Daily Papers

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse domains, but their training remains resource- and time-intensive, requiring massive compute power and careful orchestration of training procedures. Model souping-the practice of averaging weights from multiple models of the same architecture-has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. In this paper, we introduce Soup Of Category Experts (SoCE), a principled approach for model souping that utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous uniform-averaging approaches, our method leverages the observation that benchmark categories often exhibit low inter-correlations in model performance. SoCE identifies "expert" models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that the proposed method improves performance and robustness across multiple domains, including multilingual capabilities, tool calling, and math and achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.

Islamic inheritance domain holds significant importance for Muslims to ensure fair distribution of shares between heirs. Manual calculation of shares under numerous scenarios is complex, time-consuming, and error-prone. Recent advancements in Large Language Models (LLMs) have sparked interest in their potential to assist with complex legal reasoning tasks. This study evaluates the reasoning capabilities of state-of-the-art LLMs to interpret and apply Islamic inheritance laws. We utilized the dataset proposed in the ArabicNLP QIAS 2025 challenge, which includes inheritance case scenarios given in Arabic and derived from Islamic legal sources. Various base and fine-tuned models, are assessed on their ability to accurately identify heirs, compute shares, and justify their reasoning in alignment with Islamic legal principles. Our analysis reveals that the proposed majority voting solution, leveraging three base models (Gemini Flash 2.5, Gemini Pro 2.5, and GPT o3), outperforms all other models that we utilized across every difficulty level. It achieves up to 92.7% accuracy and secures the third place overall in Task 1 of the Qias 2025 challenge.

Large language models (LLMs) have demonstrated remarkable capabilities, especially the recent advancements in reasoning, such as o1 and o3, pushing the boundaries of AI. Despite these impressive achievements in mathematics and coding, the reasoning abilities of LLMs in domains requiring cryptographic expertise remain underexplored. In this paper, we introduce CipherBank, a comprehensive benchmark designed to evaluate the reasoning capabilities of LLMs in cryptographic decryption tasks. CipherBank comprises 2,358 meticulously crafted problems, covering 262 unique plaintexts across 5 domains and 14 subdomains, with a focus on privacy-sensitive and real-world scenarios that necessitate encryption. From a cryptographic perspective, CipherBank incorporates 3 major categories of encryption methods, spanning 9 distinct algorithms, ranging from classical ciphers to custom cryptographic techniques. We evaluate state-of-the-art LLMs on CipherBank, e.g., GPT-4o, DeepSeek-V3, and cutting-edge reasoning-focused models such as o1 and DeepSeek-R1. Our results reveal significant gaps in reasoning abilities not only between general-purpose chat LLMs and reasoning-focused LLMs but also in the performance of current reasoning-focused models when applied to classical cryptographic decryption tasks, highlighting the challenges these models face in understanding and manipulating encrypted data. Through detailed analysis and error investigations, we provide several key observations that shed light on the limitations and potential improvement areas for LLMs in cryptographic reasoning. These findings underscore the need for continuous advancements in LLM reasoning capabilities.

Customer Relationship Management (CRM) systems are vital for modern enterprises, providing a foundation for managing customer interactions and data. Integrating AI agents into CRM systems can automate routine processes and enhance personalized service. However, deploying and evaluating these agents is challenging due to the lack of realistic benchmarks that reflect the complexity of real-world CRM tasks. To address this issue, we introduce CRMArena, a novel benchmark designed to evaluate AI agents on realistic tasks grounded in professional work environments. Following guidance from CRM experts and industry best practices, we designed CRMArena with nine customer service tasks distributed across three personas: service agent, analyst, and manager. The benchmark includes 16 commonly used industrial objects (e.g., account, order, knowledge article, case) with high interconnectivity, along with latent variables (e.g., complaint habits, policy violations) to simulate realistic data distributions. Experimental results reveal that state-of-the-art LLM agents succeed in less than 40% of the tasks with ReAct prompting, and less than 55% even with function-calling abilities. Our findings highlight the need for enhanced agent capabilities in function-calling and rule-following to be deployed in real-world work environments. CRMArena is an open challenge to the community: systems that can reliably complete tasks showcase direct business value in a popular work environment.

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ``Deceitful'' and ``Manipulative'', often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

With the rise of generative large language models (LLMs) like LLaMA and ChatGPT, these models have significantly transformed daily life and work by providing advanced insights. However, as jailbreak attacks continue to circumvent built-in safety mechanisms, exploiting carefully crafted scenarios or tokens, the safety risks of LLMs have come into focus. While numerous defense strategies--such as prompt detection, modification, and model fine-tuning--have been proposed to counter these attacks, a critical question arises: do these defenses compromise the utility and usability of LLMs for legitimate users? Existing research predominantly focuses on the effectiveness of defense strategies without thoroughly examining their impact on performance, leaving a gap in understanding the trade-offs between LLM safety and performance. Our research addresses this gap by conducting a comprehensive study on the utility degradation, safety elevation, and exaggerated-safety escalation of LLMs with jailbreak defense strategies. We propose USEBench, a novel benchmark designed to evaluate these aspects, along with USEIndex, a comprehensive metric for assessing overall model performance. Through experiments on seven state-of-the-art LLMs, we found that mainstream jailbreak defenses fail to ensure both safety and performance simultaneously. Although model-finetuning performs the best overall, their effectiveness varies across LLMs. Furthermore, vertical comparisons reveal that developers commonly prioritize performance over safety when iterating or fine-tuning their LLMs.

As the influence of large language models (LLMs) spans across global communities, their safety challenges in multilingual settings become paramount for alignment research. This paper examines the variations in safety challenges faced by LLMs across different languages and discusses approaches to alleviating such concerns. By comparing how state-of-the-art LLMs respond to the same set of malicious prompts written in higher- vs. lower-resource languages, we observe that (1) LLMs tend to generate unsafe responses much more often when a malicious prompt is written in a lower-resource language, and (2) LLMs tend to generate more irrelevant responses to malicious prompts in lower-resource languages. To understand where the discrepancy can be attributed, we study the effect of instruction tuning with reinforcement learning from human feedback (RLHF) or supervised finetuning (SFT) on the HH-RLHF dataset. Surprisingly, while training with high-resource languages improves model alignment, training in lower-resource languages yields minimal improvement. This suggests that the bottleneck of cross-lingual alignment is rooted in the pretraining stage. Our findings highlight the challenges in cross-lingual LLM safety, and we hope they inform future research in this direction.

Ensembling multiple models has always been an effective approach to push the limits of existing performance and is widely used in classification tasks by simply averaging the classification probability vectors from multiple classifiers to achieve better accuracy. However, in the thriving open-source Large Language Model (LLM) community, ensembling methods are rare and typically limited to ensembling the full-text outputs of LLMs, such as selecting the best output using a ranker, which leads to underutilization of token-level probability information. In this paper, we treat the Generation of each token by LLMs as a Classification (GaC) for ensembling. This approach fully exploits the probability information at each generation step and better prevents LLMs from producing early incorrect tokens that lead to snowballing errors. In experiments, we ensemble state-of-the-art LLMs on several benchmarks, including exams, mathematics and reasoning, and observe that our method breaks the existing community performance ceiling. Furthermore, we observed that most of the tokens in the answer are simple and do not affect the correctness of the final answer. Therefore, we also experimented with ensembling only key tokens, and the results showed better performance with lower latency across benchmarks.

In a systematic way, we investigate a widely asked question: Do LLMs really understand what they say?, which relates to the more familiar term Stochastic Parrot. To this end, we propose a summative assessment over a carefully designed physical concept understanding task, PhysiCo. Our task alleviates the memorization issue via the usage of grid-format inputs that abstractly describe physical phenomena. The grids represents varying levels of understanding, from the core phenomenon, application examples to analogies to other abstract patterns in the grid world. A comprehensive study on our task demonstrates: (1) state-of-the-art LLMs, including GPT-4o, o1 and Gemini 2.0 flash thinking, lag behind humans by ~40%; (2) the stochastic parrot phenomenon is present in LLMs, as they fail on our grid task but can describe and recognize the same concepts well in natural language; (3) our task challenges the LLMs due to intrinsic difficulties rather than the unfamiliar grid format, as in-context learning and fine-tuning on same formatted data added little to their performance.

Large Language Model (LLM) code agents increasingly resolve repository-level issues by iteratively editing code, invoking tools, and validating candidate patches. In these workflows, agents often write tests on the fly, a paradigm adopted by many high-ranking agents on the SWE-bench leaderboard. However, we observe that GPT-5.2, which writes almost no new tests, can even achieve performance comparable to top-ranking agents. This raises the critical question: whether such tests meaningfully improve issue resolution or merely mimic human testing practices while consuming a substantial interaction budget. To reveal the impact of agent-written tests, we present an empirical study that analyzes agent trajectories across six state-of-the-art LLMs on SWE-bench Verified. Our results show that while test writing is commonly adopted, but resolved and unresolved tasks within the same model exhibit similar test-writing frequencies Furthermore, these tests typically serve as observational feedback channels, where agents prefer value-revealing print statements significantly more than formal assertion-based checks. Based on these insights, we perform a controlled experiment by revising the prompts of four agents to either increase or reduce test writing. The results suggest that changes in the volume of agent-written tests do not significantly change final outcomes. Taken together, our study reveals that current test-writing practices may provide marginal utility in autonomous software engineering tasks.

Current Large Language Models (LLMs) can assist developing program code beside many other things, but can they support working with Knowledge Graphs (KGs) as well? Which LLM is offering the best capabilities in the field of Semantic Web and Knowledge Graph Engineering (KGE)? Is this possible to determine without checking many answers manually? The LLM-KG-Bench framework in Version 3.0 is designed to answer these questions. It consists of an extensible set of tasks for automated evaluation of LLM answers and covers different aspects of working with semantic technologies. In this paper the LLM-KG-Bench framework is presented in Version 3 along with a dataset of prompts, answers and evaluations generated with it and several state-of-the-art LLMs. Significant enhancements have been made to the framework since its initial release, including an updated task API that offers greater flexibility in handling evaluation tasks, revised tasks, and extended support for various open models through the vllm library, among other improvements. A comprehensive dataset has been generated using more than 30 contemporary open and proprietary LLMs, enabling the creation of exemplary model cards that demonstrate the models' capabilities in working with RDF and SPARQL, as well as comparing their performance on Turtle and JSON-LD RDF serialization tasks.

Large Language Models (LLMs) are predominantly evaluated on Standard American English (SAE), often overlooking the diversity of global English varieties. This narrow focus may raise fairness concerns as degraded performance on non-standard varieties can lead to unequal benefits for users worldwide. Therefore, it is critical to extensively evaluate the linguistic robustness of LLMs on multiple non-standard English varieties. We introduce Trans-EnV, a framework that automatically transforms SAE datasets into multiple English varieties to evaluate the linguistic robustness. Our framework combines (1) linguistics expert knowledge to curate variety-specific features and transformation guidelines from linguistic literature and corpora, and (2) LLM-based transformations to ensure both linguistic validity and scalability. Using Trans-EnV, we transform six benchmark datasets into 38 English varieties and evaluate seven state-of-the-art LLMs. Our results reveal significant performance disparities, with accuracy decreasing by up to 46.3% on non-standard varieties. These findings highlight the importance of comprehensive linguistic robustness evaluation across diverse English varieties. Each construction of Trans-EnV was validated through rigorous statistical testing and consultation with a researcher in the field of second language acquisition, ensuring its linguistic validity. Our code and datasets are publicly available at https://github.com/jiyounglee-0523/TransEnV and https://huggingface.co/collections/jiyounglee0523/transenv-681eadb3c0c8cf363b363fb1.

In recent years, Large Language Models (LLMs) have shown great potential across a wide range of legal tasks. Despite these advances, mitigating hallucination remains a significant challenge, with state-of-the-art LLMs still frequently generating incorrect legal references. In this paper, we focus on the problem of legal citation prediction within the Australian law context, where correctly identifying and citing relevant legislations or precedents is critical. We compare several approaches: prompting general purpose and law-specialised LLMs, retrieval-only pipelines with both generic and domain-specific embeddings, task-specific instruction-tuning of LLMs, and hybrid strategies that combine LLMs with retrieval augmentation, query expansion, or voting ensembles. Our findings indicate that domain-specific pre-training alone is insufficient for achieving satisfactory citation accuracy even after law-specialised pre-training. In contrast, instruction tuning on our task-specific dataset dramatically boosts performance reaching the best results across all settings. We also highlight that database granularity along with the type of embeddings play a critical role in the performance of retrieval systems. Among retrieval-based approaches, hybrid methods consistently outperform retrieval-only setups, and among these, ensemble voting delivers the best result by combining the predictive quality of instruction-tuned LLMs with the retrieval system.

Current Large Language Models (LLMs) are predominantly designed with English as the primary language, and even the few that are multilingual tend to exhibit strong English-centric biases. Much like speakers who might produce awkward expressions when learning a second language, LLMs often generate unnatural outputs in non-English languages, reflecting English-centric patterns in both vocabulary and grammar. Despite the importance of this issue, the naturalness of multilingual LLM outputs has received limited attention. In this paper, we address this gap by introducing novel automatic corpus-level metrics to assess the lexical and syntactic naturalness of LLM outputs in a multilingual context. Using our new metrics, we evaluate state-of-the-art LLMs on a curated benchmark in French and Chinese, revealing a tendency towards English-influenced patterns. To mitigate this issue, we also propose a simple and effective alignment method to improve the naturalness of an LLM in a target language and domain, achieving consistent improvements in naturalness without compromising the performance on general-purpose benchmarks. Our work highlights the importance of developing multilingual metrics, resources and methods for the new wave of multilingual LLMs.

Existing code generation benchmarks primarily evaluate functional correctness, with limited focus on code efficiency and often restricted to a single language like Python. To address this gap, we introduce EffiBench-X, the first multi-language benchmark designed to measure the efficiency of LLM-generated code. EffiBench-X supports Python, C++, Java, JavaScript, Ruby, and Golang. It comprises competitive programming tasks with human-expert solutions as efficiency baselines. Evaluating state-of-the-art LLMs on EffiBench-X reveals that while models generate functionally correct code, they consistently underperform human experts in efficiency. Even the most efficient LLM-generated solutions (Qwen3-32B) achieve only around 62\% of human efficiency on average, with significant language-specific variations. LLMs show better efficiency in Python, Ruby, and JavaScript than in Java, C++, and Golang. For instance, DeepSeek-R1's Python code is significantly more efficient than its Java code. These results highlight the critical need for research into LLM optimization techniques to improve code efficiency across diverse languages. The dataset and evaluation infrastructure are submitted and available at https://github.com/EffiBench/EffiBench-X.git and https://huggingface.co/datasets/EffiBench/effibench-x.

Automated environment configuration is a critical bottleneck in scaling software engineering (SWE) automation. To provide a reliable evaluation standard for this task, we present Multi-Docker-Eval benchmark. It includes 40 real-world repositories spanning 9 programming languages and measures both success in achieving executable states and efficiency under realistic constraints. Our extensive evaluation of state-of-the-art LLMs and agent frameworks reveals key insights: (1) the overall success rate of current models is low (F2P at most 37.7%), with environment construction being the primary bottleneck; (2) model size and reasoning length are not decisive factors, and open-source models like DeepSeek-V3.1 and Kimi-K2 are competitive in both efficiency and effectiveness; (3) agent framework and programming language also have significantly influence on success rate. These findings provide actionable guidelines for building scalable, fully automated SWE pipelines.

Rigorous security-focused evaluation of large language model (LLM) agents is imperative for establishing trust in their safe deployment throughout the software development lifecycle. However, existing benchmarks largely rely on synthetic challenges or simplified vulnerability datasets that fail to capture the complexity and ambiguity encountered by security engineers in practice. We introduce SEC-bench, the first fully automated benchmarking framework for evaluating LLM agents on authentic security engineering tasks. SEC-bench employs a novel multi-agent scaffold that automatically constructs code repositories with harnesses, reproduces vulnerabilities in isolated environments, and generates gold patches for reliable evaluation. Our framework automatically creates high-quality software vulnerability datasets with reproducible artifacts at a cost of only $0.87 per instance. Using SEC-bench, we implement two critical software security tasks to rigorously evaluate LLM agents' capabilities: proof-of-concept (PoC) generation and vulnerability patching. A comprehensive evaluation of state-of-the-art LLM code agents reveals significant performance gaps, achieving at most 18.0% success in PoC generation and 34.0% in vulnerability patching on our complete dataset. These results highlight the crucial steps needed toward developing LLM agents that are more practical, intelligent, and autonomous for security engineering.

Organic reaction mechanisms are the stepwise elementary reactions by which reactants form intermediates and products, and are fundamental to understanding chemical reactivity and designing new molecules and reactions. Although large language models (LLMs) have shown promise in understanding chemical tasks such as synthesis design, it is unclear to what extent this reflects genuine chemical reasoning capabilities, i.e., the ability to generate valid intermediates, maintain chemical consistency, and follow logically coherent multi-step pathways. We address this by introducing oMeBench, the first large-scale, expert-curated benchmark for organic mechanism reasoning in organic chemistry. It comprises over 10,000 annotated mechanistic steps with intermediates, type labels, and difficulty ratings. Furthermore, to evaluate LLM capability more precisely and enable fine-grained scoring, we propose oMeS, a dynamic evaluation framework that combines step-level logic and chemical similarity. We analyze the performance of state-of-the-art LLMs, and our results show that although current models display promising chemical intuition, they struggle with correct and consistent multi-step reasoning. Notably, we find that using prompting strategy and fine-tuning a specialist model on our proposed dataset increases performance by 50% over the leading closed-source model. We hope that oMeBench will serve as a rigorous foundation for advancing AI systems toward genuine chemical reasoning.

The evaluation of Large Language Models (LLMs) on mathematical reasoning has largely focused on elementary problems, competition-style questions, or formal theorem proving, leaving graduate-level and computational mathematics relatively underexplored. We introduce CompMath-MCQ, a new benchmark dataset for assessing LLMs on advanced mathematical reasoning in a multiple-choice setting. The dataset consists of 1{,}500 originally authored questions by professors of graduate-level courses, covering topics including Linear Algebra, Numerical Optimization, Vector Calculus, Probability, and Python-based scientific computing. Three option choices are provided for each question, with exactly one of them being correct. To ensure the absence of data leakage, all questions are newly created and not sourced from existing materials. The validity of questions is verified through a procedure based on cross-LLM disagreement, followed by manual expert review. By adopting a multiple-choice format, our dataset enables objective, reproducible, and bias-free evaluation through lm_eval library. Baseline results with state-of-the-art LLMs indicate that advanced computational mathematical reasoning remains a significant challenge. We release CompMath-MCQ at the following link: https://github.com/biancaraimondi/CompMath-MCQ.git

Post-training quantization (PTQ) is a cornerstone for efficiently deploying large language models (LLMs), where a small calibration set critically affects quantization performance. However, conventional practices rely on random sequences of fixed length, overlooking the variable-length nature of LLM inputs. Input length directly influences the activation distribution and, consequently, the weight importance captured by the Hessian, which in turn affects quantization outcomes. As a result, Hessian estimates derived from fixed-length calibration may fail to represent the true importance of weights across diverse input scenarios. We propose MaCa (Matryoshka Calibration), a simple yet effective method for length-aware Hessian construction. MaCa (i) incorporates multi-scale sequence length information into Hessian estimation and (ii) regularizes each sequence as an independent sample, yielding a more stable and fruitful Hessian for accurate quantization. Experiments on state-of-the-art LLMs (e.g., Qwen3, Gemma3, LLaMA3) demonstrate that MaCa consistently improves accuracy under low bit quantization, offering a lightweight enhancement compatible with existing PTQ frameworks. To the best of our knowledge, this is the first work to systematically highlight the role of multi-scale calibration in LLM quantization.

Autonomous agent applications powered by large language models (LLMs) have recently risen to prominence as effective tools for addressing complex real-world tasks. At their core, agentic workflows rely on LLMs to plan and execute the use of tools and external Application Programming Interfaces (APIs) in sequence to arrive at the answer to a user's request. Various benchmarks and leaderboards have emerged to evaluate an LLM's capabilities for tool and API use; however, most of these evaluations only track single or multiple isolated API calling capabilities. In this paper, we present NESTFUL, a benchmark to evaluate LLMs on nested sequences of API calls, i.e., sequences where the output of one API call is passed as input to a subsequent call. NESTFUL has a total of 300 human annotated samples divided into two types - executable and non-executable. The executable samples are curated manually by crawling Rapid-APIs whereas the non-executable samples are hand picked by human annotators from data synthetically generated using an LLM. We evaluate state-of-the-art LLMs with function calling abilities on NESTFUL. Our results show that most models do not perform well on nested APIs in NESTFUL as compared to their performance on the simpler problem settings available in existing benchmarks.

Scaling high-quality tutoring remains a major challenge in education. Due to growing demand, many platforms employ novice tutors who, unlike experienced educators, struggle to address student mistakes and thus fail to seize prime learning opportunities. Our work explores the potential of large language models (LLMs) to close the novice-expert knowledge gap in remediating math mistakes. We contribute Bridge, a method that uses cognitive task analysis to translate an expert's latent thought process into a decision-making model for remediation. This involves an expert identifying (A) the student's error, (B) a remediation strategy, and (C) their intention before generating a response. We construct a dataset of 700 real tutoring conversations, annotated by experts with their decisions. We evaluate state-of-the-art LLMs on our dataset and find that the expert's decision-making model is critical for LLMs to close the gap: responses from GPT4 with expert decisions (e.g., "simplify the problem") are +76% more preferred than without. Additionally, context-sensitive decisions are critical to closing pedagogical gaps: random decisions decrease GPT4's response quality by -97% than expert decisions. Our work shows the potential of embedding expert thought processes in LLM generations to enhance their capability to bridge novice-expert knowledge gaps. Our dataset and code can be found at: https://github.com/rosewang2008/bridge.

Scientific research relies on accurate citation for attribution and integrity, yet large language models (LLMs) introduce a new risk: fabricated references that appear plausible but correspond to no real publications. Such hallucinated citations have already been observed in submissions and accepted papers at major machine learning venues, exposing vulnerabilities in peer review. Meanwhile, rapidly growing reference lists make manual verification impractical, and existing automated tools remain fragile to noisy and heterogeneous citation formats and lack standardized evaluation. We present the first comprehensive benchmark and detection framework for hallucinated citations in scientific writing. Our multi-agent verification pipeline decomposes citation checking into claim extraction, evidence retrieval, passage matching, reasoning, and calibrated judgment to assess whether a cited source truly supports its claim. We construct a large-scale human-validated dataset across domains and define unified metrics for citation faithfulness and evidence alignment. Experiments with state-of-the-art LLMs reveal substantial citation errors and show that our framework significantly outperforms prior methods in both accuracy and interpretability. This work provides the first scalable infrastructure for auditing citations in the LLM era and practical tools to improve the trustworthiness of scientific references.

Large Language models are revolutionizing the conversational recommender systems through their impressive capabilities in instruction comprehension, reasoning, and human interaction. A core factor underlying effective recommendation dialogue is the ability to infer and reason about users' mental states (such as desire, intention, and belief), a cognitive capacity commonly referred to as Theory of Mind. Despite growing interest in evaluating ToM in LLMs, current benchmarks predominantly rely on synthetic narratives inspired by Sally-Anne test, which emphasize physical perception and fail to capture the complexity of mental state inference in realistic conversational settings. Moreover, existing benchmarks often overlook a critical component of human ToM: behavioral prediction, the ability to use inferred mental states to guide strategic decision-making and select appropriate conversational actions for future interactions. To better align LLM-based ToM evaluation with human-like social reasoning, we propose RecToM, a novel benchmark for evaluating ToM abilities in recommendation dialogues. RecToM focuses on two complementary dimensions: Cognitive Inference and Behavioral Prediction. The former focus on understanding what has been communicated by inferring the underlying mental states. The latter emphasizes what should be done next, evaluating whether LLMs can leverage these inferred mental states to predict, select, and assess appropriate dialogue strategies. Extensive experiments on state-of-the-art LLMs demonstrate that RecToM poses a significant challenge. While the models exhibit partial competence in recognizing mental states, they struggle to maintain coherent, strategic ToM reasoning throughout dynamic recommendation dialogues, particularly in tracking evolving intentions and aligning conversational strategies with inferred mental states.

Large Language Models (LLMs) are increasingly used to automate hardware design tasks, including the generation of Verilog code. While early benchmarks focus primarily on functional correctness, efficient hardware design demands additional optimization for synthesis metrics such as area, delay, and power. Existing benchmarks fall short in evaluating these aspects comprehensively: they often lack optimized baselines or testbenches for verification. To address these gaps, we present Pluto, a benchmark and evaluation framework designed to assess the efficiency of LLM-generated Verilog designs. Pluto presents a comprehensive evaluation set of 114 problems with self-checking testbenches and multiple Pareto-optimal reference implementations. Experimental results show that state-of-the-art LLMs can achieve high functional correctness, reaching 78.3\% at pass@1, but their synthesis efficiency still lags behind expert-crafted implementations, with area efficiency of 63.8\%, delay efficiency of 65.9\%, and power efficiency of 64.0\% at eff@1. This highlights the need for efficiency-aware evaluation frameworks such as Pluto to drive progress in hardware-focused LLM research.

Large Language Models (LLMs) have achieved human-level proficiency across diverse tasks, but their ability to perform rigorous mathematical problem solving remains an open challenge. In this work, we investigate a fundamental yet computationally intractable problem: determining whether a given multivariate polynomial is nonnegative. This problem, closely related to Hilbert's Seventeenth Problem, plays a crucial role in global polynomial optimization and has applications in various fields. First, we introduce SoS-1K, a meticulously curated dataset of approximately 1,000 polynomials, along with expert-designed reasoning instructions based on five progressively challenging criteria. Evaluating multiple state-of-the-art LLMs, we find that without structured guidance, all models perform only slightly above the random guess baseline 50%. However, high-quality reasoning instructions significantly improve accuracy, boosting performance up to 81%. Furthermore, our 7B model, SoS-7B, fine-tuned on SoS-1K for just 4 hours, outperforms the 671B DeepSeek-V3 and GPT-4o-mini in accuracy while only requiring 1.8% and 5% of the computation time needed for letters, respectively. Our findings highlight the potential of LLMs to push the boundaries of mathematical reasoning and tackle NP-hard problems.

Expert persona prompting -- assigning roles such as expert in math to language models -- is widely used for task improvement. However, prior work shows mixed results on its effectiveness, and does not consider when and why personas should improve performance. We analyze the literature on persona prompting for task improvement and distill three desiderata: 1) performance advantage of expert personas, 2) robustness to irrelevant persona attributes, and 3) fidelity to persona attributes. We then evaluate 9 state-of-the-art LLMs across 27 tasks with respect to these desiderata. We find that expert personas usually lead to positive or non-significant performance changes. Surprisingly, models are highly sensitive to irrelevant persona details, with performance drops of almost 30 percentage points. In terms of fidelity, we find that while higher education, specialization, and domain-relatedness can boost performance, their effects are often inconsistent or negligible across tasks. We propose mitigation strategies to improve robustness -- but find they only work for the largest, most capable models. Our findings underscore the need for more careful persona design and for evaluation schemes that reflect the intended effects of persona usage.

While Large Language Models (LLMs) demonstrate impressive reasoning capabilities, growing evidence suggests much of their success stems from memorized answer-reasoning patterns rather than genuine inference. In this work, we investigate a central question: are LLMs primarily anchored to final answers or to the textual pattern of reasoning chains? We propose a five-level answer-visibility prompt framework that systematically manipulates answer cues and probes model behavior through indirect, behavioral analysis. Experiments across state-of-the-art LLMs reveal a strong and consistent reliance on explicit answers. The performance drops by 26.90\% when answer cues are masked, even with complete reasoning chains. These findings suggest that much of the reasoning exhibited by LLMs may reflect post-hoc rationalization rather than true inference, calling into question their inferential depth. Our study uncovers the answer-anchoring phenomenon with rigorous empirical validation and underscores the need for a more nuanced understanding of what constitutes reasoning in LLMs.

The capabilities of Large Language Models (LLMs) in code generation, particularly for implementing target functionalities from natural language descriptions, have been extensively studied. As an alternative form of natural language, input-output examples (I/O examples) provide an accessible, unambiguous, and flexible way to describe functionalities, but the diversity, sparseness, and incompleteness of I/O examples also place challenges on understanding and implementing requirements. Therefore, generating code from input-output examples (i.e., example-based code generation) provides a new perspective, allowing us to evaluate LLMs' capability to infer target functionalities from limited information and to process new-form requirements. However, related research about LLMs in example-based code generation remains largely unexplored. To fill this gap, this paper presents the first comprehensive study on example-based code generation using LLMs. To address the incorrectness caused by the incompleteness of I/O examples, we adopt an iterative evaluation framework and formalize the objective of example-based code generation as two sequential sub-objectives: generating code conforming to given examples and generating code that successfully implements the target functionalities from (iteratively) given examples. We assess six state-of-the-art LLMs using a new benchmark of 168 diverse target functionalities. The results demonstrate that when requirements were described using iterative I/O examples rather than natural language, the LLMs' score decreased by over 60%, indicating that example-based code generation remains challenging for the evaluated LLMs. More interestingly, the vast majority (even over 95%) of successfully implemented functionalities are achieved in the first round of iterations, suggesting that the LLMs struggle to effectively utilize the iteratively supplemented requirements.

At the core of Deep Research is knowledge mining, the task of extracting structured information from massive unstructured text in response to user instructions. Large language models (LLMs) excel at interpreting such instructions but are prohibitively expensive to deploy at scale, while traditional pipelines of classifiers and extractors remain efficient yet brittle and unable to generalize to new tasks. We introduce Falconer, a collaborative framework that combines the agentic reasoning of LLMs with lightweight proxy models for scalable knowledge mining. In Falconer, LLMs act as planners, decomposing user instructions into executable pipelines, and as annotators, generating supervision to train small proxies. The framework unifies classification and extraction into two atomic operations, get label and get span, enabling a single instruction-following model to replace multiple task-specific components. To evaluate the consistency between proxy models incubated by Falconer and annotations provided by humans and large models, we construct new benchmarks covering both planning and end-to-end execution. Experiments show that Falconer closely matches state-of-the-art LLMs in instruction-following accuracy while reducing inference cost by up to 90% and accelerating large-scale knowledge mining by more than 20x, offering an efficient and scalable foundation for Deep Research.

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.

Robustly evaluating the long-form storytelling capabilities of Large Language Models (LLMs) remains a significant challenge, as existing benchmarks often lack the necessary scale, diversity, or objective measures. To address this, we introduce WebNovelBench, a novel benchmark specifically designed for evaluating long-form novel generation. WebNovelBench leverages a large-scale dataset of over 4,000 Chinese web novels, framing evaluation as a synopsis-to-story generation task. We propose a multi-faceted framework encompassing eight narrative quality dimensions, assessed automatically via an LLM-as-Judge approach. Scores are aggregated using Principal Component Analysis and mapped to a percentile rank against human-authored works. Our experiments demonstrate that WebNovelBench effectively differentiates between human-written masterpieces, popular web novels, and LLM-generated content. We provide a comprehensive analysis of 24 state-of-the-art LLMs, ranking their storytelling abilities and offering insights for future development. This benchmark provides a scalable, replicable, and data-driven methodology for assessing and advancing LLM-driven narrative generation.

Large language models (LLMs) based on decoder-only transformers have demonstrated superior text understanding capabilities compared to CLIP and T5-series models. However, the paradigm for utilizing current advanced LLMs in text-to-image diffusion models remains to be explored. We observed an unusual phenomenon: directly using a large language model as the prompt encoder significantly degrades the prompt-following ability in image generation. We identified two main obstacles behind this issue. One is the misalignment between the next token prediction training in LLM and the requirement for discriminative prompt features in diffusion models. The other is the intrinsic positional bias introduced by the decoder-only architecture. To deal with this issue, we propose a novel framework to fully harness the capabilities of LLMs. Through the carefully designed usage guidance, we effectively enhance the text representation capability for prompt encoding and eliminate its inherent positional bias. This allows us to integrate state-of-the-art LLMs into the text-to-image generation model flexibly. Furthermore, we also provide an effective manner to fuse multiple LLMs into our framework. Considering the excellent performance and scaling capabilities demonstrated by the transformer architecture, we further design an LLM-Infused Diffusion Transformer (LI-DiT) based on the framework. We conduct extensive experiments to validate LI-DiT across model size and data size. Benefiting from the inherent ability of the LLMs and our innovative designs, the prompt understanding performance of LI-DiT easily surpasses state-of-the-art open-source models as well as mainstream closed-source commercial models including Stable Diffusion 3, DALL-E 3, and Midjourney V6. The powerful LI-DiT-10B will be available after further optimization and security checks.

The rapid evolution of software libraries presents a significant challenge for code generation models, which must adapt to frequent version updates while maintaining compatibility with previous versions. Existing code completion benchmarks often overlook this dynamic aspect, and the one that does consider it relies on static code prediction tasks without execution-based evaluation, offering a limited perspective on a model's practical usability. To address this gap, we introduce \GitChameleon{}, a novel, manually curated dataset comprising 116 Python code completion problems, each conditioned on specific library versions and accompanied by executable unit tests. is designed to rigorously assess the ability of modern large language models (LLMs) to generate version-specific code that is not only syntactically correct but also functionally accurate upon execution. Our comprehensive evaluations reveal that state-of-the-art LLMs struggle with this task; for instance, GPT-4o achieves a pass@10 of only 39.9\% (43.7\% when provided with error feedback), highlighting the complexity of the problem and the limitations of current models. By providing an execution-based benchmark that emphasizes the dynamic nature of code libraries, serves as a critical tool to advance the development of more adaptable and reliable code generation models. For facilitation for further exploration of version-conditioned code generation, we make our code repository publicly accessible at https://github.com/NizarIslah/GitChameleon.

While large language models (LLMs) with reasoning capabilities are progressing rapidly on high-school math competitions and coding, can they reason effectively through complex, open-ended challenges found in frontier physics research? And crucially, what kinds of reasoning tasks do physicists want LLMs to assist with? To address these questions, we present the CritPt (Complex Research using Integrated Thinking - Physics Test, pronounced "critical point"), the first benchmark designed to test LLMs on unpublished, research-level reasoning tasks that broadly covers modern physics research areas, including condensed matter, quantum physics, atomic, molecular & optical physics, astrophysics, high energy physics, mathematical physics, statistical physics, nuclear physics, nonlinear dynamics, fluid dynamics and biophysics. CritPt consists of 71 composite research challenges designed to simulate full-scale research projects at the entry level, which are also decomposed to 190 simpler checkpoint tasks for more fine-grained insights. All problems are newly created by 50+ active physics researchers based on their own research. Every problem is hand-curated to admit a guess-resistant and machine-verifiable answer and is evaluated by an automated grading pipeline heavily customized for advanced physics-specific output formats. We find that while current state-of-the-art LLMs show early promise on isolated checkpoints, they remain far from being able to reliably solve full research-scale challenges: the best average accuracy among base models is only 4.0% , achieved by GPT-5 (high), moderately rising to around 10% when equipped with coding tools. Through the realistic yet standardized evaluation offered by CritPt, we highlight a large disconnect between current model capabilities and realistic physics research demands, offering a foundation to guide the development of scientifically grounded AI tools.

Recent studies in large language models (LLMs) shed light on their multilingual ability and safety, beyond conventional tasks in language modeling. Still, current benchmarks reveal their inability to comprehensively evaluate them and are excessively dependent on manual annotations. In this paper, we introduce code-switching red-teaming (CSRT), a simple yet effective red-teaming technique that simultaneously tests multilingual understanding and safety of LLMs. We release the CSRT dataset, which comprises 315 code-switching queries combining up to 10 languages and eliciting a wide range of undesirable behaviors. Through extensive experiments with ten state-of-the-art LLMs, we demonstrate that CSRT significantly outperforms existing multilingual red-teaming techniques, achieving 46.7% more attacks than existing methods in English. We analyze the harmful responses toward the CSRT dataset concerning various aspects under ablation studies with 16K samples, including but not limited to scaling laws, unsafe behavior categories, and input conditions for optimal data generation. Additionally, we validate the extensibility of CSRT, by generating code-switching attack prompts with monolingual data.

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.

This paper explores the rapid development of a telephone call summarization system utilizing large language models (LLMs). Our approach involves initial experiments with prompting existing LLMs to generate summaries of telephone conversations, followed by the creation of a tailored synthetic training dataset utilizing stronger frontier models. We place special focus on the diversity of the generated data and on the ability to control the length of the generated summaries to meet various use-case specific requirements. The effectiveness of our method is evaluated using two state-of-the-art LLM-as-a-judge-based evaluation techniques to ensure the quality and relevance of the summaries. Our results show that fine-tuned Llama-2-7B-based summarization model performs on-par with GPT-4 in terms of factual accuracy, completeness and conciseness. Our findings demonstrate the potential for quickly bootstrapping a practical and efficient call summarization system.

Recent progress in large language models (LLMs) has showcased impressive proficiency in numerous Arabic natural language processing (NLP) applications. Nevertheless, their effectiveness in Arabic medical NLP domains has received limited investigation. This research examines the degree to which state-of-the-art LLMs demonstrate and articulate healthcare knowledge in Arabic, assessing their capabilities across a varied array of Arabic medical tasks. We benchmark several LLMs using a medical dataset proposed in the Arabic NLP AraHealthQA challenge in MedArabiQ2025 track. Various base LLMs were assessed on their ability to accurately provide correct answers from existing choices in multiple-choice questions (MCQs) and fill-in-the-blank scenarios. Additionally, we evaluated the capacity of LLMs in answering open-ended questions aligned with expert answers. Our results reveal significant variations in correct answer prediction accuracy and low variations in semantic alignment of generated answers, highlighting both the potential and limitations of current LLMs in Arabic clinical contexts. Our analysis shows that for MCQs task, the proposed majority voting solution, leveraging three base models (Gemini Flash 2.5, Gemini Pro 2.5, and GPT o3), outperforms others, achieving up to 77% accuracy and securing first place overall in the Arahealthqa 2025 shared task-track 2 (sub-task 1) challenge. Moreover, for the open-ended questions task, several LLMs were able to demonstrate excellent performance in terms of semantic alignment and achieve a maximum BERTScore of 86.44%.

This paper introduces the counter-intuitive generalization results of overfitting pre-trained large language models (LLMs) on very small datasets. In the setting of open-ended text generation, it is well-documented that LLMs tend to generate repetitive and dull sequences, a phenomenon that is especially apparent when generating using greedy decoding. This issue persists even with state-of-the-art LLMs containing billions of parameters, trained via next-token prediction on large datasets. We find that by further fine-tuning these models to achieve a near-zero training loss on a small set of samples -- a process we refer to as hyperfitting -- the long-sequence generative capabilities are greatly enhanced. Greedy decoding with these Hyperfitted models even outperform Top-P sampling over long-sequences, both in terms of diversity and human preferences. This phenomenon extends to LLMs of various sizes, different domains, and even autoregressive image generation. We further find this phenomena to be distinctly different from that of Grokking and double descent. Surprisingly, our experiments indicate that hyperfitted models rarely fall into repeating sequences they were trained on, and even explicitly blocking these sequences results in high-quality output. All hyperfitted models produce extremely low-entropy predictions, often allocating nearly all probability to a single token.

Recent advancements in the safety of Large Language Models (LLMs) have primarily focused on mitigating attacks crafted in natural language or in common encryption techniques like Base64. However, new models which often possess better reasoning capabilities, open the door to new attack vectors that were previously non-existent in older models. This seems counter-intuitive at first glance, but these advanced models can decipher more complex cryptic queries that previous models could not, making them susceptible to attacks using such prompts. To exploit this vulnerability, we propose Attacks using Custom Encryptions (ACE), a novel method to jailbreak LLMs by leveraging custom encryption schemes. We evaluate the effectiveness of ACE on four state-of-the-art LLMs, achieving Attack Success Rates (ASR) of up to 66% on close-source models and 88% on open-source models. Building upon this, we introduce Layered Attacks using Custom Encryptions (LACE), which employs multiple layers of encryption through our custom ciphers to further enhance the ASR. Our findings demonstrate that LACE significantly enhances the ability to jailbreak LLMs, increasing the ASR of GPT-4o from 40% to 78%, a 38% improvement. Our results highlight that the advanced capabilities of LLMs introduce unforeseen vulnerabilities to complex attacks. Specifically complex and layered ciphers increase the chance of jailbreaking.

The rapid advancement of Large Language Models (LLMs) has outpaced traditional evaluation methods. Static benchmarks fail to capture the depth and breadth of LLM capabilities and eventually become obsolete, while most dynamic approaches either rely too heavily on LLM-based evaluation or remain constrained by predefined test sets. We introduce Prism, a flexible, dynamic benchmarking framework designed for comprehensive LLM assessment. Prism builds on three key components: (1) a tree-based state representation that models evaluation as a Markov Decision Process, (2) a Monte Carlo Tree Search algorithm adapted to uncover challenging evaluation scenarios, and (3) a multi-agent evaluation pipeline that enables simultaneous assessment of diverse capabilities. To ensure robust evaluation, Prism integrates structural measurements of tree exploration patterns with performance metrics across difficulty levels, providing detailed diagnostics of error patterns, test coverage, and solution approaches. Through extensive experiments on five state-of-the-art LLMs, we analyze how model architecture and scale influence code generation performance across varying task difficulties. Our results demonstrate Prism's effectiveness as a dynamic benchmark that evolves with model advancements while offering deeper insights into their limitations.

Large language models (LLMs) are increasingly deployed in medical contexts, raising critical concerns about safety, alignment, and susceptibility to adversarial manipulation. While prior benchmarks assess model refusal capabilities for harmful prompts, they often lack clinical specificity, graded harmfulness levels, and coverage of jailbreak-style attacks. We introduce CARES (Clinical Adversarial Robustness and Evaluation of Safety), a benchmark for evaluating LLM safety in healthcare. CARES includes over 18,000 prompts spanning eight medical safety principles, four harm levels, and four prompting styles: direct, indirect, obfuscated, and role-play, to simulate both malicious and benign use cases. We propose a three-way response evaluation protocol (Accept, Caution, Refuse) and a fine-grained Safety Score metric to assess model behavior. Our analysis reveals that many state-of-the-art LLMs remain vulnerable to jailbreaks that subtly rephrase harmful prompts, while also over-refusing safe but atypically phrased queries. Finally, we propose a mitigation strategy using a lightweight classifier to detect jailbreak attempts and steer models toward safer behavior via reminder-based conditioning. CARES provides a rigorous framework for testing and improving medical LLM safety under adversarial and ambiguous conditions.

Can Large Language Models (LLMs) simulate human behavior in complex environments? LLMs have recently been shown to exhibit advanced reasoning skills but much of NLP evaluation still relies on static benchmarks. Answering this requires evaluation environments that probe strategic reasoning in competitive, dynamic scenarios that involve long-term planning. We introduce AucArena, a novel simulation environment for evaluating LLMs within auctions, a setting chosen for being highly unpredictable and involving many skills related to resource and risk management, while also being easy to evaluate. We conduct several controlled simulations using state-of-the-art LLMs as bidding agents. We find that through simple prompting, LLMs do indeed demonstrate many of the skills needed for effectively engaging in auctions (e.g., managing budget, adhering to long-term goals and priorities), skills that we find can be sharpened by explicitly encouraging models to be adaptive and observe strategies in past auctions. These results are significant as they show the potential of using LLM agents to model intricate social dynamics, especially in competitive settings. However, we also observe considerable variability in the capabilities of individual LLMs. Notably, even our most advanced models (GPT-4) are occasionally surpassed by heuristic baselines and human agents, highlighting the potential for further improvements in the design of LLM agents and the important role that our simulation environment can play in further testing and refining agent architectures.

Large language models (LLMs) have demonstrated remarkable performance across various machine learning tasks. Yet the substantial memory footprint of LLMs significantly hinders their deployment. In this paper, we improve the accessibility of LLMs through BitMoD, an algorithm-hardware co-design solution that enables efficient LLM acceleration at low weight precision. On the algorithm side, BitMoD introduces fine-grained data type adaptation that uses a different numerical data type to quantize a group of (e.g., 128) weights. Through the careful design of these new data types, BitMoD is able to quantize LLM weights to very low precision (e.g., 4 bits and 3 bits) while maintaining high accuracy. On the hardware side, BitMoD employs a bit-serial processing element to easily support multiple numerical precisions and data types; our hardware design includes two key innovations: First, it employs a unified representation to process different weight data types, thus reducing the hardware cost. Second, it adopts a bit-serial dequantization unit to rescale the per-group partial sum with minimal hardware overhead. Our evaluation on six representative LLMs demonstrates that BitMoD significantly outperforms state-of-the-art LLM quantization and acceleration methods. For discriminative tasks, BitMoD can quantize LLM weights to 4-bit with <!0.5% accuracy loss on average. For generative tasks, BitMoD is able to quantize LLM weights to 3-bit while achieving better perplexity than prior LLM quantization scheme. Combining the superior model performance with an efficient accelerator design, BitMoD achieves an average of 1.69times and 1.48times speedups compared to prior LLM accelerators ANT and OliVe, respectively.

Large Language Models (LLMs) are increasingly being deployed in interactive contexts that involve direct user engagement, such as chatbots and writing assistants. These deployments are increasingly plagued by prompt injection and jailbreaking (collectively, prompt hacking), in which models are manipulated to ignore their original instructions and instead follow potentially malicious ones. Although widely acknowledged as a significant security threat, there is a dearth of large-scale resources and quantitative studies on prompt hacking. To address this lacuna, we launch a global prompt hacking competition, which allows for free-form human input attacks. We elicit 600K+ adversarial prompts against three state-of-the-art LLMs. We describe the dataset, which empirically verifies that current LLMs can indeed be manipulated via prompt hacking. We also present a comprehensive taxonomical ontology of the types of adversarial prompts.

Foundation models are becoming valuable tools in medicine. Yet despite their promise, the best way to leverage Large Language Models (LLMs) in complex medical tasks remains an open question. We introduce a novel multi-agent framework, named Medical Decision-making Agents (MDAgents) that helps address this gap by automatically assigning a collaboration structure to a team of LLMs. The assigned solo or group collaboration structure is tailored to the medical task at hand, emulating real-world medical decision-making processes adapted to tasks of varying complexities. We evaluate our framework and baseline methods using state-of-the-art LLMs across a suite of real-world medical knowledge and medical diagnosis benchmarks, including a comparison of LLMs' medical complexity classification against human physicians. MDAgents achieved the best performance in seven out of ten benchmarks on tasks requiring an understanding of medical knowledge and multi-modal reasoning, showing a significant improvement of up to 4.2% (p < 0.05) compared to previous methods' best performances. Ablation studies reveal that MDAgents effectively determines medical complexity to optimize for efficiency and accuracy across diverse medical tasks. Notably, the combination of moderator review and external medical knowledge in group collaboration resulted in an average accuracy improvement of 11.8%. Our code can be found at https://github.com/mitmedialab/MDAgents.

The rapid advancement of Large Language Models (LLMs) has demonstrated remarkable progress in complex reasoning tasks. However, a significant discrepancy persists between benchmark performances and real-world applications. We identify this gap as primarily stemming from current evaluation protocols and metrics, which inadequately capture the full spectrum of LLM capabilities, particularly in complex reasoning tasks where both accuracy and consistency are crucial. This work makes two key contributions. First, we introduce G-Pass@k, a novel evaluation metric that provides a continuous assessment of model performance across multiple sampling attempts, quantifying both the model's peak performance potential and its stability. Second, we present LiveMathBench, a dynamic benchmark comprising challenging, contemporary mathematical problems designed to minimize data leakage risks during evaluation. Through extensive experiments using G-Pass@k on state-of-the-art LLMs with LiveMathBench, we provide comprehensive insights into both their maximum capabilities and operational consistency. Our findings reveal substantial room for improvement in LLMs' "realistic" reasoning capabilities, highlighting the need for more robust evaluation methods. The benchmark and detailed results are available at: https://github.com/open-compass/GPassK.

Interacting with human via high-quality multi-turn dialogues is a key feature of large language models (LLMs). However, human-based evaluation of such capability involves intensive manual labor. This report provides a preliminary evaluation of existing large language models for human-style multi-turn chatting, through an LLM-based approach. We start from real-world human dialogues and keep the very first utterances as the ChatSEED. Then we prompt LLMs to generate a full multi-turn dialogue (tens of utterances) based on the ChatSEED, utterance by utterance. Finally, we adopt state-of-the-art LLMs (GPT-4, \etc) as the judge to evaluate the generated dialogues. With different evaluation protocols, we come to substantially identical conclusions. We find that GPT-4 can generate human-style multi-turn dialogues with impressive quality, significantly outperforms its counterparts. It's difficult for a discriminator to distinguish between GPT-4 generated dialogues and human dialogues. In contrast, other LLMs struggle to generate multi-turn dialogues of satisfactory quality due to poor instruction-following capability, tendency to generate lengthy utterances, or limited general capability. All data and codes will be provided in https://github.com/open-compass/BotChat/ and we hope they can serve as a valuable resource for evaluating multi-turn chatting capabilities of LLMs.

Large language models (LLMs) have achieved striking successes on many benchmarks, yet recent studies continue to expose fundamental weaknesses. In particular, tasks that require abstract reasoning remain challenging, often because they use representations such as grids, symbols, or visual patterns that differ from the natural language data LLMs are trained on. In this paper, we introduce Concept, a simple word-guessing board game, as a benchmark for probing abductive reasoning in a representation that is much closer to LLM pre-training data: natural language. Our results show that this game, easily solved by humans (with a success rate of over 90\%), is still very challenging for state-of-the-art LLMs (no model exceeds 40\% success rate). Specifically, we observe that LLMs struggle with interpreting other players' strategic intents, and with correcting initial hypotheses given sequential information updates. In addition, we extend the evaluation across multiple languages, and find that the LLM performance drops further in lower-resource languages (Dutch, French, and Spanish) compared to English.

Mixture-of-Experts (MoE) architectures face challenges such as high memory consumption and redundancy in experts. Pruning MoE can reduce network weights while maintaining model performance. Motivated by the recent observation of emergent large magnitude features in Large Language Models (LLM) and MoE routing policy, we propose MoE-Pruner, a method that prunes weights with the smallest magnitudes multiplied by the corresponding input activations and router weights, on each output neuron. Our pruning method is one-shot, requiring no retraining or weight updates. We evaluate our method on Mixtral-8x7B and Mixtral-8x22B across multiple language benchmarks. Experimental results show that our pruning method significantly outperforms state-of-the-art LLM pruning methods. Furthermore, our pruned MoE models can benefit from a pretrained teacher model through expert-wise knowledge distillation, improving performance post-pruning. Experimental results demonstrate that the Mixtral-8x7B model with 50% sparsity maintains 99% of the performance of the original model after the expert-wise knowledge distillation.

LLM-as-a-Judge has been widely adopted as an evaluation method and served as supervised rewards in model training. However, existing benchmarks for LLM-as-a-Judge are mainly relying on human-annotated ground truth, which introduces human bias that undermines the assessment of reliability and imposes scalability constraints. To overcome these limitations, we introduce Sage, a novel evaluation suite that assesses the quality of LLM judges without necessitating any human annotation. Inspired by axioms of rational choice theory, Sage introduces two new lenses for measuring LLM-as-a-Judge: local self-consistency (pair-wise preference stability) and global logical consistency (transitivity across a full set of preferences). We curate a dataset of 650 questions by combining structured benchmark problems with real-world user queries. Our experiments demonstrate both the stability of our metrics and their high correlation with supervised benchmarks like LLMBar and RewardBench2, confirming Sage's reliability as an evaluation suite for the robustness and accuracy of LLM-as-a-Judge. Based on Sage, we reveal that current state-of-the-art LLMs exhibit significant reliability problems when acting as judges in both scoring and pairwise settings; even the top-performing models, Gemini-2.5-Pro and GPT-5, fail to maintain consistent preferences in nearly a quarter of difficult cases. We attribute this to a new phenomenon called situational preference, which explains why explicit rubrics or criteria can help the model judge consistently across answer pairs. Our further analysis shows that finetuned LLM-as-a-Judge is a feasible method to boost performance, and the panel-based judge as well as deep reasoning can enhance the judging consistency. We also find substantial inconsistency in human judgments, which indicates that human annotation may not be a reliable gold standard.

The advent of large language models (LLMs) has transformed text-based services, enabling capabilities ranging from real-time translation to AI-driven chatbots. However, existing serving systems primarily focus on optimizing server-side aggregate metrics like token generation throughput, ignoring individual user experience with streamed text. As a result, under high and/or bursty load, a significant number of users can receive unfavorable service quality or poor Quality-of-Experience (QoE). In this paper, we first formally define QoE of text streaming services, where text is delivered incrementally and interactively to users, by considering the end-to-end token delivery process throughout the entire interaction with the user. Thereafter, we propose Andes, a QoE-aware serving system that enhances user experience for LLM-enabled text streaming services. At its core, Andes strategically allocates contended GPU resources among multiple requests over time to optimize their QoE. Our evaluations demonstrate that, compared to the state-of-the-art LLM serving systems like vLLM, Andes improves the average QoE by up to 3.2times under high request rate, or alternatively, it attains up to 1.6times higher request rate while preserving high QoE.

Theory-of-Mind (ToM) is a fundamental psychological capability that allows humans to understand and interpret the mental states of others. Humans infer others' thoughts by integrating causal cues and indirect clues from broad contextual information, often derived from past interactions. In other words, human ToM heavily relies on the understanding about the backgrounds and life stories of others. Unfortunately, this aspect is largely overlooked in existing benchmarks for evaluating machines' ToM capabilities, due to their usage of short narratives without global backgrounds. In this paper, we verify the importance of understanding long personal backgrounds in ToM and assess the performance of LLMs in such realistic evaluation scenarios. To achieve this, we introduce a novel benchmark, CharToM-QA, comprising 1,035 ToM questions based on characters from classic novels. Our human study reveals a significant disparity in performance: the same group of educated participants performs dramatically better when they have read the novels compared to when they have not. In parallel, our experiments on state-of-the-art LLMs, including the very recent o1 model, show that LLMs still perform notably worse than humans, despite that they have seen these stories during pre-training. This highlights the limitations of current LLMs in capturing the nuanced contextual information required for ToM reasoning.

Formulating optimization problems for industrial applications demands significant manual effort and domain expertise. While Large Language Models (LLMs) show promise in automating this process, evaluating their performance remains difficult due to the absence of robust metrics. Existing solver-based approaches often face inconsistency, infeasibility issues, and high computational costs. To address these issues, we propose ORGEval, a graph-theoretic evaluation framework for assessing LLMs' capabilities in formulating linear and mixed-integer linear programs. ORGEval represents optimization models as graphs, reducing equivalence detection to graph isomorphism testing. We identify and prove a sufficient condition, when the tested graphs are symmetric decomposable (SD), under which the Weisfeiler-Lehman (WL) test is guaranteed to correctly detect isomorphism. Building on this, ORGEval integrates a tailored variant of the WL-test with an SD detection algorithm to evaluate model equivalence. By focusing on structural equivalence rather than instance-level configurations, ORGEval is robust to numerical variations. Experimental results show that our method can successfully detect model equivalence and produce 100\% consistent results across random parameter configurations, while significantly outperforming solver-based methods in runtime, especially on difficult problems. Leveraging ORGEval, we construct the Bench4Opt dataset and benchmark state-of-the-art LLMs on optimization modeling. Our results reveal that although optimization modeling remains challenging for all LLMs, DeepSeek-V3 and Claude-Opus-4 achieve the highest accuracies under direct prompting, outperforming even leading reasoning models.

LLMs demonstrate promising performance in software vulnerability detection after fine-tuning. However, it remains unclear whether these gains reflect a genuine understanding of vulnerability root causes or merely an exploitation of functional patterns. In this paper, we identify a critical failure mode termed the "semantic trap," where fine-tuned LLMs achieve high detection scores by associating certain functional domains with vulnerability likelihood rather than reasoning about the underlying security semantics. To systematically evaluate this phenomenon, we propose TrapEval, a comprehensive evaluation framework designed to disentangle vulnerability root cause from functional pattern. TrapEval introduces two complementary datasets derived from real-world open-source projects: V2N, which pairs vulnerable code with unrelated benign code, and V2P, which pairs vulnerable code with its corresponding patched version, forcing models to distinguish near-identical code that differs only in subtle security-critical logic. Using TrapEval, we fine-tune five representative state-of-the-art LLMs across three model families and evaluate them under cross-dataset testing, semantic-preserving perturbations, and varying degrees of semantic gap measured by CodeBLEU. Our empirical results reveal that, despite improvements in metrics, fine-tuned LLMs consistently struggle to distinguish vulnerable code from its patched counterpart, exhibit severe robustness degradation under minor semantic-preserving transformations, and rely heavily on functional-context shortcuts when the semantic gap is small. These findings provide strong evidence that current fine-tuning practices often fail to impart true vulnerability reasoning. Our findings serve as a wake-up call: high benchmark scores on traditional datasets may be illusory, masking the model's inability to understand the true causal logic of vulnerabilities.

While Large Language Models (LLMs) display versatile functionality, they continue to generate harmful, biased, and toxic content, as demonstrated by the prevalence of human-designed jailbreaks. In this work, we present Tree of Attacks with Pruning (TAP), an automated method for generating jailbreaks that only requires black-box access to the target LLM. TAP utilizes an LLM to iteratively refine candidate (attack) prompts using tree-of-thoughts reasoning until one of the generated prompts jailbreaks the target. Crucially, before sending prompts to the target, TAP assesses them and prunes the ones unlikely to result in jailbreaks. Using tree-of-thought reasoning allows TAP to navigate a large search space of prompts and pruning reduces the total number of queries sent to the target. In empirical evaluations, we observe that TAP generates prompts that jailbreak state-of-the-art LLMs (including GPT4 and GPT4-Turbo) for more than 80% of the prompts using only a small number of queries. This significantly improves upon the previous state-of-the-art black-box method for generating jailbreaks.

We demonstrate the first cross-domain cross-layer level-4 autonomous optical network via a multi-AI-agent system. Field trials show 98 percent task completion rate across the distributed AI training lifecycle-3.2x higher than single agents using state-of-the-art LLMs.

Large language models (LLMs) are capable of solving a wide range of tasks, yet they have struggled with reasoning. To address this, we propose Additional Logic Training (ALT), which aims to enhance LLMs' reasoning capabilities by program-generated logical reasoning samples. We first establish principles for designing high-quality samples by integrating symbolic logic theory and previous empirical insights. Then, based on these principles, we construct a synthetic corpus named Formal Logic Deduction Diverse (FLD^{times 2}), comprising numerous samples of multi-step deduction with unknown facts, diverse reasoning rules, diverse linguistic expressions, and challenging distractors. Finally, we empirically show that ALT on FLD^{times2} substantially enhances the reasoning capabilities of state-of-the-art LLMs, including LLaMA-3.1-70B. Improvements include gains of up to 30 points on logical reasoning benchmarks, up to 10 points on math and coding benchmarks, and 5 points on the benchmark suite BBH.

We are releasing a new suite of security benchmarks for LLMs, CYBERSECEVAL 3, to continue the conversation on empirically measuring LLM cybersecurity risks and capabilities. CYBERSECEVAL 3 assesses 8 different risks across two broad categories: risk to third parties, and risk to application developers and end users. Compared to previous work, we add new areas focused on offensive security capabilities: automated social engineering, scaling manual offensive cyber operations, and autonomous offensive cyber operations. In this paper we discuss applying these benchmarks to the Llama 3 models and a suite of contemporaneous state-of-the-art LLMs, enabling us to contextualize risks both with and without mitigations in place.

Large language models (LLMs) are helping millions of users write texts about diverse issues, and in doing so expose users to different ideas and perspectives. This creates concerns about issue bias, where an LLM tends to present just one perspective on a given issue, which in turn may influence how users think about this issue. So far, it has not been possible to measure which issue biases LLMs actually manifest in real user interactions, making it difficult to address the risks from biased LLMs. Therefore, we create IssueBench: a set of 2.49m realistic prompts for measuring issue bias in LLM writing assistance, which we construct based on 3.9k templates (e.g. "write a blog about") and 212 political issues (e.g. "AI regulation") from real user interactions. Using IssueBench, we show that issue biases are common and persistent in state-of-the-art LLMs. We also show that biases are remarkably similar across models, and that all models align more with US Democrat than Republican voter opinion on a subset of issues. IssueBench can easily be adapted to include other issues, templates, or tasks. By enabling robust and realistic measurement, we hope that IssueBench can bring a new quality of evidence to ongoing discussions about LLM biases and how to address them.

Detecting logical fallacies in texts can help users spot argument flaws, but automating this detection is not easy. Manually annotating fallacies in large-scale, real-world text data to create datasets for developing and validating detection models is costly. This paper introduces CoCoLoFa, the largest known logical fallacy dataset, containing 7,706 comments for 648 news articles, with each comment labeled for fallacy presence and type. We recruited 143 crowd workers to write comments embodying specific fallacy types (e.g., slippery slope) in response to news articles. Recognizing the complexity of this writing task, we built an LLM-powered assistant into the workers' interface to aid in drafting and refining their comments. Experts rated the writing quality and labeling validity of CoCoLoFa as high and reliable. BERT-based models fine-tuned using CoCoLoFa achieved the highest fallacy detection (F1=0.86) and classification (F1=0.87) performance on its test set, outperforming the state-of-the-art LLMs. Our work shows that combining crowdsourcing and LLMs enables us to more effectively construct datasets for complex linguistic phenomena that crowd workers find challenging to produce on their own.

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.

As Large Language Models (LLMs) and generative AI become more widespread, the content safety risks associated with their use also increase. We find a notable deficiency in high-quality content safety datasets and benchmarks that comprehensively cover a wide range of critical safety areas. To address this, we define a broad content safety risk taxonomy, comprising 13 critical risk and 9 sparse risk categories. Additionally, we curate AEGISSAFETYDATASET, a new dataset of approximately 26, 000 human-LLM interaction instances, complete with human annotations adhering to the taxonomy. We plan to release this dataset to the community to further research and to help benchmark LLM models for safety. To demonstrate the effectiveness of the dataset, we instruction-tune multiple LLM-based safety models. We show that our models (named AEGISSAFETYEXPERTS), not only surpass or perform competitively with the state-of-the-art LLM-based safety models and general purpose LLMs, but also exhibit robustness across multiple jail-break attack categories. We also show how using AEGISSAFETYDATASET during the LLM alignment phase does not negatively impact the performance of the aligned models on MT Bench scores. Furthermore, we propose AEGIS, a novel application of a no-regret online adaptation framework with strong theoretical guarantees, to perform content moderation with an ensemble of LLM content safety experts in deployment

The field of machine learning (ML) has gained widespread adoption, leading to a significant demand for adapting ML to specific scenarios, which is yet expensive and non-trivial. The predominant approaches towards the automation of solving ML tasks (e.g., AutoML) are often time consuming and hard to understand for human developers. In contrast, though human engineers have the incredible ability to understand tasks and reason about solutions, their experience and knowledge are often sparse and difficult to utilize by quantitative approaches. In this paper, we aim to bridge the gap between machine intelligence and human knowledge by introducing a novel framework MLCopilot, which leverages the state-of-the-art LLMs to develop ML solutions for novel tasks. We showcase the possibility of extending the capability of LLMs to comprehend structured inputs and perform thorough reasoning for solving novel ML tasks. And we find that, after some dedicated design, the LLM can (i) observe from the existing experiences of ML tasks and (ii) reason effectively to deliver promising results for new tasks. The solution generated can be used directly to achieve high levels of competitiveness.

The democratization of AI is currently hindered by the immense computational costs required to train Large Language Models (LLMs) for low-resource languages. This paper presents Persian-Phi, a 3.8B parameter model that challenges the assumption that robust multilingual capabilities require massive model sizes or multilingual baselines. We demonstrate how Microsoft Phi-3 Mini -- originally a monolingual English model -- can be effectively adapted to Persian through a novel, resource-efficient curriculum learning pipeline. Our approach employs a unique "warm-up" stage using bilingual narratives (Tiny Stories) to align embeddings prior to heavy training, followed by continual pretraining and instruction tuning via Parameter-Efficient Fine-Tuning (PEFT). Despite its compact size, Persian-Phi achieves competitive results on Open Persian LLM Leaderboard in HuggingFace. Our findings provide a validated, scalable framework for extending the reach of state-of-the-art LLMs to underrepresented languages with minimal hardware resources. The Persian-Phi model is publicly available at https://huggingface.co/amirakhlaghiqqq/PersianPhi.

Post-Training Quantization (PTQ) enhances the efficiency of Large Language Models (LLMs) by enabling faster operation and compatibility with more accessible hardware through reduced memory usage, at the cost of small performance drops. We explore the role of calibration sets in PTQ, specifically their effect on hidden activations in various notable open-source LLMs. Calibration sets are crucial for evaluating activation magnitudes and identifying outliers, which can distort the quantization range and negatively impact performance. Our analysis reveals a marked contrast in quantization effectiveness across models. The older OPT model, upon which much of the quantization literature is based, shows significant performance deterioration and high susceptibility to outliers with varying calibration sets. In contrast, newer models like Llama-2 7B, Llama-3 8B, Command-R 35B, and Mistral 7B demonstrate strong robustness, with Mistral 7B showing near-immunity to outliers and stable activations. These findings suggest a shift in PTQ strategies might be needed. As advancements in pre-training methods reduce the relevance of outliers, there is an emerging need to reassess the fundamentals of current quantization literature. The emphasis should pivot towards optimizing inference speed, rather than primarily focusing on outlier preservation, to align with the evolving characteristics of state-of-the-art LLMs.

When using Large Language Models (LLMs) to support Knowledge Graph Engineering (KGE), one of the first indications when searching for an appropriate model is its size. According to the scaling laws, larger models typically show higher capabilities. However, in practice, resource costs are also an important factor and thus it makes sense to consider the ratio between model performance and costs. The LLM-KG-Bench framework enables the comparison of LLMs in the context of KGE tasks and assesses their capabilities of understanding and producing KGs and KG queries. Based on a dataset created in an LLM-KG-Bench run covering 26 open state-of-the-art LLMs, we explore the model size scaling laws specific to KGE tasks. In our analyses, we assess how benchmark scores evolve between different model size categories. Additionally, we inspect how the general score development of single models and families of models correlates to their size. Our analyses revealed that, with a few exceptions, the model size scaling laws generally also apply to the selected KGE tasks. However, in some cases, plateau or ceiling effects occurred, i.e., the task performance did not change much between a model and the next larger model. In these cases, smaller models could be considered to achieve high cost-effectiveness. Regarding models of the same family, sometimes larger models performed worse than smaller models of the same family. These effects occurred only locally. Hence it is advisable to additionally test the next smallest and largest model of the same family.

Despite advances in scientific AI, a coherent framework for Scientific General Intelligence (SGI)-the ability to autonomously conceive, investigate, and reason across scientific domains-remains lacking. We present an operational SGI definition grounded in the Practical Inquiry Model (PIM: Deliberation, Conception, Action, Perception) and operationalize it via four scientist-aligned tasks: deep research, idea generation, dry/wet experiments, and experimental reasoning. SGI-Bench comprises over 1,000 expert-curated, cross-disciplinary samples inspired by Science's 125 Big Questions, enabling systematic evaluation of state-of-the-art LLMs. Results reveal gaps: low exact match (10--20%) in deep research despite step-level alignment; ideas lacking feasibility and detail; high code executability but low execution result accuracy in dry experiments; low sequence fidelity in wet protocols; and persistent multimodal comparative-reasoning challenges. We further introduce Test-Time Reinforcement Learning (TTRL), which optimizes retrieval-augmented novelty rewards at inference, enhancing hypothesis novelty without reference answer. Together, our PIM-grounded definition, workflow-centric benchmark, and empirical insights establish a foundation for AI systems that genuinely participate in scientific discovery.

Large language models (LLMs) face significant challenges with needle-in-a-haystack tasks, where relevant information ("the needle") must be drawn from a large pool of irrelevant context ("the haystack"). Previous studies have highlighted positional bias and distractor quantity as critical factors affecting model performance, yet the influence of gold context size has received little attention. We address this gap by systematically studying how variations in gold context length impact LLM performance on long-context question answering tasks. Our experiments reveal that LLM performance drops sharply when the gold context is shorter, i.e., smaller gold contexts consistently degrade model performance and amplify positional sensitivity, posing a major challenge for agentic systems that must integrate scattered, fine-grained information of varying lengths. This pattern holds across three diverse domains (general knowledge, biomedical reasoning, and mathematical reasoning) and seven state-of-the-art LLMs of various sizes and architectures. Our work provides clear insights to guide the design of robust, context-aware LLM-driven systems.

To enable Large Language Models (LLMs) to function as conscious agents with generalizable reasoning capabilities, it is crucial that they possess the reasoning ability to comprehend situational changes (transitions) in distribution triggered by environmental factors or actions from other agents. Despite its fundamental significance, this ability remains underexplored due to the complexity of modeling infinite possible changes in an event and their associated distributions, coupled with the lack of benchmark data with situational transitions. Addressing these gaps, we propose a novel formulation of reasoning with distributional changes as a three-step discriminative process, termed as MetAphysical ReaSoning. We then introduce the first-ever benchmark, MARS, comprising three tasks corresponding to each step. These tasks systematically assess LLMs' capabilities in reasoning the plausibility of (i) changes in actions, (ii) states caused by changed actions, and (iii) situational transitions driven by changes in action. Extensive evaluations with 20 (L)LMs of varying sizes and methods indicate that all three tasks in this process pose significant challenges, even for state-of-the-art LLMs and LMs after fine-tuning. Further analyses reveal potential causes for the underperformance of LLMs and demonstrate that pre-training them on large-scale conceptualization taxonomies can potentially enhance their metaphysical reasoning capabilities. Our data and models are publicly accessible at https://github.com/HKUST-KnowComp/MARS.

Natural language explanations represent a proxy for evaluating explanation-based and multi-step Natural Language Inference (NLI) models. However, assessing the validity of explanations for NLI is challenging as it typically involves the crowd-sourcing of apposite datasets, a process that is time-consuming and prone to logical errors. To address existing limitations, this paper investigates the verification and refinement of natural language explanations through the integration of Large Language Models (LLMs) and Theorem Provers (TPs). Specifically, we present a neuro-symbolic framework, named Explanation-Refiner, that integrates TPs with LLMs to generate and formalise explanatory sentences and suggest potential inference strategies for NLI. In turn, the TP is employed to provide formal guarantees on the logical validity of the explanations and to generate feedback for subsequent improvements. We demonstrate how Explanation-Refiner can be jointly used to evaluate explanatory reasoning, autoformalisation, and error correction mechanisms of state-of-the-art LLMs as well as to automatically enhance the quality of explanations of variable complexity in different domains.

Large language models (LLMs) are rapidly transitioning into medical clinical use, yet their reliability under realistic, multi-turn interactions remains poorly understood. Existing evaluation frameworks typically assess single-turn question answering under idealized conditions, overlooking the complexities of medical consultations where conflicting input, misleading context, and authority influence are common. We introduce MedQA-Followup, a framework for systematically evaluating multi-turn robustness in medical question answering. Our approach distinguishes between shallow robustness (resisting misleading initial context) and deep robustness (maintaining accuracy when answers are challenged across turns), while also introducing an indirect-direct axis that separates contextual framing (indirect) from explicit suggestion (direct). Using controlled interventions on the MedQA dataset, we evaluate five state-of-the-art LLMs and find that while models perform reasonably well under shallow perturbations, they exhibit severe vulnerabilities in multi-turn settings, with accuracy dropping from 91.2% to as low as 13.5% for Claude Sonnet 4. Counterintuitively, indirect, context-based interventions are often more harmful than direct suggestions, yielding larger accuracy drops across models and exposing a significant vulnerability for clinical deployment. Further compounding analyses reveal model differences, with some showing additional performance drops under repeated interventions while others partially recovering or even improving. These findings highlight multi-turn robustness as a critical but underexplored dimension for safe and reliable deployment of medical LLMs.

Refusal training is widely used to prevent LLMs from generating harmful, undesirable, or illegal outputs. We reveal a curious generalization gap in the current refusal training approaches: simply reformulating a harmful request in the past tense (e.g., "How to make a Molotov cocktail?" to "How did people make a Molotov cocktail?") is often sufficient to jailbreak many state-of-the-art LLMs. We systematically evaluate this method on Llama-3 8B, Claude-3.5 Sonnet, GPT-3.5 Turbo, Gemma-2 9B, Phi-3-Mini, GPT-4o mini, GPT-4o, and R2D2 models using GPT-3.5 Turbo as a reformulation model. For example, the success rate of this simple attack on GPT-4o increases from 1% using direct requests to 88% using 20 past tense reformulation attempts on harmful requests from JailbreakBench with GPT-4 as a jailbreak judge. Interestingly, we also find that reformulations in the future tense are less effective, suggesting that refusal guardrails tend to consider past historical questions more benign than hypothetical future questions. Moreover, our experiments on fine-tuning GPT-3.5 Turbo show that defending against past reformulations is feasible when past tense examples are explicitly included in the fine-tuning data. Overall, our findings highlight that the widely used alignment techniques -- such as SFT, RLHF, and adversarial training -- employed to align the studied models can be brittle and do not always generalize as intended. We provide code and jailbreak artifacts at https://github.com/tml-epfl/llm-past-tense.

Chain-of-Thought (CoT) reasoning enables Large Language Models (LLMs) to solve complex reasoning tasks by generating intermediate reasoning steps. However, most existing approaches focus on hard token decoding, which constrains reasoning within the discrete vocabulary space and may not always be optimal. While recent efforts explore continuous-space reasoning, they often suffer from catastrophic forgetting, limiting their applicability to state-of-the-art LLMs that already perform well in zero-shot settings with a proper instruction. To address this challenge, we propose a novel approach for continuous-space reasoning that does not require modifying the underlying LLM. Specifically, we employ a lightweight assistant model to generate instance-specific soft thought tokens speculatively as the initial chain of thoughts, which are then mapped into the LLM's representation space via a projection module. Experimental results on five reasoning benchmarks demonstrate that our method enhances LLM reasoning performance through supervised, parameter-efficient fine-tuning.

Personalization of Large Language Models (LLMs) often assumes users hold static preferences that reflect globally in all tasks. In reality, humans hold dynamic preferences that change depending on the context. As users interact with an LLM in various contexts, they naturally reveal their contextual preferences, which a model must infer and apply in future contexts to ensure alignment. To assess this, we introduce CUPID, a benchmark of 756 human-curated interaction session histories between users and LLM-based chat assistants. In each interaction session, the user provides a request in a specific context and expresses their preference through multi-turn feedback. Given a new user request and prior interaction sessions, our benchmark assesses whether LLMs can infer the preference relevant to this request and generate a response that satisfies this preference. With CUPID, we evaluated 10 open and proprietary LLMs, revealing that state-of-the-art LLMs struggle to infer preferences from multi-turn interactions and fail to discern what previous context is relevant to a new request -- under 50% precision and 65% recall. Our work highlights the need to advance LLM capabilities for more contextually personalized interactions and proposes CUPID as a resource to drive these improvements.

We propose Zero-Error Horizon (ZEH) for trustworthy LLMs, which represents the maximum range that a model can solve without any errors. While ZEH itself is simple, we demonstrate that evaluating the ZEH of state-of-the-art LLMs yields abundant insights. For example, by evaluating the ZEH of GPT-5.2, we found that GPT-5.2 cannot even compute the parity of a short string like 11000, and GPT-5.2 cannot determine whether the parentheses in ((((()))))) are balanced. This is surprising given the excellent capabilities of GPT-5.2. The fact that LLMs make mistakes on such simple problems serves as an important lesson when applying LLMs to safety-critical domains. By applying ZEH to Qwen2.5 and conducting detailed analysis, we found that while ZEH correlates with accuracy, the detailed behaviors differ, and ZEH provides clues about the emergence of algorithmic capabilities. Finally, while computing ZEH incurs significant computational cost, we discuss how to mitigate this cost by achieving up to one order of magnitude speedup using tree structures and online softmax.

The EU's Artificial Intelligence Act (AI Act) is a significant step towards responsible AI development, but lacks clear technical interpretation, making it difficult to assess models' compliance. This work presents COMPL-AI, a comprehensive framework consisting of (i) the first technical interpretation of the EU AI Act, translating its broad regulatory requirements into measurable technical requirements, with the focus on large language models (LLMs), and (ii) an open-source Act-centered benchmarking suite, based on thorough surveying and implementation of state-of-the-art LLM benchmarks. By evaluating 12 prominent LLMs in the context of COMPL-AI, we reveal shortcomings in existing models and benchmarks, particularly in areas like robustness, safety, diversity, and fairness. This work highlights the need for a shift in focus towards these aspects, encouraging balanced development of LLMs and more comprehensive regulation-aligned benchmarks. Simultaneously, COMPL-AI for the first time demonstrates the possibilities and difficulties of bringing the Act's obligations to a more concrete, technical level. As such, our work can serve as a useful first step towards having actionable recommendations for model providers, and contributes to ongoing efforts of the EU to enable application of the Act, such as the drafting of the GPAI Code of Practice.

Research on language technology for the development of medical applications is currently a hot topic in Natural Language Understanding and Generation. Thus, a number of large language models (LLMs) have recently been adapted to the medical domain, so that they can be used as a tool for mediating in human-AI interaction. While these LLMs display competitive performance on automated medical texts benchmarks, they have been pre-trained and evaluated with a focus on a single language (English mostly). This is particularly true of text-to-text models, which typically require large amounts of domain-specific pre-training data, often not easily accessible for many languages. In this paper, we address these shortcomings by compiling, to the best of our knowledge, the largest multilingual corpus for the medical domain in four languages, namely English, French, Italian and Spanish. This new corpus has been used to train Medical mT5, the first open-source text-to-text multilingual model for the medical domain. Additionally, we present two new evaluation benchmarks for all four languages with the aim of facilitating multilingual research in this domain. A comprehensive evaluation shows that Medical mT5 outperforms both encoders and similarly sized text-to-text models for the Spanish, French, and Italian benchmarks, while being competitive with current state-of-the-art LLMs in English.

Mixture-of-Experts (MoE) architectures have become the key to scaling modern LLMs, yet little is understood about how their sparse routing dynamics respond to multilingual data. In this work, we analyze expert routing patterns using parallel multilingual datasets and present highly interpretable layer-wise phenomena. We find that MoE models route tokens in language-specific ways in the early and late decoder layers but exhibit significant cross-lingual routing alignment in middle layers, mirroring parameter-sharing trends observed in dense LLMs. In particular, we reveal a clear, strong correlation between a model's performance in a given language and how similarly its tokens are routed to English in these layers. Extending beyond correlation, we explore inference-time interventions that induce higher cross-lingual routing alignment. We introduce a method that steers the router by promoting middle-layer task experts frequently activated in English, and it successfully increases multilingual performance. These 1-2% gains are remarkably consistent across two evaluation tasks, three models, and 15+ languages, especially given that these simple interventions override routers of extensively trained, state-of-the-art LLMs. In comparison, interventions outside of the middle layers or targeting multilingual-specialized experts only yield performance degradation. Altogether, we present numerous findings that explain how MoEs process non-English text and demonstrate that generalization is limited by the model's ability to leverage language-universal experts in all languages.

Large language models (LLMs) are a new and powerful tool for a wide span of applications involving natural language and demonstrate impressive code generation abilities. In this paper, we explore the capabilitity of state-of-the-art LLMs, including closed-source options like OpenAI GPT-4 and open-source alternatives like Meta AI Codellama, to automatically generate tests and use these tests to validate and verify compiler implementations of a directive-based programming paradigm, OpenACC. Our approach entails exploring various prompt engineering techniques including a code template, retrieval-augmented generation (RAG) with code template, expressive prompt using RAG with code template, one-shot example, and RAG with one-shot example. This paper focusses on (a) exploring the capabilities of the latest LLMs for code generation, (b) investigating prompt and fine tuning methods, and (c) analyzing the outcome of LLMs generated tests

While Large Language Models (LLMs) demonstrate impressive performance in mathematics, existing math benchmarks come with significant limitations. Many focus on problems with fixed ground-truth answers, and are often saturated due to problem simplicity or the viability of guessing or memorization. Crucially, they capture only a narrow subset of relevant math problems. To address this research gap, we introduce \mc, a new benchmark of 126 challenging problems sourced from various math competitions, which targets constructive proofs, a widely encountered problem type requiring the construction of mathematical objects with specific properties. These proofs are particularly suitable for LLM evaluation, as solution correctness can be easily verified. Our automated verifiers also enable MathConstruct to generate problem variations, used to evaluate robustness. State-of-the-art LLMs solve only 54% of MathConstruct problems, highlighting its complexity and importance for LLM evaluation.

Large language models (LLMs) have recently achieved human-level performance on a range of professional and academic benchmarks. The accessibility of these models has lagged behind their performance. State-of-the-art LLMs require costly infrastructure; are only accessible via rate-limited, geo-locked, and censored web interfaces; and lack publicly available code and technical reports. In this paper, we tell the story of GPT4All, a popular open source repository that aims to democratize access to LLMs. We outline the technical details of the original GPT4All model family, as well as the evolution of the GPT4All project from a single model into a fully fledged open source ecosystem. It is our hope that this paper acts as both a technical overview of the original GPT4All models as well as a case study on the subsequent growth of the GPT4All open source ecosystem.

The design of complex machines stands as both a marker of human intelligence and a foundation of engineering practice. Given recent advances in large language models (LLMs), we ask whether they, too, can learn to create. We approach this question through the lens of compositional machine design: a task in which machines are assembled from standardized components to meet functional demands like locomotion or manipulation in a simulated physical environment. To support this investigation, we introduce BesiegeField, a testbed built on the machine-building game Besiege, which enables part-based construction, physical simulation and reward-driven evaluation. Using BesiegeField, we benchmark state-of-the-art LLMs with agentic workflows and identify key capabilities required for success, including spatial reasoning, strategic assembly, and instruction-following. As current open-source models fall short, we explore reinforcement learning (RL) as a path to improvement: we curate a cold-start dataset, conduct RL finetuning experiments, and highlight open challenges at the intersection of language, machine design, and physical reasoning.

Named entity recognition (NER) is a core task for historical research in automatically establishing all references to people, places, events and the like. Yet, do to the high linguistic and genre diversity of sources, only limited canonisation of spellings, the level of required historical domain knowledge, and the scarcity of annotated training data, established approaches to natural language processing (NLP) have been both extremely expensive and yielded only unsatisfactory results in terms of recall and precision. Our paper introduces a new approach. We demonstrate how readily-available, state-of-the-art LLMs significantly outperform two leading NLP frameworks, spaCy and flair, for NER in historical documents by seven to twentytwo percent higher F1-Scores. Our ablation study shows how providing historical context to the task and a bit of persona modelling that turns focus away from a purely linguistic approach are core to a successful prompting strategy. We also demonstrate that, contrary to our expectations, providing increasing numbers of examples in few-shot approaches does not improve recall or precision below a threshold of 16-shot. In consequence, our approach democratises access to NER for all historians by removing the barrier of scripting languages and computational skills required for established NLP tools and instead leveraging natural language prompts and consumer-grade tools and frontends.

Large Language Models (LLMs) have been profusely evaluated on their ability to answer questions on many topics and their performance on different natural language understanding tasks. Those tests are usually conducted in English, but most LLM users are not native English speakers. Therefore, it is of interest to analyze how LLMs understand other languages at different levels: from paragraphs to morphems. In this paper, we evaluate the performance of state-of-the-art LLMs in TELEIA, a recently released benchmark with similar questions to those of Spanish exams for foreign students, covering topics such as reading comprehension, word formation, meaning and compositional semantics, and grammar. The results show that LLMs perform well at understanding Spanish but are still far from achieving the level of a native speaker in terms of grammatical competence.

As large language models (LLMs) have been deployed in various real-world settings, concerns about the harm they may propagate have grown. Various jailbreaking techniques have been developed to expose the vulnerabilities of these models and improve their safety. This work reveals that many state-of-the-art LLMs are vulnerable to malicious requests hidden behind scientific language. Specifically, our experiments with GPT4o, GPT4o-mini, GPT-4, LLama3-405B-Instruct, Llama3-70B-Instruct, Cohere, Gemini models demonstrate that, the models' biases and toxicity substantially increase when prompted with requests that deliberately misinterpret social science and psychological studies as evidence supporting the benefits of stereotypical biases. Alarmingly, these models can also be manipulated to generate fabricated scientific arguments claiming that biases are beneficial, which can be used by ill-intended actors to systematically jailbreak these strong LLMs. Our analysis studies various factors that contribute to the models' vulnerabilities to malicious requests in academic language. Mentioning author names and venues enhances the persuasiveness of models, and the bias scores increase as dialogues progress. Our findings call for a more careful investigation on the use of scientific data for training LLMs.

Since state-of-the-art LLMs often underperform in languages other than English or Chinese, improving the capability of LLMs in new languages has become an essential task. Moreover, LLMs' entire end-to-end training process remains largely unknown to the public due to proprietary reasons, technical complexity, inconsistent documentation, and ethical considerations. The complete picture remains a closely guarded secret within the industry. This paper presents methods to adapt an existing English-based LLM to Korean in a low-budget scenario. We describe the entire end-to-end process: collecting Korean datasets, preprocessing the data, training the model, creating downstream benchmarks, and conducting evaluations. The evaluation results indicate that our method can effectively and cost-efficiently add new language capabilities to existing LLMs. Our new bilingual models, Thunder-LLM and Thunder-LLM-Ins, achieve superior Korean performance compared to state-of-the-art models while utilizing minimal data and computational resources. We share our comprehensive experience and make the code publicly available.

Autonomous agents that address day-to-day digital tasks (e.g., ordering groceries for a household), must not only operate multiple apps (e.g., notes, messaging, shopping app) via APIs, but also generate rich code with complex control flow in an iterative manner based on their interaction with the environment. However, existing benchmarks for tool use are inadequate, as they only cover tasks that require a simple sequence of API calls. To remedy this gap, we built AppWorld Engine, a high-quality execution environment (60K lines of code) of 9 day-to-day apps operable via 457 APIs and populated with realistic digital activities simulating the lives of ~100 fictitious users. We then created AppWorld Benchmark (40K lines of code), a suite of 750 natural, diverse, and challenging autonomous agent tasks requiring rich and interactive code generation. It supports robust programmatic evaluation with state-based unit tests, allowing for different ways of completing a task while also checking for unexpected changes, i.e., collateral damage. The state-of-the-art LLM, GPT-4o, solves only ~49% of our 'normal' tasks and ~30% of 'challenge' tasks, while other models solve at least 16% fewer. This highlights the benchmark's difficulty and AppWorld's potential to push the frontiers of interactive coding agents. The project website is available at https://appworld.dev/.

The number of pretrained Large Language Models (LLMs) is increasing steadily, though the majority are designed predominantly for the English language. While state-of-the-art LLMs can handle other languages, due to language contamination or some degree of multilingual pretraining data, they are not optimized for non-English languages, leading to inefficient encoding (high token "fertility") and slower inference speed. In this work, we thoroughly compare a variety of vocabulary adaptation techniques for optimizing English LLMs for the Italian language, and put forward Semantic Alignment Vocabulary Adaptation (SAVA), a novel method that leverages neural mapping for vocabulary substitution. SAVA achieves competitive performance across multiple downstream tasks, enhancing grounded alignment strategies. We adapt two LLMs: Mistral-7b-v0.1, reducing token fertility by 25\%, and Llama-3.1-8B, optimizing the vocabulary and reducing the number of parameters by 1 billion. We show that, following the adaptation of the vocabulary, these models can recover their performance with a relatively limited stage of continual training on the target language. Finally, we test the capabilities of the adapted models on various multi-choice and generative tasks.

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.

Recent advances in large language models (LLMs) have fueled the vision of automated scientific discovery, often called AI Co-Scientists. To date, prior work casts these systems as generative co-authors responsible for crafting hypotheses, synthesizing code, or drafting manuscripts. In this work, we explore a complementary application: using LLMs as verifiers to automate the academic verification of scientific manuscripts. To that end, we introduce SPOT, a dataset of 83 published papers paired with 91 errors significant enough to prompt errata or retraction, cross-validated with actual authors and human annotators. Evaluating state-of-the-art LLMs on SPOT, we find that none surpasses 21.1\% recall or 6.1\% precision (o3 achieves the best scores, with all others near zero). Furthermore, confidence estimates are uniformly low, and across eight independent runs, models rarely rediscover the same errors, undermining their reliability. Finally, qualitative analysis with domain experts reveals that even the strongest models make mistakes resembling student-level misconceptions derived from misunderstandings. These findings highlight the substantial gap between current LLM capabilities and the requirements for dependable AI-assisted academic verification.

Humans are influenced by how information is presented, a phenomenon known as the framing effect. Previous work has shown that LLMs may also be susceptible to framing but has done so on synthetic data and did not compare to human behavior. We introduce WildFrame, a dataset for evaluating LLM responses to positive and negative framing, in naturally-occurring sentences, and compare humans on the same data. WildFrame consists of 1,000 texts, first selecting real-world statements with clear sentiment, then reframing them in either positive or negative light, and lastly, collecting human sentiment annotations. By evaluating eight state-of-the-art LLMs on WildFrame, we find that all models exhibit framing effects similar to humans (rgeq0.57), with both humans and models being more influenced by positive rather than negative reframing. Our findings benefit model developers, who can either harness framing or mitigate its effects, depending on the downstream application.

Recent studies have demonstrated the potential of Large Language Models (LLMs) in generating GPU Kernels. Current benchmarks focus on the translation of high-level languages into CUDA, overlooking the more general and challenging task of text-to-CUDA generation. Furthermore, given the hardware-specific and performance-critical features of GPU programming, accurately assessing the performance of LLM-generated GPU programs is nontrivial. In this work, we introduce CUDABench, a comprehensive benchmark designed to evaluate the text-to-CUDA capabilities of LLMs. First, we construct CUDABench-Set, which covers Breadth-Depth-Difficulty evaluation space in diverse application domains, including artificial intelligence, scientific computing, and data analytics, etc. Furthermore, we propose CUDABench-Score and Generative Verification Pipeline that assess (1) compilation correctness, (2) functional consistency through execution-based verification, and (3) a novel roofline-based metric, Performance-Score. Benchmarking state-of-the-art LLMs reveals insightful findings and challenges of text-to-CUDA, such as a notable mismatch between high compilation success rates and low functional correctness, a lack of domain-specific algorithmic knowledge, and suboptimal utilization of GPU hardware resources. Our benchmark is available at https://github.com/CUDA-Bench/CUDABench.

Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, 5.7 times 10^2 or 580?" This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers). Furthermore, the hidden state after reading a pair of numerals encodes their ranking, with a linear classifier achieving over 90% accuracy. Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective. Finally, we show that incorporating the classifier probe's log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models' internal magnitude representations can enhance their numerical reasoning capabilities.

Large Language Models have significantly advanced natural language processing, achieving remarkable performance in tasks such as language generation, translation, and reasoning. However, their substantial computational requirements restrict deployment to high-end systems, limiting accessibility on consumer-grade devices. This challenge is especially pronounced for under-resourced languages like those spoken in the Iberian Peninsula, where relatively limited linguistic resources and benchmarks hinder effective evaluation. This work presents a comprehensive evaluation of compact state-of-the-art LLMs across several essential NLP tasks tailored for Iberian languages. The results reveal that while some models consistently excel in certain tasks, significant performance gaps remain, particularly for languages such as Basque. These findings highlight the need for further research on balancing model compactness with robust multilingual performance

As information retrieval systems continue to evolve, accurate evaluation and benchmarking of these systems become pivotal. Web search datasets, such as MS MARCO, primarily provide short keyword queries without accompanying intent or descriptions, posing a challenge in comprehending the underlying information need. This paper proposes an approach to augmenting such datasets to annotate informative query descriptions, with a focus on two prominent benchmark datasets: TREC-DL-21 and TREC-DL-22. Our methodology involves utilizing state-of-the-art LLMs to analyze and comprehend the implicit intent within individual queries from benchmark datasets. By extracting key semantic elements, we construct detailed and contextually rich descriptions for these queries. To validate the generated query descriptions, we employ crowdsourcing as a reliable means of obtaining diverse human perspectives on the accuracy and informativeness of the descriptions. This information can be used as an evaluation set for tasks such as ranking, query rewriting, or others.

The pace of evolution of Large Language Models (LLMs) necessitates new approaches for rigorous and comprehensive evaluation. Traditional human annotation is increasingly impracticable due to the complexities and costs involved in generating high-quality, challenging problems. In this work, we introduce CHASE, a unified framework to synthetically generate challenging problems using LLMs without human involvement. For a given task, our approach builds a hard problem in a bottom-up manner from simpler components. Moreover, our framework decomposes the generation process into independently verifiable sub-tasks, thereby ensuring a high level of quality and correctness. We implement CHASE to create evaluation benchmarks across three diverse domains: (1) document-based question answering, (2) repository-level code completion, and (3) math reasoning. The performance of state-of-the-art LLMs on these synthetic benchmarks lies in the range of 40-60% accuracy, thereby demonstrating the effectiveness of our framework at generating challenging problems. We publicly release our benchmarks and code.

Large language models (LLMs) have shown great potential for healthcare applications. However, existing evaluation benchmarks provide minimal coverage of Alzheimer's Disease and Related Dementias (ADRD). To address this gap, we introduce ADRD-Bench, the first ADRD-specific benchmark dataset designed for rigorous evaluation of LLMs. ADRD-Bench has two components: 1) ADRD Unified QA, a synthesis of 1,352 questions consolidated from seven established medical benchmarks, providing a unified assessment of clinical knowledge; and 2) ADRD Caregiving QA, a novel set of 149 questions derived from the Aging Brain Care (ABC) program, a widely used, evidence-based brain health management program. Guided by a program with national expertise in comprehensive ADRD care, this new set was designed to mitigate the lack of practical caregiving context in existing benchmarks. We evaluated 33 state-of-the-art LLMs on the proposed ADRD-Bench. Results showed that the accuracy of open-weight general models ranged from 0.63 to 0.93 (mean: 0.78; std: 0.09). The accuracy of open-weight medical models ranged from 0.48 to 0.93 (mean: 0.82; std: 0.13). The accuracy of closed-source general models ranged from 0.83 to 0.91 (mean: 0.89; std: 0.03). While top-tier models achieved high accuracies (>0.9), case studies revealed that inconsistent reasoning quality and stability limit their reliability, highlighting a critical need for domain-specific improvement to enhance LLMs' knowledge and reasoning grounded in daily caregiving data. The entire dataset is available at https://github.com/IIRL-ND/ADRD-Bench.

Large Language Models (LLMs) have shown growing potential in molecular sciences, but they often produce chemically inaccurate descriptions and struggle to recognize or justify potential errors. This raises important concerns about their robustness and reliability in scientific applications. To support more rigorous evaluation of LLMs in chemical reasoning, we present the MolErr2Fix benchmark, designed to assess LLMs on error detection and correction in molecular descriptions. Unlike existing benchmarks focused on molecule-to-text generation or property prediction, MolErr2Fix emphasizes fine-grained chemical understanding. It tasks LLMs with identifying, localizing, explaining, and revising potential structural and semantic errors in molecular descriptions. Specifically, MolErr2Fix consists of 1,193 fine-grained annotated error instances. Each instance contains quadruple annotations, i.e,. (error type, span location, the explanation, and the correction). These tasks are intended to reflect the types of reasoning and verification required in real-world chemical communication. Evaluations of current state-of-the-art LLMs reveal notable performance gaps, underscoring the need for more robust chemical reasoning capabilities. MolErr2Fix provides a focused benchmark for evaluating such capabilities and aims to support progress toward more reliable and chemically informed language models. All annotations and an accompanying evaluation API will be publicly released to facilitate future research.

Large Language Models (LLMs) have made significant strides in natural language generation but often face challenges in tasks requiring precise calculations and structural analysis. This paper investigates the performance of state-of-the-art LLMs on language complexity measurement tasks, through the computation of the LIX readability metric and Average Dependency Distance (ADD). Using Swedish high school and university-level essays, we evaluate the models' abilities to compute LIX scores and perform dependency parsing, comparing their results to established ground truths. Our findings reveal that while all models demonstrate some capacity for these tasks, ChatGPT-o1-mini performs most consistently, achieving the highest accuracy in both LIX computation and dependency parsing. Additionally, we observe a strong significant correlation -0.875 p 0.026 (N=6) between the models' accuracy in computing LIX and their overall performance on the Massive Multitask Language Understanding (MMLU) benchmark. These results suggest that language complexity measurement abilities can serve as a noisy zero-shot proxies for assessing the general capabilities of LLMs, providing a practical method for model evaluation without the need for extensive benchmarking datasets.

LLMs are transforming software development, yet current code generation and code repair benchmarks mainly assess syntactic and functional correctness in simple, single-error cases. LLMs' capabilities to autonomously find and fix runtime logical errors in complex data science code remain largely unexplored. To address this gap, we introduce DSDBench: the Data Science Debugging Benchmark, the first benchmark for systematic evaluation of LLMs on multi-hop error tracing and multi-bug detection in data science code debugging. DSDBench adapts datasets from existing data science task benchmarks, such as DABench and MatPlotBench, featuring realistic data science debugging tasks with automatically synthesized multi-hop, multi-bug code snippets. DSDBench includes 1,117 annotated samples with 741 cause-effect error pairs and runtime error messages. Evaluations of state-of-the-art LLMs on DSDBench show significant performance gaps, highlighting challenges in debugging logical runtime errors in data science code. DSDBench offers a crucial resource to evaluate and improve LLMs' debugging and reasoning capabilities, enabling more reliable AI-assisted data science in the future. DSDBench is publicly available at github.com/KevinCL16/DSDBench.

Large Language Models (LLMs) have demonstrated impressive capabilities in various tasks, including instruction following, which is crucial for aligning model outputs with user expectations. However, evaluating LLMs' ability to follow instructions remains challenging due to the complexity and subjectivity of human language. Current benchmarks primarily focus on single-turn, monolingual instructions, which do not adequately reflect the complexities of real-world applications that require handling multi-turn and multilingual interactions. To address this gap, we introduce Multi-IF, a new benchmark designed to assess LLMs' proficiency in following multi-turn and multilingual instructions. Multi-IF, which utilizes a hybrid framework combining LLM and human annotators, expands upon the IFEval by incorporating multi-turn sequences and translating the English prompts into another 7 languages, resulting in a dataset of 4,501 multilingual conversations, where each has three turns. Our evaluation of 14 state-of-the-art LLMs on Multi-IF reveals that it presents a significantly more challenging task than existing benchmarks. All the models tested showed a higher rate of failure in executing instructions correctly with each additional turn. For example, o1-preview drops from 0.877 at the first turn to 0.707 at the third turn in terms of average accuracy over all languages. Moreover, languages with non-Latin scripts (Hindi, Russian, and Chinese) generally exhibit higher error rates, suggesting potential limitations in the models' multilingual capabilities. We release Multi-IF prompts and the evaluation code base to encourage further research in this critical area.

Existing benchmarks for large language models (LLMs) increasingly struggle to differentiate between top-performing models, underscoring the need for more challenging evaluation frameworks. We introduce MMLU-Pro+, an enhanced benchmark building upon MMLU-Pro to assess shortcut learning and higher-order reasoning in LLMs. By incorporating questions with multiple correct answers across diverse domains, MMLU-Pro+ tests LLMs' ability to engage in complex reasoning and resist simplistic problem-solving strategies. Our results show that MMLU-Pro+ maintains MMLU-Pro's difficulty while providing a more rigorous test of model discrimination, particularly in multi-correct answer scenarios. We introduce novel metrics like shortcut selection ratio and correct pair identification ratio, offering deeper insights into model behavior and anchoring bias. Evaluations of six state-of-the-art LLMs reveal significant performance gaps, highlighting variations in reasoning abilities and bias susceptibility. We release the dataset and evaluation codes at https://github.com/asgsaeid/mmlu-pro-plus.

Formal mathematical reasoning remains a critical challenge for artificial intelligence, hindered by limitations of existing benchmarks in scope and scale. To address this, we present FormalMATH, a large-scale Lean4 benchmark comprising 5,560 formally verified problems spanning from high-school Olympiad challenges to undergraduate-level theorems across diverse domains (e.g., algebra, applied mathematics, calculus, number theory, and discrete mathematics). To mitigate the inefficiency of manual formalization, we introduce a novel human-in-the-loop autoformalization pipeline that integrates: (1) specialized large language models (LLMs) for statement autoformalization, (2) multi-LLM semantic verification, and (3) negation-based disproof filtering strategies using off-the-shelf LLM-based provers. This approach reduces expert annotation costs by retaining 72.09% of statements before manual verification while ensuring fidelity to the original natural-language problems. Our evaluation of state-of-the-art LLM-based theorem provers reveals significant limitations: even the strongest models achieve only 16.46% success rate under practical sampling budgets, exhibiting pronounced domain bias (e.g., excelling in algebra but failing in calculus) and over-reliance on simplified automation tactics. Notably, we identify a counterintuitive inverse relationship between natural-language solution guidance and proof success in chain-of-thought reasoning scenarios, suggesting that human-written informal reasoning introduces noise rather than clarity in the formal reasoning settings. We believe that FormalMATH provides a robust benchmark for benchmarking formal mathematical reasoning.

The recent exponential growth of Large Language Models (LLMs) has relied on GPU-based systems. However, CPUs are emerging as a flexible and lower-cost alternative, especially when targeting inference and reasoning workloads. RISC-V is rapidly gaining traction in this area, given its open and vendor-neutral ISA. However, the RISC-V hardware for LLM workloads and the corresponding software ecosystem are not fully mature and streamlined, given the requirement of domain-specific tuning. This paper aims at filling this gap, focusing on optimizing LLM inference on the Sophon SG2042, the first commercially available many-core RISC-V CPU with vector processing capabilities. On two recent state-of-the-art LLMs optimized for reasoning, DeepSeek R1 Distill Llama 8B and DeepSeek R1 Distill QWEN 14B, we achieve 4.32/2.29 token/s for token generation and 6.54/3.68 token/s for prompt processing, with a speed up of up 2.9x/3.0x compared to our baseline.

Large language models (LLMs) achieve promising performance, yet their ability to reason remains poorly understood. Existing evaluations largely emphasize task-level accuracy, often conflating pattern matching with reasoning capability. We present X-RAY, an explainable reasoning analysis system that maps the LLM reasoning capability using calibrated, formally verified probes. We model reasoning capability as a function of extractable structure, operationalized through formal properties such as constraint interaction, reasoning depth, and solution-space geometry. X-Ray generates probes via formal tools with controlled structural variations, enabling precise isolation of incremental structural information through formal calibration and verification. We evaluate state-of-the-art LLMs on problems ranging from junior-level to advanced in mathematics, physics, and chemistry. Our analysis reveals a systematic asymmetry in LLM reasoning: models are relatively robust to constraint refinement, where additional conditions shrink an existing solution space, but degrade sharply under solution-space restructuring, where modifications alter the underlying structural form of the solution manifold. Moreover, calibrated formal probes differentiate models that appear indistinguishable on standard benchmarks and reveal failure modes that are structurally interpretable rather than opaque. Beyond evaluation, our framework is contamination-free and supports the training and testing of reasoning models.

Traditional customer support systems, such as Interactive Voice Response (IVR), rely on rigid scripts and lack the flexibility required for handling complex, policy-driven tasks. While large language model (LLM) agents offer a promising alternative, evaluating their ability to act in accordance with business rules and real-world support workflows remains an open challenge. Existing benchmarks primarily focus on tool usage or task completion, overlooking an agent's capacity to adhere to multi-step policies, navigate task dependencies, and remain robust to unpredictable user or environment behavior. In this work, we introduce JourneyBench, a benchmark designed to assess policy-aware agents in customer support. JourneyBench leverages graph representations to generate diverse, realistic support scenarios and proposes the User Journey Coverage Score, a novel metric to measure policy adherence. We evaluate multiple state-of-the-art LLMs using two agent designs: a Static-Prompt Agent (SPA) and a Dynamic-Prompt Agent (DPA) that explicitly models policy control. Across 703 conversations in three domains, we show that DPA significantly boosts policy adherence, even allowing smaller models like GPT-4o-mini to outperform more capable ones like GPT-4o. Our findings demonstrate the importance of structured orchestration and establish JourneyBench as a critical resource to advance AI-driven customer support beyond IVR-era limitations.

High-level synthesis (HLS) enables software developers to describe and implement hardware at a higher level of abstraction by using C/C++ instead of traditional hardware description languages to automatically generate FPGA-ready designs. However, generating HLS code significantly differs from standard C/C++: it disallows certain coding idioms, relies on specialized libraries, and critically requires fine-grained transformations and the insertion of optimization directives (pragmas) to achieve high performance. Large language models (LLMs) have shown promise in automating such transformations, yet existing open-source datasets lack sufficient complexity and optimization diversity. To address this gap, we introduce the HLStrans dataset, a comprehensive collection of 137 distinct real word programs, each annotated with a variety of C-to-HLS transformations that yield over 23K labeled design variants. These include a broad spectrum of pragmas and code-level optimizations. We benchmark state-of-the-art LLMs on this dataset to evaluate their ability to generate synthesizable, high-performance HLS code. As part of an ongoing effort, we plan to expand the HLStrans dataset in both scale and program variety, further empowering research at the intersection of AI and hardware synthesis.

With powerful large language models (LLMs) demonstrating superhuman reasoning capabilities, a critical question arises: Do LLMs genuinely reason, or do they merely recall answers from their extensive, web-scraped training datasets? Publicly released benchmarks inevitably become contaminated once incorporated into subsequent LLM training sets, undermining their reliability as faithful assessments. To address this, we introduce KUMO, a generative evaluation framework designed specifically for assessing reasoning in LLMs. KUMO synergistically combines LLMs with symbolic engines to dynamically produce diverse, multi-turn reasoning tasks that are partially observable and adjustable in difficulty. Through an automated pipeline, KUMO continuously generates novel tasks across open-ended domains, compelling models to demonstrate genuine generalization rather than memorization. We evaluated 23 state-of-the-art LLMs on 5,000 tasks across 100 domains created by KUMO, benchmarking their reasoning abilities against university students. Our findings reveal that many LLMs have outperformed university-level performance on easy reasoning tasks, and reasoning-scaled LLMs reach university-level performance on complex reasoning challenges. Moreover, LLM performance on KUMO tasks correlates strongly with results on newly released real-world reasoning benchmarks, underscoring KUMO's value as a robust, enduring assessment tool for genuine LLM reasoning capabilities.

While Large Language Models (LLMs) have shown promise in software engineering, their application to unit testing remains largely confined to isolated test generation or oracle prediction, neglecting the broader challenge of test suite maintenance. We introduce TAM-Eval (Test Automated Maintenance Evaluation), a framework and benchmark designed to evaluate model performance across three core test maintenance scenarios: creation, repair, and updating of test suites. Unlike prior work limited to function-level tasks, TAM-Eval operates at the test file level, while maintaining access to full repository context during isolated evaluation, better reflecting real-world maintenance workflows. Our benchmark comprises 1,539 automatically extracted and validated scenarios from Python, Java, and Go projects. TAM-Eval supports system-agnostic evaluation of both raw LLMs and agentic workflows, using a reference-free protocol based on test suite pass rate, code coverage, and mutation testing. Empirical results indicate that state-of-the-art LLMs have limited capabilities in realistic test maintenance processes and yield only marginal improvements in test effectiveness. We release TAM-Eval as an open-source framework to support future research in automated software testing. Our data and code are publicly available at https://github.com/trndcenter/TAM-Eval.

Arabic poetry is one of the richest and most culturally rooted forms of expression in the Arabic language, known for its layered meanings, stylistic diversity, and deep historical continuity. Although large language models (LLMs) have demonstrated strong performance across languages and tasks, their ability to understand Arabic poetry remains largely unexplored. In this work, we introduce Fann or Flop, the first benchmark designed to assess the comprehension of Arabic poetry by LLMs in 12 historical eras, covering 14 core poetic genres and a variety of metrical forms, from classical structures to contemporary free verse. The benchmark comprises a curated corpus of poems with explanations that assess semantic understanding, metaphor interpretation, prosodic awareness, and cultural context. We argue that poetic comprehension offers a strong indicator for testing how good the LLM understands classical Arabic through Arabic poetry. Unlike surface-level tasks, this domain demands deeper interpretive reasoning and cultural sensitivity. Our evaluation of state-of-the-art LLMs shows that most models struggle with poetic understanding despite strong results on standard Arabic benchmarks. We release "Fann or Flop" along with the evaluation suite as an open-source resource to enable rigorous evaluation and advancement for Arabic language models. Code is available at: https://github.com/mbzuai-oryx/FannOrFlop.

Large Language Models (LLMs) have revolutionized natural language processing and hold immense potential for advancing Artificial Intelligence. However, the core architecture of most mainstream LLMs -- the Transformer -- has inherent limitations in computational depth, rendering them theoretically incapable of solving many reasoning tasks that demand increasingly deep computations. Chain of Thought (CoT) prompting has emerged as a technique to address these architectural limitations, as evidenced by several theoretical studies. It offers a promising approach to solving complex reasoning tasks that were previously beyond the capabilities of these models. Despite its successes, CoT and its variants (such as Tree of Thought, Graph of Thought, etc.) rely on a "one-prompt-for-all" approach, using a single prompt structure (e.g., "think step by step") for a wide range of tasks -- from counting and sorting to solving mathematical and algorithmic problems. This approach poses significant challenges for models to generate the correct reasoning steps, as the model must navigate through a vast prompt template space to find the appropriate template for each task. In this work, we build upon previous theoretical analyses of CoT to demonstrate how the one-prompt-for-all approach can negatively affect the computability of LLMs. We partition the solution search space into two: the prompt space and the answer space. Our findings show that task-specific supervision is essential for navigating the prompt space accurately and achieving optimal performance. Through experiments with state-of-the-art LLMs, we reveal a gap in reasoning performance when supervision is applied versus when it is not.

Social deduction games like Werewolf combine language, reasoning, and strategy, providing a testbed for studying natural language and social intelligence. However, most studies reduce the game to LLM-based self-play, yielding templated utterances and anecdotal cases that overlook the richness of social gameplay. Evaluation further relies on coarse metrics such as survival time or subjective scoring due to the lack of quality reference data. To address these gaps, we curate a high-quality, human-verified multimodal Werewolf dataset containing over 100 hours of video, 32.4M utterance tokens, and 15 rule variants. Based on this dataset, we propose a novel strategy-alignment evaluation that leverages the winning faction's strategies as ground truth in two stages: 1) Speech evaluation, formulated as multiple-choice-style tasks that assess whether the model can adopt appropriate stances across five dimensions of social ability; and 2) Decision evaluation, which assesses the model's voting choices and opponent-role inferences. This framework enables a fine-grained evaluation of models' linguistic and reasoning capabilities, while capturing their ability to generate strategically coherent gameplay. Our experiments show that state-of-the-art LLMs show diverse performance, with roughly half remain below 0.50, revealing clear gaps in deception and counterfactual reasoning. We hope our dataset further inspires research on language, reasoning, and strategy in multi-agent interaction.

As the capabilities of chatbots and their underlying LLMs continue to dramatically improve, evaluating their performance has increasingly become a major blocker to their further development. A major challenge is the available benchmarking datasets, which are largely static, outdated, and lacking in multilingual coverage, limiting their ability to capture subtle linguistic and cultural variations. This paper introduces MEDAL, an automated multi-agent framework for generating, evaluating, and curating more representative and diverse open-domain dialogue evaluation benchmarks. Our approach leverages several state-of-the-art LLMs to generate user-chatbot multilingual dialogues, conditioned on varied seed contexts. A strong LLM (GPT-4.1) is then used for a multidimensional analysis of the performance of the chatbots, uncovering noticeable cross-lingual performance differences. Guided by this large-scale evaluation, we curate a new meta-evaluation multilingual benchmark and human-annotate samples with nuanced quality judgments. This benchmark is then used to assess the ability of several reasoning and non-reasoning LLMs to act as evaluators of open-domain dialogues. We find that current LLMs struggle to detect nuanced issues, particularly those involving empathy and reasoning.

Do large language models (LLMs) have theory of mind? A plethora of papers and benchmarks have been introduced to evaluate if current models have been able to develop this key ability of social intelligence. However, all rely on limited datasets with simple patterns that can potentially lead to problematic blind spots in evaluation and an overestimation of model capabilities. We introduce ExploreToM, the first framework to allow large-scale generation of diverse and challenging theory of mind data for robust training and evaluation. Our approach leverages an A* search over a custom domain-specific language to produce complex story structures and novel, diverse, yet plausible scenarios to stress test the limits of LLMs. Our evaluation reveals that state-of-the-art LLMs, such as Llama-3.1-70B and GPT-4o, show accuracies as low as 0% and 9% on ExploreToM-generated data, highlighting the need for more robust theory of mind evaluation. As our generations are a conceptual superset of prior work, fine-tuning on our data yields a 27-point accuracy improvement on the classic ToMi benchmark (Le et al., 2019). ExploreToM also enables uncovering underlying skills and factors missing for models to show theory of mind, such as unreliable state tracking or data imbalances, which may contribute to models' poor performance on benchmarks.

While humans naturally develop theory of mind (ToM), the capability to understand other people's mental states and beliefs, state-of-the-art large language models (LLMs) underperform on simple ToM benchmarks. We posit that we can extend our understanding of LLMs' ToM abilities by evaluating key human ToM precursors -- perception inference and perception-to-belief inference -- in LLMs. We introduce two datasets, Percept-ToMi and Percept-FANToM, to evaluate these precursory inferences for ToM in LLMs by annotating characters' perceptions on ToMi and FANToM, respectively. Our evaluation of eight state-of-the-art LLMs reveals that the models generally perform well in perception inference while exhibiting limited capability in perception-to-belief inference (e.g., lack of inhibitory control). Based on these results, we present PercepToM, a novel ToM method leveraging LLMs' strong perception inference capability while supplementing their limited perception-to-belief inference. Experimental results demonstrate that PercepToM significantly enhances LLM's performance, especially in false belief scenarios.

This paper investigates the rationality of large language models (LLMs) in strategic decision-making contexts, specifically within the framework of game theory. We evaluate several state-of-the-art LLMs across a spectrum of complete-information and incomplete-information games. Our findings reveal that LLMs frequently deviate from rational strategies, particularly as the complexity of the game increases with larger payoff matrices or deeper sequential trees. To address these limitations, we design multiple game-theoretic workflows that guide the reasoning and decision-making processes of LLMs. These workflows aim to enhance the models' ability to compute Nash Equilibria and make rational choices, even under conditions of uncertainty and incomplete information. Experimental results demonstrate that the adoption of these workflows significantly improves the rationality and robustness of LLMs in game-theoretic tasks. Specifically, with the workflow, LLMs exhibit marked improvements in identifying optimal strategies, achieving near-optimal allocations in negotiation scenarios, and reducing susceptibility to exploitation during negotiations. Furthermore, we explore the meta-strategic considerations of whether it is rational for agents to adopt such workflows, recognizing that the decision to use or forgo the workflow constitutes a game-theoretic issue in itself. Our research contributes to a deeper understanding of LLMs' decision-making capabilities in strategic contexts and provides insights into enhancing their rationality through structured workflows. The findings have implications for the development of more robust and strategically sound AI agents capable of navigating complex interactive environments. Code and data supporting this study are available at https://github.com/Wenyueh/game_theory.

The advanced function-calling capabilities of foundation models open up new possibilities for deploying agents to perform complex API tasks. However, managing large amounts of data and interacting with numerous APIs makes function calling hardware-intensive and costly, especially on edge devices. Current Large Language Models (LLMs) struggle with function calling at the edge because they cannot handle complex inputs or manage multiple tools effectively. This results in low task-completion accuracy, increased delays, and higher power consumption. In this work, we introduce Less-is-More, a novel fine-tuning-free function-calling scheme for dynamic tool selection. Our approach is based on the key insight that selectively reducing the number of tools available to LLMs significantly improves their function-calling performance, execution time, and power efficiency on edge devices. Experimental results with state-of-the-art LLMs on edge hardware show agentic success rate improvements, with execution time reduced by up to 70% and power consumption by up to 40%.

Existing large language models (LLMs) for register transfer level code generation face challenges like compilation failures and suboptimal power, performance, and area (PPA) efficiency. This is due to the lack of PPA awareness in conventional transformer decoding algorithms. In response, we present an automated transformer decoding algorithm that integrates Monte Carlo tree-search for lookahead, guiding the transformer to produce compilable, functionally correct, and PPA-optimized code. Empirical evaluation with a fine-tuned language model on RTL codesets shows that our proposed technique consistently generates functionally correct code compared to prompting-only methods and effectively addresses the PPA-unawareness drawback of naive large language models. For the largest design generated by the state-of-the-art LLM (16-bit adder), our technique can achieve a 31.8% improvement in the area-delay product.