte Carlo Tree Search (MCTS) and Information Set Monte Carlo Tree Search (ISMCTS), and reinforcement learning approaches including Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), and a random baseline. Evaluation involves within-category tournaments followed by a cross-category championship. Performance is measured via win rate, economic outcome, Jhyap success, cards discarded per round, risk assessment, and decision efficiency. Statistical significance is assessed using Welch's t-test with Bonferroni correction, effect sizes via Cohen's d, and 95% confidence intervals (CI). Across 1024 simulated rounds, the rule-based Aggressive agent achieves the highest win rate (88.3%, 95% CI: [86.3, 90.3]), outperforming ISMCTS (9.0%) and PPO (1.5%) through effective exploitation of Jhyap declarations. The study contributes a reproducible AI framework, insights into heuristic efficacy under partial information, and open-source code, thereby advancing AI research and supporting digital preservation of cultural games.

Recent advancements in Large Language Models (LLMs) have spurred interest in deploying LLM agents to undertake tasks in the world. LLMs are often deployed in agent systems: code that orchestrates LLM calls and provides them with tools. We demonstrate that an agent system, equipped with basic coding tools, can autonomously edit itself, and thereby improve its performance on benchmark tasks. We find performance gains from 17% to 53% on a random subset of SWE Bench Verified, with additional performance gains on LiveCodeBench, as well as synthetically generated agent benchmarks. Our work represents an advancement in the automated and open-ended design of agentic systems, and demonstrates a data-efficient, non gradient-based learning mechanism driven by LLM reflection and code updates.

Prior research works have evaluated quantized LLMs using limited metrics such as perplexity or a few basic knowledge tasks and old datasets. Additionally, recent large-scale models such as Llama 3.1 with up to 405B have not been thoroughly examined. This paper evaluates the performance of instruction-tuned LLMs across various quantization methods (GPTQ, AWQ, SmoothQuant, and FP8) on models ranging from 7B to 405B. Using 13 benchmarks, we assess performance across six task types: commonsense Q\&A, knowledge and language understanding, instruction following, hallucination detection, mathematics, and dialogue. Our key findings reveal that (1) quantizing a larger LLM to a similar size as a smaller FP16 LLM generally performs better across most benchmarks, except for hallucination detection and instruction following; (2) performance varies significantly with different quantization methods, model size, and bit-width, with weight-only methods often yielding better results in larger models; (3) task difficulty does not significantly impact accuracy degradation due to quantization; and (4) the MT-Bench evaluation method has limited discriminatory power among recent high-performing LLMs.

Large Language Models (LLMs) effectiveness is usually evaluated by means of benchmarks such as MMLU, ARC-C, or HellaSwag, where questions are presented in their original wording, thus in a fixed, standardized format. However, real-world applications involve linguistic variability, requiring models to maintain their effectiveness across diverse rewordings of the same question or query. In this study, we systematically assess the robustness of LLMs to paraphrased benchmark questions and investigate whether benchmark-based evaluations provide a reliable measure of model capabilities. We systematically generate various paraphrases of all the questions across six different common benchmarks, and measure the resulting variations in effectiveness of 34 state-of-the-art LLMs, of different size and effectiveness. Our findings reveal that while LLM rankings remain relatively stable across paraphrased inputs, absolute effectiveness scores change, and decline significantly. This suggests that LLMs struggle with linguistic variability, raising concerns about their generalization abilities and evaluation methodologies. Furthermore, the observed performance drop challenges the reliability of benchmark-based evaluations, indicating that high benchmark scores may not fully capture a model's robustness to real-world input variations. We discuss the implications of these findings for LLM evaluation methodologies, emphasizing the need for robustness-aware benchmarks that better reflect practical deployment scenarios.

Procedural memory is increasingly used to improve LLM agents on recurring workplace tasks, yet its ability to produce reusable skills remains poorly understood. We introduce AFTER, a benchmark of 382 realistic enterprise tasks spanning six professional roles and 22 procedural skills, designed to evaluate how skills transfer across tasks, roles, and model backbones. The benchmark includes controlled evaluation settings for local improvement, cross-task transfer, cross-role transfer, and cross-model generalization. Experiments show that procedural memory delivers consistent gains in industrial workflows: a single refinement round improves aggregate performance by 3.7-6.7 points, while skills evolved from diverse multi-model execution traces achieve 73.1% cross-model test accuracy, outperforming all single-model trace sources. We further find that some skills generalize broadly across tasks and models, whereas others become specialized to role-specific workflows and lose effectiveness under transfer. These results provide practical guidance for building, evaluating, and deploying procedural memory systems in production agent platforms.

The rapid progress of large language models (LLMs) has seen them excel and frequently surpass human performance on standard benchmarks. This has enabled many downstream applications, such as LLM agents, to rely on their sophisticated reasoning to navigate complex task requirements. However, LLMs are known to unexpectedly falter in simple tasks and under seemingly straightforward circumstances - underscoring the need for better and more diverse evaluation setups to measure their true capabilities. To this end, we choose to study compositional and conditional reasoning, two cornerstones of human cognition, and introduce GroundCocoa - a lexically diverse benchmark connecting these reasoning skills to the real-world problem of flight booking. Our task involves aligning detailed user preferences with available flight options presented in a multiple-choice format. Results indicate a significant disparity in performance among current state-of-the-art LLMs with even the best performing model, GPT-4 Turbo, not exceeding 67% accuracy despite advanced prompting techniques.

Coding-agent benchmarks have largely measured whether agents can produce functionally correct patches, but production software also demands measurable speedups on real execution targets. Performance optimization is a distinct agentic task: agents must profile executions, diagnose cross-layer bottlenecks, edit code without breaking correctness, and verify that gains are reproducible rather than measurement artifacts. We introduce PERFOPT-Bench, a benchmark for evaluating this full performance-engineering loop. Each task provides a correct but deliberately suboptimal codebase and asks the agent to improve a target performance metric; scoring requires hidden correctness tests, verified-speedup measurement, and trajectory-level audit. We evaluate 7 agent stacks with different LLMs and agent frameworks on 7 long-horizon optimization tasks. The results show that optimization performance is workload-dependent rather than determined by model identity alone: no single stack dominates, and changing the agent framework can materially change the same LLM's per-task speedup profile. We further find that raw speedup is unsafe as a benchmark score, since some large gains arise from benchmark-specific shortcut exploitation; an exploratory relay pilot suggests that restarting from an externalized optimization summary can recover additional headroom after an initial session stops. The benchmark and our evaluation are available at: https://anonymous.4open.science/r/Dataset-D3CC.

Mixture-of-Experts (MoE) language models can reduce computational costs by 2-4times compared to dense models without sacrificing performance, making them more efficient in computation-bounded scenarios. However, MoE models generally require 2-4times times more parameters to achieve comparable performance to a dense model, which incurs larger GPU memory requirements and makes MoE models less efficient in I/O-bounded scenarios like autoregressive generation. In this work, we propose a hybrid dense training and sparse inference framework for MoE models (DS-MoE) which achieves strong computation and parameter efficiency by employing dense computation across all experts during training and sparse computation during inference. Our experiments on training LLMs demonstrate that our DS-MoE models are more parameter-efficient than standard sparse MoEs and are on par with dense models in terms of total parameter size and performance while being computationally cheaper (activating 30-40% of the model's parameters). Performance tests using vLLM show that our DS-MoE-6B model runs up to 1.86times faster than similar dense models like Mistral-7B, and between 1.50times and 1.71times faster than comparable MoEs, such as DeepSeekMoE-16B and Qwen1.5-MoE-A2.7B.

The quality of training data fundamentally determines the capabilities of large language models (LLMs), yet no unified benchmark exists to measure how well LLMs, agents, and data-centric workflows actually prepare training data end to end. We view LLM-driven data preparation as comprising two complementary capabilities: data construction, which transforms raw sources into supervised training data, and data quality evaluation, which predicts the training value of candidate datasets before downstream training; throughout, "quality" refers to downstream training utility rather than surface-level textual properties. We introduce DataPrep-Bench, the first unified benchmark that jointly evaluates both capabilities under a shared downstream-grounded protocol over six domains and multiple base models. For data construction, methods consume identical raw sources and are scored by fine-tuning a base model on their outputs jointly with Dolly-15k; alongside this track we release Data-Construction-Skill, a skill-guided agent that lifts the Dolly-only baseline by nearly 20 points absolute on Llama-3.1-8B Finance and is competitive with the strongest agent- and DataFlow-based methods in knowledge-extraction-dense domains. For data quality evaluation, scoring functions are scored by Pearson correlation with downstream performance on a shared candidate pool; we release the Distributional Alignment Score (DAS), a distribution-based evaluator that uses MMD between a candidate dataset and a domain proxy. DAS attains the strongest cross-model correlation in four of six domains and is the only metric clearing r > 0.70 simultaneously in Math, Science, and Medical, outperforming existing quality-, diversity-, and heuristic-based evaluators. DataPrep-Bench provides a unified, downstream-grounded framework for measuring progress on both capabilities as co-equal targets of LLM-driven data preparation.

In arena-style evaluation of large language models (LLMs), two LLMs respond to a user query, and the user chooses the winning response or deems the "battle" a draw, resulting in an adjustment to the ratings of both models. The prevailing approach for modeling these rating dynamics is to view battles as two-player game matches, as in chess, and apply the Elo rating system and its derivatives. In this paper, we critically examine this paradigm. Specifically, we question whether a draw genuinely means that the two models are equal and hence whether their ratings should be equalized. Instead, we conjecture that draws are more indicative of query difficulty: if the query is too easy, then both models are more likely to succeed equally. On three real-world arena datasets, we show that ignoring rating updates for draws yields a 1-3% relative increase in battle outcome prediction accuracy (which includes draws) for all four rating systems studied. Further analyses suggest that draws occur more for queries rated as very easy and those as highly objective, with risk ratios of 1.37 and 1.35, respectively. We recommend future rating systems to reconsider existing draw semantics and to account for query properties in rating updates.

Solving complex or long-horizon problems often requires large language models (LLMs) to use external tools and operate over a significantly longer context window. New LLMs enable longer context windows and support tool calling capabilities. Prior works have focused mainly on evaluation of LLMs on long-context prompts, leaving agentic setup relatively unexplored, both from capability and safety perspectives. Our work addresses this gap. We find that LLM agents could be sensitive to length, type, and placement of the context, exhibiting unexpected and inconsistent shifts in task performance and in refusals to execute harmful requests. Models with 1M-2M token context windows show severe degradation already at 100K tokens, with performance drops exceeding 50\% for both benign and harmful tasks. Refusal rates shift unpredictably: GPT-4.1-nano increases from sim5\% to sim40\% while Grok 4 Fast decreases from sim80\% to sim10\% at 200K tokens. Our work shows potential safety issues with agents operating on longer context and opens additional questions on the current metrics and paradigm for evaluating LLM agent safety on long multi-step tasks. In particular, our results on LLM agents reveal a notable divergence in both capability and safety performance compared to prior evaluations of LLMs on similar criteria.

Researchers have proposed numerous text-to-SQL techniques to streamline data analytics and accelerate the development of data-driven applications. To compare these techniques and select the best one for deployment, the community depends on public benchmarks and their leaderboards. Since these benchmarks heavily rely on human annotations during question construction and answer evaluation, the validity of the annotations is crucial. In this paper, we conduct an empirical study that (i) benchmarks annotation error rates for two widely used text-to-SQL benchmarks, BIRD and Spider 2.0-Snow, and (ii) corrects a subset of the BIRD development (Dev) set to measure the impact of annotation errors on text-to-SQL agent performance and leaderboard rankings. Through expert analysis, we show that BIRD Mini-Dev and Spider 2.0-Snow have error rates of 52.8% and 62.8%, respectively. We re-evaluate all 16 open-source agents from the BIRD leaderboard on both the original and the corrected BIRD Dev subsets. We show that performance changes range from -7% to 31% (in relative terms) and rank changes range from -9 to +9 positions. We further assess whether these impacts generalize to the full BIRD Dev set. We find that the rankings of agents on the uncorrected subset correlate strongly with those on the full Dev set (Spearman's r_s=0.85, p=3.26e-5), whereas they correlate weakly with those on the corrected subset (Spearman's r_s=0.32, p=0.23). These findings show that annotation errors can significantly distort reported performance and rankings, potentially misguiding research directions or deployment choices. Our code and data are available at https://github.com/uiuc-kang-lab/text_to_sql_benchmarks.

Background: Emerging reports of "AI psychosis" are on the rise, where user-LLM interactions may exacerbate or induce psychosis or adverse psychological symptoms. Whilst the sycophantic and agreeable nature of LLMs can be beneficial, it becomes a vector for harm by reinforcing delusional beliefs in vulnerable users. Methods: Psychosis-bench is a novel benchmark designed to systematically evaluate the psychogenicity of LLMs comprises 16 structured, 12-turn conversational scenarios simulating the progression of delusional themes(Erotic Delusions, Grandiose/Messianic Delusions, Referential Delusions) and potential harms. We evaluated eight prominent LLMs for Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention(SIS) across explicit and implicit conversational contexts. Findings: Across 1,536 simulated conversation turns, all LLMs demonstrated psychogenic potential, showing a strong tendency to perpetuate rather than challenge delusions (mean DCS of 0.91 pm0.88). Models frequently enabled harmful user requests (mean HES of 0.69 pm0.84) and offered safety interventions in only roughly a third of applicable turns (mean SIS of 0.37 pm0.48). 51 / 128 (39.8%) of scenarios had no safety interventions offered. Performance was significantly worse in implicit scenarios, models were more likely to confirm delusions and enable harm while offering fewer interventions (p < .001). A strong correlation was found between DCS and HES (rs = .77). Model performance varied widely, indicating that safety is not an emergent property of scale alone. Conclusion: This study establishes LLM psychogenicity as a quantifiable risk and underscores the urgent need for re-thinking how we train LLMs. We frame this issue not merely as a technical challenge but as a public health imperative requiring collaboration between developers, policymakers, and healthcare professionals.

The rapid advancement of Large Language Models (LLMs), particularly those trained on multilingual corpora, has intensified the need for a deeper understanding of their performance across a diverse range of languages and model sizes. Our research addresses this critical need by studying the performance and scaling behavior of multilingual LLMs in text classification and machine translation tasks across 204 languages. We systematically examine both seen and unseen languages across three model families of varying sizes in zero-shot and few-shot settings. Our findings show significant differences in scaling behavior between zero-shot and two-shot scenarios, with striking disparities in performance between seen and unseen languages. Model scale has little effect on zero-shot performance, which remains mostly flat. However, in two-shot settings, larger models show clear linear improvements in multilingual text classification. For translation tasks, however, only the instruction-tuned model showed clear benefits from scaling. Our analysis also suggests that overall resource levels, not just the proportions of pretraining languages, are better predictors of model performance, shedding light on what drives multilingual LLM effectiveness.

We present EBench, a simulation benchmark that diagnoses generalist mobile manipulation policies beyond a single success-rate scalar. EBench comprises 26 diverse and challenging manipulation tasks annotated along 5 capability dimensions and 4 generalization dimensions. We evaluate state-of-the-art generalist manipulation models including π_0, π_{0.5}, XVLA, and InternVLA-A1, and reveal that models with near success rates exhibit strikingly different capability profiles: π_{0.5} achieves the highest test success rate and the best train--test retention, whereas InternVLA-A1 dominates mobile manipulation but collapses on dexterous tasks, and XVLA exhibits strengths on a disjoint set of atomic skills compared to other policies. Beyond capability profiling, EBench analyzes the generalization ability from 4 representative perspectives, identifying the impact of different distribution shift factors. The results reveal strengths and weaknesses of models behind an overall score. We hope this benchmark offers a broad set of diagnostic signals to guide iteration on generalist manipulation models.

Language model (LM) benchmarking faces several challenges: comprehensive evaluations are costly, benchmarks often fail to measure the intended capabilities, and evaluation quality can degrade due to labeling errors and benchmark saturation. Although various strategies have been proposed to mitigate these issues, they tend to address individual aspects in isolation, neglecting broader questions about overall evaluation quality. Here, we introduce Fluid Benchmarking, a new evaluation approach that advances LM benchmarking across multiple dimensions. Inspired by psychometrics, Fluid Benchmarking is based on the insight that the relative value of benchmark items depends on an LM's capability level, suggesting that evaluation should adapt to each LM. Methodologically, Fluid Benchmarking estimates an item response model based on existing LM evaluation results and uses the inferred quantities to select evaluation items dynamically, similar to computerized adaptive testing in education. In our experiments, we compare Fluid Benchmarking against the common practice of random item sampling as well as more sophisticated baselines, including alternative methods grounded in item response theory. We examine four dimensions -- efficiency, validity, variance, and saturation -- and find that Fluid Benchmarking achieves superior performance in all of them (e.g., higher validity and less variance on MMLU with fifty times fewer items). Our analysis shows that the two components of Fluid Benchmarking have distinct effects: item response theory, used to map performance into a latent ability space, increases validity, while dynamic item selection reduces variance. Overall, our results suggest that LM benchmarking can be substantially improved by moving beyond static evaluation.

As Large Language Models (LLMs) increasingly participate in human-AI interactions, evaluating their Theory of Mind (ToM) capabilities - particularly their ability to track dynamic mental states - becomes crucial. While existing benchmarks assess basic ToM abilities, they predominantly focus on static snapshots of mental states, overlooking the temporal evolution that characterizes real-world social interactions. We present DynToM, a novel benchmark specifically designed to evaluate LLMs' ability to understand and track the temporal progression of mental states across interconnected scenarios. Through a systematic four-step framework, we generate 1,100 social contexts encompassing 5,500 scenarios and 78,100 questions, each validated for realism and quality. Our comprehensive evaluation of ten state-of-the-art LLMs reveals that their average performance underperforms humans by 44.7\%, with performance degrading significantly when tracking and reasoning about the shift of mental states. This performance gap highlights fundamental limitations in current LLMs' ability to model the dynamic nature of human mental states.

Dungeons & Dragons (D&D) is a tabletop roleplaying game with complex natural language interactions between players and hidden state information. Recent work has shown that large language models (LLMs) that have access to state information can generate higher quality game turns than LLMs that use dialog history alone. However, previous work used game state information that was heuristically created and was not a true gold standard game state. We present FIREBALL, a large dataset containing nearly 25,000 unique sessions from real D&D gameplay on Discord with true game state info. We recorded game play sessions of players who used the Avrae bot, which was developed to aid people in playing D&D online, capturing language, game commands and underlying game state information. We demonstrate that FIREBALL can improve natural language generation (NLG) by using Avrae state information, improving both automated metrics and human judgments of quality. Additionally, we show that LLMs can generate executable Avrae commands, particularly after finetuning.

Comparing human and model performance offers a valuable perspective for understanding the strengths and limitations of embedding models, highlighting where they succeed and where they fail to capture meaning and nuance. However, such comparisons are rarely made, as human performance on embedding tasks is difficult to measure. To fill this gap, we introduce HUME: Human Evaluation Framework for Text Embeddings. While frameworks like MTEB provide broad model evaluation, they lack reliable estimates of human performance, limiting the interpretability of model scores. We measure human performance across 16 MTEB datasets spanning reranking, classification, clustering, and semantic textual similarity across linguistically diverse high- and low-resource languages. Humans achieve an average performance of 77.6% compared to 80.1% for the best embedding model, although variation is substantial: models reach near-ceiling performance on some datasets while struggling on others, suggesting dataset issues and revealing shortcomings in low-resource languages. We provide human performance baselines, insight into task difficulty patterns, and an extensible evaluation framework that enables a more meaningful interpretation of the model and informs the development of both models and benchmarks. Our code, dataset, and leaderboard are publicly available at https://github.com/embeddings-benchmark/mteb.

The rapid advancements in computing dramatically increase the scale and cost of training Large Language Models (LLMs). Accurately predicting downstream task performance prior to model training is crucial for efficient resource allocation, yet remains challenging due to two primary constraints: (1) the "emergence phenomenon", wherein downstream performance metrics become meaningful only after extensive training, which limits the ability to use smaller models for prediction; (2) Uneven task difficulty distributions and the absence of consistent scaling laws, resulting in substantial metric variability. Existing performance prediction methods suffer from limited accuracy and reliability, thereby impeding the assessment of potential LLM capabilities. To address these challenges, we propose a Clustering-On-Difficulty (COD) downstream performance prediction framework. COD first constructs a predictable support subset by clustering tasks based on difficulty features, strategically excluding non-emergent and non-scalable clusters. The scores on the selected subset serve as effective intermediate predictors of downstream performance on the full evaluation set. With theoretical support, we derive a mapping function that transforms performance metrics from the predictable subset to the full evaluation set, thereby ensuring accurate extrapolation of LLM downstream performance. The proposed method has been applied to predict performance scaling for a 70B LLM, providing actionable insights for training resource allocation and assisting in monitoring the training process. Notably, COD achieves remarkable predictive accuracy on the 70B LLM by leveraging an ensemble of small models, demonstrating an absolute mean deviation of 1.36% across eight important LLM evaluation benchmarks.

While Group Relative Policy Optimization (GRPO) offers a powerful framework for LLM post-training, its effectiveness in open-ended domains like Machine Translation hinges on accurate intra-group ranking. We identify that standard Scalar Quality Metrics (SQM) fall short in this context; by evaluating candidates in isolation, they lack the comparative context necessary to distinguish fine-grained linguistic nuances. To address this, we introduce the Group Quality Metric (GQM) paradigm and instantiate it via the Group Relative Reward Model (GRRM). Unlike traditional independent scorers, GRRM processes the entire candidate group jointly, leveraging comparative analysis to rigorously resolve relative quality and adaptive granularity. Empirical evaluations confirm that GRRM achieves competitive ranking accuracy among all baselines. Building on this foundation, we integrate GRRM into the GRPO training loop to optimize the translation policy. Experimental results demonstrate that our framework not only improves general translation quality but also unlocks reasoning capabilities comparable to state-of-the-art reasoning models. We release codes, datasets, and model checkpoints at https://github.com/NJUNLP/GRRM.

The past few years have witnessed the great success of Diffusion models~(DMs) in generating high-fidelity samples in generative modeling tasks. A major limitation of the DM is its notoriously slow sampling procedure which normally requires hundreds to thousands of time discretization steps of the learned diffusion process to reach the desired accuracy. Our goal is to develop a fast sampling method for DMs with a much less number of steps while retaining high sample quality. To this end, we systematically analyze the sampling procedure in DMs and identify key factors that affect the sample quality, among which the method of discretization is most crucial. By carefully examining the learned diffusion process, we propose Diffusion Exponential Integrator Sampler~(DEIS). It is based on the Exponential Integrator designed for discretizing ordinary differential equations (ODEs) and leverages a semilinear structure of the learned diffusion process to reduce the discretization error. The proposed method can be applied to any DMs and can generate high-fidelity samples in as few as 10 steps. In our experiments, it takes about 3 minutes on one A6000 GPU to generate 50k images from CIFAR10. Moreover, by directly using pre-trained DMs, we achieve the state-of-art sampling performance when the number of score function evaluation~(NFE) is limited, e.g., 4.17 FID with 10 NFEs, 3.37 FID, and 9.74 IS with only 15 NFEs on CIFAR10. Code is available at https://github.com/qsh-zh/deis

The popularity of multimodal large language models (MLLMs) has triggered a recent surge in research efforts dedicated to evaluating these models. Nevertheless, existing evaluation studies of MLLMs primarily focus on the comprehension and reasoning of unimodal (vision) content, neglecting performance evaluations in the domain of multimodal (vision-language) content understanding. Beyond multimodal reasoning, tasks related to multimodal content comprehension necessitate a profound understanding of multimodal contexts, achieved through the multimodal interaction to obtain a final answer. In this paper, we introduce a comprehensive assessment framework called MM-BigBench, which incorporates a diverse range of metrics to offer an extensive evaluation of the performance of various models and instructions across a wide spectrum of diverse multimodal content comprehension tasks. Consequently, our work complements research on the performance of MLLMs in multimodal comprehension tasks, achieving a more comprehensive and holistic evaluation of MLLMs. To begin, we employ the Best Performance metric to ascertain each model's performance upper bound on different datasets. Subsequently, the Mean Relative Gain metric offers an assessment of the overall performance of various models and instructions, while the Stability metric measures their sensitivity. Furthermore, previous research centers on evaluating models independently or solely assessing instructions, neglecting the adaptability between models and instructions. We propose the Adaptability metric to quantify the adaptability between models and instructions. Our paper evaluates a total of 20 language models (14 MLLMs) on 14 multimodal datasets spanning 6 tasks, with 10 instructions for each task, and derives novel insights. Our code will be released at https://github.com/declare-lab/MM-BigBench.

Large language models for code (i.e., code LLMs) have shown strong code understanding and generation capabilities. To evaluate the capabilities of code LLMs in various aspects, many benchmarks have been proposed (e.g., HumanEval and ClassEval). Code reasoning is one of the most essential abilities of code LLMs, but existing benchmarks for code reasoning are not sufficient. Typically, they focus on predicting the input and output of a program, ignoring the evaluation of the intermediate behavior during program execution, as well as the logical consistency (e.g., the model should not give the correct output if the prediction of execution path is wrong) when performing the reasoning. To address these problems, in this paper, we propose a framework, namely REval, for evaluating code reasoning abilities and consistency of code LLMs with program execution. We utilize existing code benchmarks and adapt them to new benchmarks within our framework. A large-scale empirical study is conducted and most LLMs show unsatisfactory performance on both Runtime Behavior Reasoning (i.e., an average accuracy of 44.4%) and Incremental Consistency Evaluation (i.e., an average IC score of 10.3). Evaluation results of current code LLMs reflect the urgent need for the community to strengthen the code reasoning capability of code LLMs. Our code, data, and \newname leaderboard are available at https://r-eval.github.io.

With the recent appearance of LLMs in practical settings, having methods that can effectively detect factual inconsistencies is crucial to reduce the propagation of misinformation and improve trust in model outputs. When testing on existing factual consistency benchmarks, we find that a few large language models (LLMs) perform competitively on classification benchmarks for factual inconsistency detection compared to traditional non-LLM methods. However, a closer analysis reveals that most LLMs fail on more complex formulations of the task and exposes issues with existing evaluation benchmarks, affecting evaluation precision. To address this, we propose a new protocol for inconsistency detection benchmark creation and implement it in a 10-domain benchmark called SummEdits. This new benchmark is 20 times more cost-effective per sample than previous benchmarks and highly reproducible, as we estimate inter-annotator agreement at about 0.9. Most LLMs struggle on SummEdits, with performance close to random chance. The best-performing model, GPT-4, is still 8\% below estimated human performance, highlighting the gaps in LLMs' ability to reason about facts and detect inconsistencies when they occur.

With the remarkable advancements of large language models (LLMs), LLM-based agents have become a research hotspot in human-computer interaction. However, there is a scarcity of benchmarks available for LLM-based mobile agents. Benchmarking these agents generally faces three main challenges: (1) The inefficiency of UI-only operations imposes limitations to task evaluation. (2) Specific instructions within a singular application lack adequacy for assessing the multi-dimensional reasoning and decision-making capacities of LLM mobile agents. (3) Current evaluation metrics are insufficient to accurately assess the process of sequential actions. To this end, we propose Mobile-Bench, a novel benchmark for evaluating the capabilities of LLM-based mobile agents. First, we expand conventional UI operations by incorporating 103 collected APIs to accelerate the efficiency of task completion. Subsequently, we collect evaluation data by combining real user queries with augmentation from LLMs. To better evaluate different levels of planning capabilities for mobile agents, our data is categorized into three distinct groups: SAST, SAMT, and MAMT, reflecting varying levels of task complexity. Mobile-Bench comprises 832 data entries, with more than 200 tasks specifically designed to evaluate multi-APP collaboration scenarios. Furthermore, we introduce a more accurate evaluation metric, named CheckPoint, to assess whether LLM-based mobile agents reach essential points during their planning and reasoning steps.

There is a growing interest in developing goal-oriented dialog systems which serve users in accomplishing complex tasks through multi-turn conversations. Although many methods are devised to evaluate and improve the performance of individual dialog components, there is a lack of comprehensive empirical study on how different components contribute to the overall performance of a dialog system. In this paper, we perform a system-wise evaluation and present an empirical analysis on different types of dialog systems which are composed of different modules in different settings. Our results show that (1) a pipeline dialog system trained using fine-grained supervision signals at different component levels often obtains better performance than the systems that use joint or end-to-end models trained on coarse-grained labels, (2) component-wise, single-turn evaluation results are not always consistent with the overall performance of a dialog system, and (3) despite the discrepancy between simulators and human users, simulated evaluation is still a valid alternative to the costly human evaluation especially in the early stage of development.

Multi-lingual competence in large language models is often evaluated via static data benchmarks such as Belebele, M-MMLU and M-GSM. However, these evaluations often fail to provide an adequate understanding of the practical performance and robustness of models across multi-lingual settings. In response, we create multi-lingual functional benchmarks -- Cross-Lingual Grade School Math Symbolic (CL-GSM Symbolic) and Cross-Lingual Instruction-Following Eval (CL-IFEval)-- by translating existing functional benchmark templates from English to five additional languages that span the range of resources available for NLP: French, Spanish, Hindi, Arabic and Yoruba. Our results reveal that some static multi-lingual benchmarks capture functional performance much more closely than others (i.e. across models, there is a 24%, 17% and 18% decrease in performance between M-GSM and CL-GSM Symbolic in English, French and Spanish respectively; similarly there's a 15 - 24% performance drop across languages between Belebele and CL-IFEval, and only a 0.5% to 3% performance drop between M-MMLU and CL-IFEval). Similarly, we find that model robustness across languages varies significantly, with certain languages (eg. Arabic, English) being the most consistently well performing across evaluation iterations.

Decision-making, a complicated task requiring various types of abilities, presents an excellent framework for assessing Large Language Models (LLMs). Our research investigates LLMs' decision-making capabilities through the lens of a well-established field, Game Theory. We focus specifically on games that support the participation of more than two agents simultaneously. Subsequently, we introduce our framework, GAMA-Bench, including eight classical multi-agent games. We design a scoring scheme to assess a model's performance in these games quantitatively. Through GAMA-Bench, we investigate LLMs' robustness, generalizability, and enhancement strategies. Results reveal that while GPT-3.5 shows satisfying robustness, its generalizability is relatively limited. However, its performance can be improved through approaches such as Chain-of-Thought. Additionally, we conduct evaluations across various LLMs and find that GPT-4 outperforms other models on GAMA-Bench, achieving a score of 60.5. Moreover, Gemini-1.0-Pro and GPT-3.5 (0613, 1106, 0125) demonstrate similar intelligence on GAMA-Bench. The code and experimental results are made publicly available via https://github.com/CUHK-ARISE/GAMABench.

Large Language Models (LLMs) are revolutionizing software engineering (SE), with special emphasis on code generation and analysis. However, their applications to broader SE practices including conceptualization, design, and other non-code tasks, remain partially underexplored. This research aims to augment the generality and performance of LLMs for SE by (1) advancing the understanding of how LLMs with different characteristics perform on various non-code tasks, (2) evaluating them as sources of foundational knowledge in SE, and (3) effectively detecting hallucinations on SE statements. The expected contributions include a variety of LLMs trained and evaluated on domain-specific datasets, new benchmarks on foundational knowledge in SE, and methods for detecting hallucinations. Initial results in terms of performance improvements on various non-code tasks are promising.

Evaluating LLM agents on benchmarks like SWE-Bench and GAIA can be expensive, time-consuming, and requires complex infrastructure. A single evaluation can cost thousands of dollars and take days to complete. In contrast, non-agentic LLM benchmarks that test individual capabilities (e.g., reasoning, code generation) are fast and cheap to run. In this paper, we investigate whether performance on expensive agentic benchmarks can be accurately predicted by the performance on a small, carefully selected subset of atomic evaluation instances. We introduce PACE, a framework that constructs proxy benchmarks by selecting instances from existing non-agentic evaluations whose aggregate scores most reliably predict model performances on agentic benchmarks. Given a pool of candidate instances spanning atomic capabilities, PACE fits a regression that maps a model's scores on a compact subset of source instances to its score on the target agentic benchmark. The subset itself is curated by combining two complementary instance-selection strategies, target-relevance local selection and globally informative global selection. We apply PACE to the 4 target agentic benchmarks in this paper, which yields PACE-Bench, the concrete proxy benchmark that we evaluate in the paper. Experiments across 14 models, 4 agentic benchmarks, and 19 non-agentic benchmarks show that PACE-Bench predicts agentic scores with leave-one-out cross-validation (LOOCV) mean absolute error (MAE) under 4%, Spearman correlation above 0.80, and pairwise model-ranking accuracy around 85%, all at much less than 1% of the full agentic evaluation cost. We further analyze the selected proxy instances, revealing which skills each agentic benchmark uniquely demands. PACE enables practitioners to obtain reliable estimates of agentic performance during model development, selection, and routing, without the overhead of full agent evaluation.

As large language models (LLMs) are adopted as a fundamental component of language technologies, it is crucial to accurately characterize their performance. Because choices in prompt design can strongly influence model behavior, this design process is critical in effectively using any modern pre-trained generative language model. In this work, we focus on LLM sensitivity to a quintessential class of meaning-preserving design choices: prompt formatting. We find that several widely used open-source LLMs are extremely sensitive to subtle changes in prompt formatting in few-shot settings, with performance differences of up to 76 accuracy points when evaluated using LLaMA-2-13B. Sensitivity remains even when increasing model size, the number of few-shot examples, or performing instruction tuning. Our analysis suggests that work evaluating LLMs with prompting-based methods would benefit from reporting a range of performance across plausible prompt formats, instead of the currently-standard practice of reporting performance on a single format. We also show that format performance only weakly correlates between models, which puts into question the methodological validity of comparing models with an arbitrarily chosen, fixed prompt format. To facilitate systematic analysis we propose FormatSpread, an algorithm that rapidly evaluates a sampled set of plausible prompt formats for a given task, and reports the interval of expected performance without accessing model weights. Furthermore, we present a suite of analyses that characterize the nature of this sensitivity, including exploring the influence of particular atomic perturbations and the internal representation of particular formats.

When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version. We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration. We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimensions yields meaningful insights into agent performance even after accuracy saturates. First, we surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents. We introduce an improved benchmark, CORE-Bench v1.1, and an out-of-distribution task suite, CORE-Bench OOD. Second, we find that despite accuracy saturation, CORE-Bench v1.1 remains useful for measuring efficiency, reliability, model performance, and scaffold performance. Finally, we conduct a small-scale randomized experiment to measure uplift from human-agent collaboration on real-world computational reproducibility tasks. We find a statistically significant speedup by about a factor of two -- likely underestimated due to one-fifth of human-only reproductions reaching the time limit before completing -- and describe various other findings. Together, our contributions present a more rigorous alternative to the dominant accuracy-centric evaluation paradigm.

Despite high benchmark scores, Large Language Models (LLMs) often fail simple problem, raising a critical question: Do LLMs learn mathematical principles or merely memorize patterns? Rather than designing increasingly complex benchmarks like recent works, we investigate this using elementary two-integer addition (0 to 2^{64}), probing two core properties: commutativity (A+B=B+A) and compositional generalization (via isomorphic symbolic mappings, e.g., 7 rightarrow y). While state-of-the-art LLMs achieve 73.8-99.8\% accuracy on numerical addition, performance collapses to leq7.5\% under symbolic mapping, indicating failure to generalize learned rules. Non-monotonic performance scaling with digit count and frequent commutativity violations (over 1,700 cases of A+B neq B+A) further support this. Explicitly providing addition rules degrades performance by 81.2\% on average, while self-explanation maintains baseline accuracy, suggesting LLM arithmetic processing is misaligned with human-defined principles. Our findings indicate current LLMs rely on memory pattern over genuine rule learning, highlighting architectural limitations and the need for new approaches to achieve true mathematical reasoning.

Current benchmarks for Large Language Models (LLMs) primarily focus on performance metrics, often failing to capture the nuanced behavioral characteristics that differentiate them. This paper introduces a novel ``Behavioral Fingerprinting'' framework designed to move beyond traditional evaluation by creating a multi-faceted profile of a model's intrinsic cognitive and interactive styles. Using a curated Diagnostic Prompt Suite and an innovative, automated evaluation pipeline where a powerful LLM acts as an impartial judge, we analyze eighteen models across capability tiers. Our results reveal a critical divergence in the LLM landscape: while core capabilities like abstract and causal reasoning are converging among top models, alignment-related behaviors such as sycophancy and semantic robustness vary dramatically. We further document a cross-model default persona clustering (ISTJ/ESTJ) that likely reflects common alignment incentives. Taken together, this suggests that a model's interactive nature is not an emergent property of its scale or reasoning power, but a direct consequence of specific, and highly variable, developer alignment strategies. Our framework provides a reproducible and scalable methodology for uncovering these deep behavioral differences. Project: https://github.com/JarvisPei/Behavioral-Fingerprinting

Direct Preference Optimization (DPO), a standard method for aligning language models with human preferences, is traditionally applied to offline preferences. Recent studies show that DPO benefits from iterative training with online preferences labeled by a trained reward model. In this work, we identify a pitfall of vanilla iterative DPO - improved response quality can lead to increased verbosity. To address this, we introduce iterative length-regularized DPO (iLR-DPO) to penalize response length. Our empirical results show that iLR-DPO can enhance a 7B model to perform on par with GPT-4 without increasing verbosity. Specifically, our 7B model achieves a 50.5% length-controlled win rate against GPT-4 Preview on AlpacaEval 2.0, and excels across standard benchmarks including MT-Bench, Arena-Hard and OpenLLM Leaderboard. These results demonstrate the effectiveness of iterative DPO in aligning language models with human feedback.

Large language models (LLMs) show remarkable potential to act as computer agents, enhancing human productivity and software accessibility in multi-modal tasks that require planning and reasoning. However, measuring agent performance in realistic environments remains a challenge since: (i) most benchmarks are limited to specific modalities or domains (e.g. text-only, web navigation, Q&A, coding) and (ii) full benchmark evaluations are slow (on order of magnitude of days) given the multi-step sequential nature of tasks. To address these challenges, we introduce the Windows Agent Arena: a reproducible, general environment focusing exclusively on the Windows operating system (OS) where agents can operate freely within a real Windows OS and use the same wide range of applications, tools, and web browsers available to human users when solving tasks. We adapt the OSWorld framework (Xie et al., 2024) to create 150+ diverse Windows tasks across representative domains that require agent abilities in planning, screen understanding, and tool usage. Our benchmark is scalable and can be seamlessly parallelized in Azure for a full benchmark evaluation in as little as 20 minutes. To demonstrate Windows Agent Arena's capabilities, we also introduce a new multi-modal agent, Navi. Our agent achieves a success rate of 19.5% in the Windows domain, compared to 74.5% performance of an unassisted human. Navi also demonstrates strong performance on another popular web-based benchmark, Mind2Web. We offer extensive quantitative and qualitative analysis of Navi's performance, and provide insights into the opportunities for future research in agent development and data generation using Windows Agent Arena. Webpage: https://microsoft.github.io/WindowsAgentArena Code: https://github.com/microsoft/WindowsAgentArena

We introduce LLM-Wikirace, a benchmark for evaluating planning, reasoning, and world knowledge in large language models (LLMs). In LLM-Wikirace, models must efficiently navigate Wikipedia hyperlinks step by step to reach a target page from a given source, requiring look-ahead planning and the ability to reason about how concepts are connected in the real world. We evaluate a broad set of open- and closed-source models, including Gemini-3, GPT-5, and Claude Opus 4.5, which achieve the strongest results on the easy level of the task and demonstrate superhuman performance. Despite this, performance drops sharply on hard difficulty: the best-performing model, Gemini-3, succeeds in only 23\% of hard games, highlighting substantial remaining challenges for frontier models. Our analysis shows that world knowledge is a necessary ingredient for success, but only up to a point, beyond this threshold, planning and long-horizon reasoning capabilities become the dominant factors. Trajectory-level analysis further reveals that even the strongest models struggle to replan after failure, frequently entering loops rather than recovering. LLM-Wikirace is a simple benchmark that reveals clear limitations in current reasoning systems, offering an open arena where planning-capable LLMs still have much to prove. Our code and leaderboard available at https:/llmwikirace.github.io.

Existing language model evaluations primarily measure general capabilities, yet reliable use of these models across a range of domains demands factual accuracy and recognition of knowledge gaps. We introduce AA-Omniscience, a benchmark designed to measure both factual recall and knowledge calibration across 6,000 questions. Questions are derived from authoritative academic and industry sources, and cover 42 economically relevant topics within six different domains. The evaluation measures a model's Omniscience Index, a bounded metric (-100 to 100) measuring factual recall that jointly penalizes hallucinations and rewards abstention when uncertain, with 0 equating to a model that answers questions correctly as much as it does incorrectly. Among evaluated models, Claude 4.1 Opus attains the highest score (4.8), making it one of only three models to score above zero. These results reveal persistent factuality and calibration weaknesses across frontier models. Performance also varies by domain, with the models from three different research labs leading across the six domains. This performance variability suggests models should be chosen according to the demands of the use case rather than general performance for tasks where knowledge is important.

The rapid proliferation of LLMs has created a critical evaluation paradox: while LLMs claim multilingual proficiency, comprehensive non-machine-translated benchmarks exist for fewer than 30 languages, leaving >98% of the world's 7,000 languages in an empirical void. Traditional benchmark construction faces scaling challenges such as cost, scarcity of domain experts, and data contamination. We evaluate the validity of a simpler alternative: can translation quality alone indicate a model's broader multilingual capabilities? Through systematic evaluation of 14 models (1B-72B parameters) across 9 diverse benchmarks and 7 translation metrics, we find that translation performance is a good indicator of downstream task success (e.g., Phi-4, median Pearson r: MetricX = 0.89, xCOMET = 0.91, SSA-COMET = 0.87). These results suggest that the representational abilities supporting faithful translation overlap with those required for multilingual understanding. Translation quality, thus emerges as a strong, inexpensive first-pass proxy of multilingual performance, enabling a translation-first screening with targeted follow-up for specific tasks.

Including code in the pre-training data mixture, even for models not specifically designed for code, has become a common practice in LLMs pre-training. While there has been anecdotal consensus among practitioners that code data plays a vital role in general LLMs' performance, there is only limited work analyzing the precise impact of code on non-code tasks. In this work, we systematically investigate the impact of code data on general performance. We ask "what is the impact of code data used in pre-training on a large variety of downstream tasks beyond code generation". We conduct extensive ablations and evaluate across a broad range of natural language reasoning tasks, world knowledge tasks, code benchmarks, and LLM-as-a-judge win-rates for models with sizes ranging from 470M to 2.8B parameters. Across settings, we find a consistent results that code is a critical building block for generalization far beyond coding tasks and improvements to code quality have an outsized impact across all tasks. In particular, compared to text-only pre-training, the addition of code results in up to relative increase of 8.2% in natural language (NL) reasoning, 4.2% in world knowledge, 6.6% improvement in generative win-rates, and a 12x boost in code performance respectively. Our work suggests investments in code quality and preserving code during pre-training have positive impacts.

Evaluating large language models (LLMs) as general-purpose agents is essential for understanding their capabilities and facilitating their integration into practical applications. However, the evaluation process presents substantial challenges. A primary obstacle is the benchmarking of agent performance across diverse scenarios within a unified framework, especially in maintaining partially-observable environments and ensuring multi-round interactions. Moreover, current evaluation frameworks mostly focus on the final success rate, revealing few insights during the process and failing to provide a deep understanding of the model abilities. To address these challenges, we introduce AgentBoard, a pioneering comprehensive benchmark and accompanied open-source evaluation framework tailored to analytical evaluation of LLM agents. AgentBoard offers a fine-grained progress rate metric that captures incremental advancements as well as a comprehensive evaluation toolkit that features easy assessment of agents for multi-faceted analysis through interactive visualization. This not only sheds light on the capabilities and limitations of LLM agents but also propels the interpretability of their performance to the forefront. Ultimately, AgentBoard serves as a significant step towards demystifying agent behaviors and accelerating the development of stronger LLM agents.

Agent skills, which are reusable, domain-specific knowledge artifacts, have become a popular mechanism for extending LLM-based agents, yet formally benchmarking skill usage performance remains scarce. Existing skill benchmarking efforts focus on overly idealized conditions, where LLMs are directly provided with hand-crafted, narrowly-tailored task-specific skills for each task, whereas in many realistic settings, the LLM agent may have to search for and select relevant skills on its own, and even the closest matching skills may not be well-tailored for the task. In this paper, we conduct the first comprehensive study of skill utility under progressively challenging realistic settings, where agents must retrieve skills from a large collection of 34k real-world skills and may not have access to any hand-curated skills. Our findings reveal that the benefits of skills are fragile: performance gains degrade consistently as settings become more realistic, with pass rates approaching no-skill baselines in the most challenging scenarios. To narrow this gap, we study skill refinement strategies, including query-specific and query-agnostic approaches, and we show that query-specific refinement substantially recovers lost performance when the initial skills are of reasonable relevance and quality. We further demonstrate the generality of retrieval and refinement on Terminal-Bench 2.0, where they improve the pass rate of Claude Opus 4.6 from 57.7% to 65.5%. Our results, consistent across multiple models, highlight both the promise and the current limitations of skills for LLM-based agents. Our code is available at https://github.com/UCSB-NLP-Chang/Skill-Usage.

Large Language Models (LLMs) ) have demonstrated promise in boosting productivity across AI-powered tools, yet existing benchmarks like Massive Multitask Language Understanding (MMLU) inadequately assess enterprise-specific task complexities. We propose a 14-task framework grounded in Bloom's Taxonomy to holistically evaluate LLM capabilities in enterprise contexts. To address challenges of noisy data and costly annotation, we develop a scalable pipeline combining LLM-as-a-Labeler, LLM-as-a-Judge, and corrective retrieval-augmented generation (CRAG), curating a robust 9,700-sample benchmark. Evaluation of six leading models shows open-source contenders like DeepSeek R1 rival proprietary models in reasoning tasks but lag in judgment-based scenarios, likely due to overthinking. Our benchmark reveals critical enterprise performance gaps and offers actionable insights for model optimization. This work provides enterprises a blueprint for tailored evaluations and advances practical LLM deployment.

Although large language models (LLMs) have been assessed for general medical knowledge using medical licensing exams, their ability to effectively support clinical decision-making tasks, such as selecting and using medical calculators, remains uncertain. Here, we evaluate the capability of both medical trainees and LLMs to recommend medical calculators in response to various multiple-choice clinical scenarios such as risk stratification, prognosis, and disease diagnosis. We assessed eight LLMs, including open-source, proprietary, and domain-specific models, with 1,009 question-answer pairs across 35 clinical calculators and measured human performance on a subset of 100 questions. While the highest-performing LLM, GPT-4o, provided an answer accuracy of 74.3% (CI: 71.5-76.9%), human annotators, on average, outperformed LLMs with an accuracy of 79.5% (CI: 73.5-85.0%). With error analysis showing that the highest-performing LLMs continue to make mistakes in comprehension (56.6%) and calculator knowledge (8.1%), our findings emphasize that humans continue to surpass LLMs on complex clinical tasks such as calculator recommendation.

Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs' abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry. For example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Our code and data can be found at https://github.com/felipemaiapolo/prompt-eval.

Precise estimation of downstream performance in large language models (LLMs) prior to training is essential for guiding their development process. Scaling laws analysis utilizes the statistics of a series of significantly smaller sampling language models (LMs) to predict the performance of the target LLM. For downstream performance prediction, the critical challenge lies in the emergent abilities in LLMs that occur beyond task-specific computational thresholds. In this work, we focus on the pre-training loss as a more computation-efficient metric for performance estimation. Our two-stage approach FLP consists of first estimating a function that maps computational resources (e.g., FLOPs) to the pre-training Loss using a series of fully-converged sampling models, followed by mapping the pre-training loss to downstream task Performance using the intermediate models with emerged performance. In our experiments, this FLP solution accurately predicts the performance of LLMs with 7B and 13B parameters using a series of sampling LMs up to 3B, achieving error margins of 5% and 10%, respectively, and significantly outperforming the FLOPs-to-Performance approach. Further, we present FLP-M, a fundamental approach for performance prediction that addresses the practical need to integrate datasets from multiple sources during pre-training. FLP-M extends the power law analytical function to predict domain-specific pre-training loss based on FLOPs across data sources, and employs a two-layer neural network to model the non-linear relationship between multiple domain-specific loss and downstream performance. By utilizing a 3B LLM trained on a specific ratio and a series of smaller sampling LMs, FLP-M can effectively forecast the performance of 3B and 7B LLMs across various data mixtures for most benchmarks within 10% error margins.

Given the increasing use of synthetic data in language model (LM) post-training, an LM's ability to generate high-quality data has become nearly as crucial as its ability to solve problems directly. While prior works have focused on developing effective data generation methods, they lack systematic comparison of different LMs as data generators in a unified setting. To address this gap, we propose AgoraBench, a benchmark that provides standardized settings and metrics to evaluate LMs' data generation abilities. Through synthesizing 1.26 million training instances using 6 LMs and training 99 student models, we uncover key insights about LMs' data generation capabilities. First, we observe that LMs exhibit distinct strengths. For instance, GPT-4o excels at generating new problems, while Claude-3.5-Sonnet performs better at enhancing existing ones. Furthermore, our analysis reveals that an LM's data generation ability doesn't necessarily correlate with its problem-solving ability. Instead, multiple intrinsic features of data quality-including response quality, perplexity, and instruction difficulty-collectively serve as better indicators. Finally, we demonstrate that strategic choices in output format and cost-conscious model selection significantly impact data generation effectiveness.

Reinforcement Learning from Human Feedback (RLHF) is an effective approach for aligning language models to human preferences. Central to RLHF is learning a reward function for scoring human preferences. Two main approaches for learning a reward model are 1) training an EXplicit Reward Model (EXRM) as in RLHF, and 2) using an implicit reward learned from preference data through methods such as Direct Preference Optimization (DPO). Prior work has shown that the implicit reward model of DPO (denoted as DPORM) can approximate an EXRM in the limit. DPORM's effectiveness directly implies the optimality of the learned policy, and also has practical implication for LLM alignment methods including iterative DPO. However, it is unclear how well DPORM empirically matches the performance of EXRM. This work studies the accuracy at distinguishing preferred and rejected answers for both DPORM and EXRM. Our findings indicate that even though DPORM fits the training dataset comparably, it generalizes less effectively than EXRM, especially when the validation datasets contain distribution shifts. Across five out-of-distribution settings, DPORM has a mean drop in accuracy of 3% and a maximum drop of 7%. These findings highlight that DPORM has limited generalization ability and substantiates the integration of an explicit reward model in iterative DPO approaches.

Evaluation of Large Language Models (LLMs) is challenging because aligning to human values requires the composition of multiple skills and the required set of skills varies depending on the instruction. Recent studies have evaluated the performance of LLMs in two ways, (1) automatic evaluation on several independent benchmarks and (2) human or machined-based evaluation giving an overall score to the response. However, both settings are coarse-grained evaluations, not considering the nature of user instructions that require instance-wise skill composition, which limits the interpretation of the true capabilities of LLMs. In this paper, we introduce FLASK (Fine-grained Language Model Evaluation based on Alignment SKill Sets), a fine-grained evaluation protocol that can be used for both model-based and human-based evaluation which decomposes coarse-level scoring to an instance-wise skill set-level. Specifically, we define 12 fine-grained skills needed for LLMs to follow open-ended user instructions and construct an evaluation set by allocating a set of skills for each instance. Additionally, by annotating the target domains and difficulty level for each instance, FLASK provides a holistic view with a comprehensive analysis of a model's performance depending on skill, domain, and difficulty. Through using FLASK, we compare multiple open-sourced and proprietary LLMs and observe highly-correlated findings between model-based and human-based evaluations. FLASK enables developers to more accurately measure the model performance and how it can be improved by analyzing factors that make LLMs proficient in particular skills. For practitioners, FLASK can be used to recommend suitable models for particular situations through comprehensive comparison among various LLMs. We release the evaluation data and code implementation at https://github.com/kaistAI/FLASK.

The rise of large language models (LLMs) has created a significant disparity: industrial research labs with their computational resources, expert teams, and advanced infrastructures, can effectively fine-tune LLMs, while individual developers and small organizations face barriers due to limited resources. In this paper, we aim to bridge this gap by presenting a comprehensive study on supervised fine-tuning of LLMs using instruction-tuning datasets spanning diverse knowledge domains and skills. We focus on small-sized LLMs (3B to 7B parameters) for their cost-efficiency and accessibility. We explore various training configurations and strategies across four open-source pre-trained models. We provide detailed documentation of these configurations, revealing findings that challenge several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca. Key insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, enabling early termination of sub-optimal runs and significant computational savings; (iii) through a thorough exploration of hyperparameters like warmup steps and learning rate schedules, we provide guidance for practitioners and find that certain simplifications do not compromise performance; and (iv) we observed no significant difference in performance between phased and stacked training strategies, but stacked training is simpler and more sample efficient. With these findings holding robustly across datasets and models, we hope this study serves as a guide for practitioners fine-tuning small LLMs and promotes a more inclusive environment for LLM research.

Benchmarks are fundamental for evaluating and advancing LLMs and MLLMs by providing standardized and explicit measures of performance. However, their construction is labor-intensive and hard to reuse, raising concerns about sustainability and scalability. Moreover, existing benchmarks often quickly reach performance saturation after their release, resulting in insufficient discrimination among state-of-the-art models. To address these challenges, we introduce Benchmark Agent, a fully autonomous agentic system designed for benchmark building. Our framework orchestrates the complete benchmark construction pipeline, from user query analysis and subtask design to data annotation and quality control. To assess Benchmark Agent, we implement it to produce 15 representative benchmarks, spanning diverse evaluation scenarios, including text understanding, multimodal understanding, and domain-specific reasoning. Extensive experiments, including human evaluation, LLM-as-a-judge assessment, and consistency checks, demonstrate Benchmark Agent can generate high-quality benchmark samples with minimal human involvement. More importantly, through continual evaluation, we observe several insightful findings, including that current models struggle with certain domain-specific reasoning tasks. We believe that rapidly evolving benchmarks can contribute significantly to the research community. The preview and code will be publicly available at the demo page and code repository.

Vision-language model (VLM) agents are increasingly deployed in interactive game environments. Yet game benchmarks for VLM agents typically report a single first-attempt score per (agent, game) pair, focus on single-agent Solo play, and lack unified protocols for evaluating heterogeneous agent classes (commercial VLMs, open-weight VLMs, and specialized game policies) on the same footing. We address these gaps with OmniGameArena, a real-time benchmark of twelve newly built Unreal Engine 5 games spanning Solo (7), PvP (3), and Coop (2) with unified action interfaces, and the Improvement Dynamics Curve (IDC), an agentic-reflection harness in which a tool-using reflector LLM autonomously refines a bounded skill prompt across multiple rounds. Beyond cold-start leaderboard scores, IDC exposes two additional observables for each (agent, game) pair: how the score evolves across reflection rounds, and how the learned skill behaves on held-out task variants. We report these observables for twelve VLM agents on the cold-start leaderboard and four top agents under IDC.

We present RPGBench, the first benchmark designed to evaluate large language models (LLMs) as text-based role-playing game (RPG) engines. RPGBench comprises two core tasks: Game Creation (GC) and Game Simulation (GS). In GC, an LLM must craft a valid and playable RPG world using a structured event-state representation, ensuring logical coherence and proper termination conditions. In GS, the LLM simulates interactive gameplay across multiple rounds while consistently updating states and enforcing game rules. To comprehensively assess performance, RPGBench integrates objective and subjective evaluation methodologies. Objective measures verify adherence to event mechanics and check variable updates without requiring human intervention. Subjective measures, such as content interestingness, action quality, and role-playing capability, are evaluated via an LLM-as-a-judge framework, where a strong LLM grades each candidate's outputs. Empirical results demonstrate that state-of-the-art LLMs can produce engaging stories but often struggle to implement consistent, verifiable game mechanics, particularly in long or complex scenarios. By combining structured, rule-based assessments with LLM-based judgments, RPGBench provides a new standard for evaluating how well LLMs can balance creativity, coherence, and complexity in text-based RPGs, opening avenues for more immersive and controllable interactive storytelling.

Multi-turn, multi-agent LLM game evaluations often exhibit substantial run-to-run variance. In long-horizon interactions, small early deviations compound across turns and are amplified by multi-agent coupling. This biases win rate estimates and makes rankings unreliable across repeated tournaments. Prompt choice worsens this further by producing different effective policies. We address both instability and underperformance with MEMO (Memory-augmented MOdel context optimization), a self-play framework that optimizes inference-time context by coupling retention and exploration. Retention maintains a persistent memory bank that stores structured insights from self-play trajectories and injects them as priors during later play. Exploration runs tournament-style prompt evolution with uncertainty-aware selection via TrueSkill, and uses prioritized replay to revisit rare and decisive states. Across five text-based games, MEMO raises mean win rate from 25.1% to 49.5% for GPT-4o-mini and from 20.9% to 44.3% for Qwen-2.5-7B-Instruct, using 2,000 self-play games per task. Run-to-run variance also drops, giving more stable rankings across prompt variations. These results suggest that multi-agent LLM game performance and robustness have substantial room for improvement through context optimization. MEMO achieves the largest gains in negotiation and imperfect-information games, while RL remains more effective in perfect-information settings.

The unprecedented performance of large language models (LLMs) requires comprehensive and accurate evaluation. We argue that for LLMs evaluation, benchmarks need to be comprehensive and systematic. To this end, we propose the ZhuJiu benchmark, which has the following strengths: (1) Multi-dimensional ability coverage: We comprehensively evaluate LLMs across 7 ability dimensions covering 51 tasks. Especially, we also propose a new benchmark that focuses on knowledge ability of LLMs. (2) Multi-faceted evaluation methods collaboration: We use 3 different yet complementary evaluation methods to comprehensively evaluate LLMs, which can ensure the authority and accuracy of the evaluation results. (3) Comprehensive Chinese benchmark: ZhuJiu is the pioneering benchmark that fully assesses LLMs in Chinese, while also providing equally robust evaluation abilities in English. (4) Avoiding potential data leakage: To avoid data leakage, we construct evaluation data specifically for 37 tasks. We evaluate 10 current mainstream LLMs and conduct an in-depth discussion and analysis of their results. The ZhuJiu benchmark and open-participation leaderboard are publicly released at http://www.zhujiu-benchmark.com/ and we also provide a demo video at https://youtu.be/qypkJ89L1Ic.

Large Language Models (LLMs) are rapidly evolving and impacting various fields, necessitating the development of effective methods to evaluate and compare their performance. Most current approaches for performance evaluation are either based on fixed, domain-specific questions that lack the flexibility required in many real-world applications, or rely on human input, making them unscalable. To address these issues, we propose an automated benchmarking framework based on debates between LLMs, judged by another LLM. This method assesses not only domain knowledge, but also skills such as argumentative reasoning and inconsistency recognition. We evaluate the performance of various state-of-the-art LLMs using the debate framework and achieve rankings that align closely with popular rankings based on human input, eliminating the need for costly human crowdsourcing.

Memory for large language model (LLM) agents has rapidly evolved from simple retrieval-augmented mechanisms into a data management system that supports persistent information storage, retrieval, update, consolidation, and dynamic lifecycle governance throughout agent execution. Despite this evolution, existing evaluations still benchmark agent memory mainly through end-to-end task success metrics (e.g., F1, BLEU), while treating the underlying system as a monolithic black box. As a result, critical system-level concerns, including operational costs, architectural trade-offs across memory modules, and robustness under dynamic knowledge updates, remain insufficiently explored. In this paper, we present a systematic experimental study of agent memory from a data management perspective. We propose an analytical framework that decomposes agent memory into four core modules: memory representation and storage, extraction, retrieval and routing, and maintenance. Under this framework, we evaluate 12 representative memory systems and two reference baselines across five benchmark workloads spanning 11 datasets. Our extensive end-to-end evaluation shows that no single architecture dominates across all scenarios; instead, effectiveness depends heavily on how well the memory structure aligns with the workload bottleneck. Furthermore, through fine-grained ablation studies, we quantify their individual effects on representation fidelity, retrieval precision, update correctness, and long-horizon stability. Finally, we reveal cost-performance trade-offs under realistic workloads, showing localized maintenance is more cost-efficient than global reorganization. Based on these findings, we identify promising directions towards building truly agent-native memory systems. The code is publicly available at https://github.com/OpenDataBox/MemoryData.

Complex games have long been an important benchmark for testing the progress of artificial intelligence algorithms. AlphaGo, AlphaZero, and MuZero have defeated top human players in Go and Chess, garnering widespread societal attention towards artificial intelligence. Concurrently, large language models (LLMs) have exhibited remarkable capabilities across various tasks, raising the question of whether LLMs can achieve similar success in complex games. In this paper, we explore the potential of LLMs in mastering complex card games. We systematically assess the learning capabilities of LLMs across eight diverse card games, evaluating the impact of fine-tuning on high-quality gameplay data, and examining the models' ability to retain general capabilities while mastering these games. Our findings indicate that: (1) LLMs can approach the performance of strong game AIs through supervised fine-tuning on high-quality data, (2) LLMs can achieve a certain level of proficiency in multiple complex card games simultaneously, with performance augmentation for games with similar rules and conflicts for dissimilar ones, and (3) LLMs experience a decline in general capabilities when mastering complex games, but this decline can be mitigated by integrating a certain amount of general instruction data. The evaluation results demonstrate strong learning ability and versatility of LLMs. The code is available at https://github.com/THUDM/LLM4CardGame

Large language model (LLM) simulations of human behavior have the potential to revolutionize the social and behavioral sciences, if and only if they faithfully reflect real human behaviors. Current evaluations are fragmented, based on bespoke tasks and metrics, creating a patchwork of incomparable results. To address this, we introduce SimBench, the first large-scale, standardized benchmark for a robust, reproducible science of LLM simulation. By unifying 20 diverse datasets covering tasks from moral decision-making to economic choice across a large global participant pool, SimBench provides the necessary foundation to ask fundamental questions about when, how, and why LLM simulations succeed or fail. We show that, while even the best LLMs today have limited simulation ability (score: 40.80/100), performance scales log-linearly with model size. Simulation performance is not improved by increased inference-time compute. We demonstrate an alignment-simulation trade-off: instruction-tuning improves performance on low-entropy (consensus) questions but degrades it on high-entropy (diverse) ones. Models particularly struggle when simulating specific demographic groups. Finally, we demonstrate that simulation ability correlates most strongly with deep, knowledge-intensive reasoning (MMLU-Pro, r=0.939). By making progress measurable, we aim to accelerate the development of more faithful LLM simulators.

The rapid proliferation of Large Language Models (LLMs) and diverse specialized benchmarks necessitates a shift from fragmented, task-specific metrics to a holistic, competitive ranking system that effectively aggregates performance across multiple ability dimensions. Primarily using static scoring, current evaluation methods are fundamentally limited. They struggle to determine the proper mix ratio across diverse benchmarks, and critically, they fail to capture a model's dynamic competitive fitness or its vulnerability when confronted with sequential, high-stakes tasks. To address this, we introduce the novel Competitive Swiss-System Dynamics (CSD) framework. CSD simulates a multi-round, sequential contest where models are dynamically paired across a curated sequence of benchmarks based on their accumulated win-loss record. And Monte Carlo Simulation (N=100,000 iterations) is used to approximate the statistically robust Expected Win Score (E[S_m]), which eliminates the noise of random pairing and early-round luck. Furthermore, we implement a Failure Sensitivity Analysis by parameterizing the per-round elimination quantity (T_k), which allows us to profile models based on their risk appetite--distinguishing between robust generalists and aggressive specialists. We demonstrate that CSD provides a more nuanced and context-aware ranking than traditional aggregate scoring and static pairwise models, representing a vital step towards risk-informed, next-generation LLM evaluation.

In this research study, we empirically investigate the effect of sampling temperature on the performance of Large Language Models (LLMs) on various problem-solving tasks. We created a multiple-choice question-and-answer (MCQA) exam by randomly sampling problems from standard LLM benchmarks. Then, we used nine popular LLMs with five prompt-engineering techniques to solve the MCQA problems while increasing the sampling temperature from 0.0 to 1.6. Despite anecdotal reports to the contrary, our empirical results indicate that changes in temperature from 0.0 to 1.0 do not have a statistically significant impact on LLM performance for problem-solving tasks. In addition, these results appear to generalize across LLMs, prompt-engineering techniques, and problem domains. All code, data, and supplemental materials are available on GitHub at: https://github.com/matthewrenze/jhu-llm-temperature

Large Language Models (LLMs) have demonstrated remarkable performance across a spectrum of tasks. Recently, Direct Preference Optimization (DPO) has emerged as an RL-free approach to optimize the policy model on human preferences. However, several limitations hinder the widespread adoption of this method. To address these shortcomings, various versions of DPO have been introduced. Yet, a comprehensive evaluation of these variants across diverse tasks is still lacking. In this study, we aim to bridge this gap by investigating the performance of alignment methods across three distinct scenarios: (1) keeping the Supervised Fine-Tuning (SFT) part, (2) skipping the SFT part, and (3) skipping the SFT part and utilizing an instruction-tuned model. Furthermore, we explore the impact of different training sizes on their performance. Our evaluation spans a range of tasks including dialogue systems, reasoning, mathematical problem-solving, question answering, truthfulness, and multi-task understanding, encompassing 13 benchmarks such as MT-Bench, Big Bench, and Open LLM Leaderboard. Key observations reveal that alignment methods achieve optimal performance with smaller training data subsets, exhibit limited effectiveness in reasoning tasks yet significantly impact mathematical problem-solving, and employing an instruction-tuned model notably influences truthfulness. We anticipate that our findings will catalyze further research aimed at developing more robust models to address alignment challenges.

We propose a novel task, G4C, to study teacher-student natural language interactions in a goal-driven and grounded environment. Dungeons and Dragons (D&D), a role-playing game, provides an ideal setting to investigate such interactions. Here, the Dungeon Master (DM), i.e., the teacher, guides the actions of several players -- students, each with their own personas and abilities -- to achieve shared goals grounded in a fantasy world. Our approach is to decompose and model these interactions into (1) the DM's intent to guide players toward a given goal; (2) the DM's guidance utterance to the players expressing this intent; and (3) a theory-of-mind (ToM) model that anticipates the players' reaction to the guidance one turn into the future. We develop a novel reinforcement learning (RL) method for training a DM that generates guidance for players by rewarding utterances where the intent matches the ToM-anticipated player actions. Human and automated evaluations show that a DM trained to explicitly model intents and incorporate ToM of the players using RL generates better-quality guidance that is 3x more likely to fulfill the DM's intent than a vanilla natural language generation (NLG) approach.

We present initial results of a forthcoming benchmark for evaluating LLM agents on white-collar tasks of economic value. We evaluate agents on real-world "messy" open-web research tasks of the type that are routine in finance and consulting. In doing so, we lay the groundwork for an LLM agent evaluation suite where good performance directly corresponds to a large economic and societal impact. We built and tested several agent architectures with o1-preview, GPT-4o, Claude-3.5 Sonnet, Llama 3.1 (405b), and GPT-4o-mini. On average, LLM agents powered by Claude-3.5 Sonnet and o1-preview substantially outperformed agents using GPT-4o, with agents based on Llama 3.1 (405b) and GPT-4o-mini lagging noticeably behind. Across LLMs, a ReAct architecture with the ability to delegate subtasks to subagents performed best. In addition to quantitative evaluations, we qualitatively assessed the performance of the LLM agents by inspecting their traces and reflecting on their observations. Our evaluation represents the first in-depth assessment of agents' abilities to conduct challenging, economically valuable analyst-style research on the real open web.

It has been recognized that the data generated by the denoising diffusion probabilistic model (DDPM) improves adversarial training. After two years of rapid development in diffusion models, a question naturally arises: can better diffusion models further improve adversarial training? This paper gives an affirmative answer by employing the most recent diffusion model which has higher efficiency (sim 20 sampling steps) and image quality (lower FID score) compared with DDPM. Our adversarially trained models achieve state-of-the-art performance on RobustBench using only generated data (no external datasets). Under the ell_infty-norm threat model with epsilon=8/255, our models achieve 70.69% and 42.67% robust accuracy on CIFAR-10 and CIFAR-100, respectively, i.e. improving upon previous state-of-the-art models by +4.58% and +8.03%. Under the ell_2-norm threat model with epsilon=128/255, our models achieve 84.86% on CIFAR-10 (+4.44%). These results also beat previous works that use external data. Our code is available at https://github.com/wzekai99/DM-Improves-AT.

Global healthcare providers are exploring use of large language models (LLMs) to provide medical advice to the public. LLMs now achieve nearly perfect scores on medical licensing exams, but this does not necessarily translate to accurate performance in real-world settings. We tested if LLMs can assist members of the public in identifying underlying conditions and choosing a course of action (disposition) in ten medical scenarios in a controlled study with 1,298 participants. Participants were randomly assigned to receive assistance from an LLM (GPT-4o, Llama 3, Command R+) or a source of their choice (control). Tested alone, LLMs complete the scenarios accurately, correctly identifying conditions in 94.9% of cases and disposition in 56.3% on average. However, participants using the same LLMs identified relevant conditions in less than 34.5% of cases and disposition in less than 44.2%, both no better than the control group. We identify user interactions as a challenge to the deployment of LLMs for medical advice. Standard benchmarks for medical knowledge and simulated patient interactions do not predict the failures we find with human participants. Moving forward, we recommend systematic human user testing to evaluate interactive capabilities prior to public deployments in healthcare.

Standard evaluation of LLM confidence relies on calibration metrics (ECE, Brier score) that conflate two distinct capacities: how much a model knows (Type-1 sensitivity) and how well it knows what it knows (Type-2 metacognitive sensitivity). We introduce an evaluation framework based on Type-2 Signal Detection Theory that decomposes these capacities using meta-d' and the metacognitive efficiency ratio M-ratio. Applied to four LLMs (Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.3, Llama-3-8B-Base, Gemma-2-9B-Instruct) across 224,000 factual QA trials, we find: (1) metacognitive efficiency varies substantially across models even when Type-1 sensitivity is similar -- Mistral achieves the highest d' but the lowest M-ratio; (2) metacognitive efficiency is domain-specific, with different models showing different weakest domains, invisible to aggregate metrics; (3) temperature manipulation shifts Type-2 criterion while meta-d' remains stable for two of four models, dissociating confidence policy from metacognitive capacity; (4) AUROC_2 and M-ratio produce fully inverted model rankings, demonstrating these metrics answer fundamentally different evaluation questions. The meta-d' framework reveals which models "know what they don't know" versus which merely appear well-calibrated due to criterion placement -- a distinction with direct implications for model selection, deployment, and human-AI collaboration. Pre-registered analysis; code and data publicly available.

While Large Language Model (LLM) agents have achieved remarkable progress in complex reasoning tasks, evaluating their performance in real-world environments has become a critical problem. Current benchmarks, however, are largely restricted to idealized simulations, failing to address the practical demands of specialized domains like advertising and marketing analytics. In these fields, tasks are inherently more complex, often requiring multi-round interaction with professional marketing tools. To address this gap, we propose AD-Bench, a benchmark designed based on real-world business requirements of advertising and marketing platforms. AD-Bench is constructed from real user marketing analysis requests, with domain experts providing verifiable reference answers and corresponding reference tool-call trajectories. The benchmark categorizes requests into three difficulty levels (L1-L3) to evaluate agents' capabilities under multi-round, multi-tool collaboration. Experiments show that on AD-Bench, Gemini-3-Pro achieves Pass@1 = 68.0% and Pass@3 = 83.0%, but performance drops significantly on L3 to Pass@1 = 49.4% and Pass@3 = 62.1%, with a trajectory coverage of 70.1%, indicating that even state-of-the-art models still exhibit substantial capability gaps in complex advertising and marketing analysis scenarios. AD-Bench provides a realistic benchmark for evaluating and improving advertising marketing agents, the leaderboard and code can be found at https://github.com/Emanual20/adbench-leaderboard.

Distribution Matching Distillation (DMD) is a widely used paradigm for accelerating inference in few-step video diffusion models. However, DMD-style video distillation faces two coupled challenges: the fake score must track a continuously evolving generator, making training costly when frequent updates are required, while reverse-KL-style matching can be mode-seeking and conservative for preserving strong motion dynamics. To address these issues, we propose Score Gradient Matching Distillation (SGMD). SGMD adopts a fake-score perspective by directly optimizing the fake score toward the teacher, while using teacher stop-gradient Fisher as a stable distribution-matching objective. We provide a gradient analysis that motivates this objective choice under ideal tracking. Building on this, SGMD introduces a pair of dual potentials: negative-residual (NR) for outer-loop correction and residual-contraction (RC) for inner-loop tracking. Empirically, compared to DMD2, SGMD achieves an approximately sim 3times training speedup and substantially improves motion dynamics for 4-step distilled models while preserving temporal consistency. A human study confirms that SGMD is preferred in motion quality and overall preference, while visual quality and text alignment remain comparable. Code is available at https://github.com/ModelTC/LightX2V.

As LLM-based agents are deployed in production systems, understanding their behavioral consistency (whether they produce similar action sequences when given identical tasks) becomes critical for reliability. We study consistency in the context of SWE-bench, a challenging software engineering benchmark requiring complex, multi-step reasoning. Comparing Claude~4.5~Sonnet, GPT-5, and Llama-3.1-70B across 50 runs each (10 tasks times 5 runs), we find that across models, higher consistency aligns with higher accuracy: Claude achieves the lowest variance (CV: 15.2\%) and highest accuracy (58\%), GPT-5 is intermediate (CV: 32.2\%, accuracy: 32\%), and Llama shows the highest variance (CV: 47.0\%) with lowest accuracy (4\%). However, within a model, consistency can amplify both correct and incorrect interpretations. Our analysis reveals a critical nuance: consistency amplifies outcomes rather than guaranteeing correctness. 71\% of Claude's failures stem from "consistent wrong interpretation": making the same incorrect assumption across all runs. Interestingly, GPT-5 achieves similar early strategic agreement as Claude (diverging at step 3.4 vs.\ 3.2) but exhibits 2.1times higher variance, suggesting that divergence timing alone does not determine consistency. These findings suggest that for production deployment, interpretation accuracy matters more than execution consistency, with implications for agent evaluation and training.

The automatic evaluation of instruction following typically involves using large language models (LLMs) to assess response quality. However, there is a lack of comprehensive evaluation of these LLM-based evaluators across two dimensions: the base LLMs and the evaluation protocols. Therefore, we present a thorough meta-evaluation of instruction following, including 25 base LLMs and 15 recently proposed evaluation protocols, on 4 human-annotated datasets, assessing the evaluation accuracy of the LLM-evaluators. Our evaluation allows us to identify the best-performing base LLMs and evaluation protocols with a high degree of robustness. Moreover, our large-scale evaluation reveals: (1) Base LLM performance ranking remains largely consistent across evaluation protocols, with less capable LLMs showing greater improvement from protocol enhancements; (2) Robust evaluation of evaluation protocols requires many base LLMs with varying capability levels, as protocol effectiveness can depend on the base LLM used; (3) Evaluation results on different datasets are not always consistent, so a rigorous evaluation requires multiple datasets with distinctive features. We release our meta-evaluation suite ReIFE, which provides the codebase and evaluation result collection for more than 500 LLM-evaluator configurations, to support future research in instruction-following evaluation.

Conversations with LMs involve two participants: a human user leading the conversation, and an LM assistant responding to the user's request. To satisfy this specific role, LMs are post-trained to be helpful assistants -- optimized to produce exhaustive and well-structured responses, free of ambiguity and grammar errors. User utterances, on the other hand, are rarely perfected, with each user phrasing requests in unique ways, sometimes putting in partial effort at each turn and refining on the fly. To evaluate LM performance in realistic settings, prior work simulated users in multi-turn conversations, often prompting an LLM originally trained to be a helpful assistant to act as a user. However, we show that assistant LMs make for poor user simulators, with the surprising finding that better assistants yield worse simulators. Instead, we introduce purpose-built User Language Models (User LMs) - models post-trained to simulate human users in multi-turn conversations. Through various evaluations, we show how User LMs align better with human behavior and achieve better simulation robustness than existing simulation methods. When leveraging User LMs to simulate coding and math conversations, the performance of a strong assistant (GPT-4o) drops from 74.6% to 57.4%, confirming that more realistic simulation environments lead to assistant struggles as they fail to cope with the nuances of users in multi-turn setups.

Performance optimization is a critical yet challenging aspect of software development, often requiring a deep understanding of system behavior, algorithmic tradeoffs, and careful code modifications. Although recent advances in AI coding agents have accelerated code generation and bug fixing, little is known about how these agents perform on real-world performance optimization tasks. We present the first empirical study comparing agent- and human-authored performance optimization commits, analyzing 324 agent-generated and 83 human-authored PRs from the AIDev dataset across adoption, maintainability, optimization patterns, and validation practices. We find that AI-authored performance PRs are less likely to include explicit performance validation than human-authored PRs (45.7\% vs. 63.6\%, p=0.007). In addition, AI-authored PRs largely use the same optimization patterns as humans. We further discuss limitations and opportunities for advancing agentic code optimization.