Daily Papers

Effective evaluation is critical for driving advancements in MLLM research. The surgical action planning (SAP) task, which aims to generate future action sequences from visual inputs, demands precise and sophisticated analytical capabilities. Unlike mathematical reasoning, surgical decision-making operates in life-critical domains and requires meticulous, verifiable processes to ensure reliability and patient safety. This task demands the ability to distinguish between atomic visual actions and coordinate complex, long-horizon procedures, capabilities that are inadequately evaluated by current benchmarks. To address this gap, we introduce SAP-Bench, a large-scale, high-quality dataset designed to enable multimodal large language models (MLLMs) to perform interpretable surgical action planning. Our SAP-Bench benchmark, derived from the cholecystectomy procedures context with the mean duration of 1137.5s, and introduces temporally-grounded surgical action annotations, comprising the 1,226 clinically validated action clips (mean duration: 68.7s) capturing five fundamental surgical actions across 74 procedures. The dataset provides 1,152 strategically sampled current frames, each paired with the corresponding next action as multimodal analysis anchors. We propose the MLLM-SAP framework that leverages MLLMs to generate next action recommendations from the current surgical scene and natural language instructions, enhanced with injected surgical domain knowledge. To assess our dataset's effectiveness and the broader capabilities of current models, we evaluate seven state-of-the-art MLLMs (e.g., OpenAI-o1, GPT-4o, QwenVL2.5-72B, Claude-3.5-Sonnet, GeminiPro2.5, Step-1o, and GLM-4v) and reveal critical gaps in next action prediction performance.

Multimodal Large Language Models (MLLMs) have shown their remarkable abilities in visual perception and understanding recently. However, how to comprehensively evaluate the capabilities of MLLMs remains a challenge. Most of the existing benchmarks predominantly focus on assessing perception, cognition, and reasoning, neglecting the abilities of self-awareness, referring to the model's recognition of its own capability boundary. In our study, we focus on self-awareness in image perception and introduce the knowledge quadrant for MLLMs, which clearly defines the knowns and unknowns in perception. Based on this, we propose a novel benchmark specifically designed to evaluate the Self-Aware capabilities in Perception for MLLMs(MM-SAP). MM-SAP encompasses three distinct sub-datasets, each focusing on different aspects of self-awareness. We evaluated eight well-known MLLMs using MM-SAP, analyzing their self-awareness and providing detailed insights. Code and data are available at https://github.com/YHWmz/MM-SAP

Recent works on structured text translation remain limited to the sentence level, as they struggle to effectively handle the complex document-level XML or HTML structures. To address this, we propose Format Reinforcement Learning (FormatRL), which employs Group Relative Policy Optimization on top of a supervised fine-tuning model to directly optimize novel structure-aware rewards: 1) TreeSim, which measures structural similarity between predicted and reference XML trees and 2) Node-chrF, which measures translation quality at the level of XML nodes. Additionally, we apply StrucAUC, a fine-grained metric distinguishing between minor errors and major structural failures. Experiments on the SAP software-documentation benchmark demonstrate improvements across six metrics and an analysis further shows how different reward functions contribute to improvements in both structural and translation quality.

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

The integration of Large Language Models (LLMs) into various global cultures fundamentally presents a cultural challenge: LLMs must navigate interactions, respect social norms, and avoid transgressing cultural boundaries. However, it is still unclear if LLMs can adapt their outputs to diverse cultural norms. Our study focuses on this aspect. We introduce NormAd, a novel dataset, which includes 2.6k stories that represent social and cultural norms from 75 countries, to assess the ability of LLMs to adapt to different granular levels of socio-cultural contexts such as the country of origin, its associated cultural values, and prevalent social norms. Our study reveals that LLMs struggle with cultural reasoning across all contextual granularities, showing stronger adaptability to English-centric cultures over those from the Global South. Even with explicit social norms, the top-performing model, Mistral-7b-Instruct, achieves only 81.8\% accuracy, lagging behind the 95.6\% achieved by humans. Evaluation on NormAd further reveals that LLMs struggle to adapt to stories involving gift-giving across cultures. Due to inherent agreement or sycophancy biases, LLMs find it considerably easier to assess the social acceptability of stories that adhere to cultural norms than those that deviate from them. Our benchmark measures the cultural adaptability (or lack thereof) of LLMs, emphasizing the potential to make these technologies more equitable and useful for global audiences. We release the NormAd dataset and its associated code on GitHub.

Theory of mind (ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity. We introduce FANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering. Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs). In particular, we formulate multiple types of questions that demand the same underlying reasoning to identify illusory or false sense of ToM capabilities in LLMs. We show that FANToM is challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.

The last couple of years have witnessed emerging research that appropriates Theory-of-Mind (ToM) tasks designed for humans to benchmark LLM's ToM capabilities as an indication of LLM's social intelligence. However, this approach has a number of limitations. Drawing on existing psychology and AI literature, we summarize the theoretical, methodological, and evaluation limitations by pointing out that certain issues are inherently present in the original ToM tasks used to evaluate human's ToM, which continues to persist and exacerbated when appropriated to benchmark LLM's ToM. Taking a human-computer interaction (HCI) perspective, these limitations prompt us to rethink the definition and criteria of ToM in ToM benchmarks in a more dynamic, interactional approach that accounts for user preferences, needs, and experiences with LLMs in such evaluations. We conclude by outlining potential opportunities and challenges towards this direction.

The realism of images generated by multimodal large language models (MLLMs), such as GPT Image2 and Nano Banana2, has improved rapidly in recent years. Compared with early generative models, current models have made clear progress in text rendering. They can produce high-quality images that closely resemble real-world application scenarios. The enhanced generation capabilities of current MLLMs pose increasingly severe challenges to AI-generated image detection. Detection is no longer limited to identifying obvious artifacts left by early generators. Instead, it requires systematic and realistic benchmarks for the new generation of generated content. However, most existing benchmarks are still built around early generative models and cannot fully evaluate the forensic challenges introduced by high-quality and multi-form generated images. To address this gap, this paper constructs a benchmark dataset for detecting images generated by MLLMs. The benchmark covers several realistic application scenarios and adopts three generation protocols to simulate direct generation, reference-based reconstruction, and local editing. Based on this benchmark, we evaluate detector degradation from traditional scenarios to MLLM-generated images and analyze false positive rates and false negative rates across three sample types, revealing the failure modes of existing methods. We further propose a structural-artifact-prior-guided dual-stream prompt framework (SAP-DSP) as a strong baseline. SAP-DSP uses dual-stream prompt learning and structure-aware routing fusion to improve representation learning. Extensive experiments show that the proposed benchmark exposes the performance degradation of existing detectors on high-quality generated images, while SAP-DSP achieves more stable detection results on this benchmark. Our code and dataset are publicly available at https://github.com/xbrainnet/SAP-DSP.

Promptable instance segmentation is widely adopted in embodied and AR systems, yet the performance of foundation models trained on perspective imagery often degrades on 360° panoramas. In this paper, we introduce Segment Any 4K Panorama (SAP), a foundation model for 4K high-resolution panoramic instance-level segmentation. We reformulate panoramic segmentation as fixed-trajectory perspective video segmentation, decomposing a panorama into overlapping perspective patches sampled along a continuous spherical traversal. This memory-aligned reformulation preserves native 4K resolution while restoring the smooth viewpoint transitions required for stable cross-view propagation. To enable large-scale supervision, we synthesize 183,440 4K-resolution panoramic images with instance segmentation labels using the InfiniGen engine. Trained under this trajectory-aligned paradigm, SAP generalizes effectively to real-world 360° images, achieving +17.2 zero-shot mIoU gain over vanilla SAM2 of different sizes on real-world 4K panorama benchmark.

We introduce Social IQa, the first largescale benchmark for commonsense reasoning about social situations. Social IQa contains 38,000 multiple choice questions for probing emotional and social intelligence in a variety of everyday situations (e.g., Q: "Jordan wanted to tell Tracy a secret, so Jordan leaned towards Tracy. Why did Jordan do this?" A: "Make sure no one else could hear"). Through crowdsourcing, we collect commonsense questions along with correct and incorrect answers about social interactions, using a new framework that mitigates stylistic artifacts in incorrect answers by asking workers to provide the right answer to a different but related question. Empirical results show that our benchmark is challenging for existing question-answering models based on pretrained language models, compared to human performance (>20% gap). Notably, we further establish Social IQa as a resource for transfer learning of commonsense knowledge, achieving state-of-the-art performance on multiple commonsense reasoning tasks (Winograd Schemas, COPA).

Designing structured visuals such as presentation slides is essential for communicative needs, necessitating both content creation and visual planning skills. In this work, we tackle the challenge of automated slide generation, where models produce slide presentations from natural language (NL) instructions. We first introduce the SlidesBench benchmark, the first benchmark for slide generation with 7k training and 585 testing examples derived from 310 slide decks across 10 domains. SlidesBench supports evaluations that are (i)reference-based to measure similarity to a target slide, and (ii)reference-free to measure the design quality of generated slides alone. We benchmark end-to-end image generation and program generation methods with a variety of models, and find that programmatic methods produce higher-quality slides in user-interactable formats. Built on the success of program generation, we create AutoPresent, an 8B Llama-based model trained on 7k pairs of instructions paired with code for slide generation, and achieve results comparable to the closed-source model GPT-4o. We further explore iterative design refinement where the model is tasked to self-refine its own output, and we found that this process improves the slide's quality. We hope that our work will provide a basis for future work on generating structured visuals.

Humans continuously infer the states, goals, and behaviors of others by perceiving their surroundings in dynamic, real-world social interactions. However, most Theory of Mind (ToM) benchmarks only evaluate static, text-based scenarios, which have a significant gap compared to real interactions. We propose the SoMi-ToM benchmark, designed to evaluate multi-perspective ToM in embodied multi-agent complex social interactions. This benchmark is based on rich multimodal interaction data generated by the interaction environment SoMi, covering diverse crafting goals and social relationships. Our framework supports multi-level evaluation: (1) first-person evaluation provides multimodal (visual, dialogue, action, etc.) input from a first-person perspective during a task for real-time state inference, (2) third-person evaluation provides complete third-person perspective video and text records after a task for goal and behavior inference. This evaluation method allows for a more comprehensive examination of a model's ToM capabilities from both the subjective immediate experience and the objective global observation. We constructed a challenging dataset containing 35 third-person perspective videos, 363 first-person perspective images, and 1225 expert-annotated multiple-choice questions (three options). On this dataset, we systematically evaluated the performance of human subjects and several state-of-the-art large vision-language models (LVLMs). The results show that LVLMs perform significantly worse than humans on SoMi-ToM: the average accuracy gap between humans and models is 40.1% in first-person evaluation and 26.4% in third-person evaluation. This indicates that future LVLMs need to further improve their ToM capabilities in embodied, complex social interactions.

The interactive use of large language models (LLMs) in AI assistants (at work, home, etc.) introduces a new set of inference-time privacy risks: LLMs are fed different types of information from multiple sources in their inputs and are expected to reason about what to share in their outputs, for what purpose and with whom, within a given context. In this work, we draw attention to the highly critical yet overlooked notion of contextual privacy by proposing ConfAIde, a benchmark designed to identify critical weaknesses in the privacy reasoning capabilities of instruction-tuned LLMs. Our experiments show that even the most capable models such as GPT-4 and ChatGPT reveal private information in contexts that humans would not, 39% and 57% of the time, respectively. This leakage persists even when we employ privacy-inducing prompts or chain-of-thought reasoning. Our work underscores the immediate need to explore novel inference-time privacy-preserving approaches, based on reasoning and theory of mind.

Recent advances in AI agents capable of solving complex, everyday tasks, from scheduling to customer service, have enabled deployment in real-world settings, but their possibilities for unsafe behavior demands rigorous evaluation. While prior benchmarks have attempted to assess agent safety, most fall short by relying on simulated environments, narrow task domains, or unrealistic tool abstractions. We introduce OpenAgentSafety, a comprehensive and modular framework for evaluating agent behavior across eight critical risk categories. Unlike prior work, our framework evaluates agents that interact with real tools, including web browsers, code execution environments, file systems, bash shells, and messaging platforms; and supports over 350 multi-turn, multi-user tasks spanning both benign and adversarial user intents. OpenAgentSafety is designed for extensibility, allowing researchers to add tools, tasks, websites, and adversarial strategies with minimal effort. It combines rule-based analysis with LLM-as-judge assessments to detect both overt and subtle unsafe behaviors. Empirical analysis of five prominent LLMs in agentic scenarios reveals unsafe behavior in 51.2% of safety-vulnerable tasks with Claude-Sonnet-3.7, to 72.7% with o3-mini, highlighting critical safety vulnerabilities and the need for stronger safeguards before real-world deployment.

Recent advances in coding agents have made them capable of planning, editing, running, and testing complex code bases. Despite their growing ability in coding tasks, these systems still struggle to infer and track user intent, especially when instructions are underspecified or context-dependent. To bridge this gap, we introduce ToM-SWE, a dual-agent architecture that pairs a primary software-engineering (SWE) agent with a lightweight theory-of-mind (ToM) partner agent dedicated to modeling the user's mental state. The ToM agent infers user goals, constraints, and preferences from instructions and interaction history, maintains a persistent memory of the user, and provides user-related suggestions to the SWE agent. In two software engineering benchmarks (ambiguous SWE-bench and stateful SWE-bench), ToM-SWE improves task success rates and user satisfaction. Notably, on the stateful SWE benchmark, a newly introduced evaluation that provides agents with a user simulator along with previous interaction histories, ToM-SWE achieves a substantially higher task success rate of 59.7\% compared to 18.1\% for OpenHands, a state-of-the-art SWE agent. Furthermore, in a three-week study with professional developers using ToM-SWE in their daily work, participants found it useful 86\% of the time, underscoring the value of stateful user modeling for practical coding agents.

Abstract Language Models (LMs) achieve substantial language capabilities when finetuned using Reinforcement Learning with Human Feedback (RLHF). However, RLHF is an unstable and data-hungry process that continually requires new high-quality LM-generated data for finetuning. We introduce Advantage-Leftover Lunch RL (A-LoL), a new class of offline policy gradient algorithms that enable RL training on any pre-existing data. By assuming the entire LM output sequence as a single action, A-LoL allows incorporating sequence-level classifiers or human-designed scoring functions as rewards. Subsequently, by using LM's internal sequence-level value estimate, A-LoL filters negative advantage (low-quality) data points during training, making it resilient to noise. Overall, A-LoL is an easy-to-implement LM training recipe that is sample-efficient and stable. We demonstrate the effectiveness of A-LoL and its variants with a set of four different language generation tasks. We compare against both online RL (PPO) and recent preference-based (DPO, PRO) and reward-based (GOLD) offline RL baselines. On the commonly-used RLHF benchmark, Helpful and Harmless Assistant (HHA), LMs trained with A-LoL methods achieve the highest diversity while also being rated more safe and helpful than baselines according to humans. Additionally, in the remaining three tasks, A-LoL could optimize multiple distinct reward functions even when using noisy or suboptimal training data. We also release our experimental code. https://github.com/abaheti95/LoL-RL

Recent advances in large language models (LLMs) have led to their extensive global deployment, and ensuring their safety calls for comprehensive and multilingual toxicity evaluations. However, existing toxicity benchmarks are overwhelmingly focused on English, posing serious risks to deploying LLMs in other languages. We address this by introducing PolygloToxicityPrompts (PTP), the first large-scale multilingual toxicity evaluation benchmark of 425K naturally occurring prompts spanning 17 languages. We overcome the scarcity of naturally occurring toxicity in web-text and ensure coverage across languages with varying resources by automatically scraping over 100M web-text documents. Using PTP, we investigate research questions to study the impact of model size, prompt language, and instruction and preference-tuning methods on toxicity by benchmarking over 60 LLMs. Notably, we find that toxicity increases as language resources decrease or model size increases. Although instruction- and preference-tuning reduce toxicity, the choice of preference-tuning method does not have any significant impact. Our findings shed light on crucial shortcomings of LLM safeguarding and highlight areas for future research.

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Truly multilingual safety moderation efforts for Large Language Models (LLMs) have been hindered by a narrow focus on a small set of languages (e.g., English, Chinese) as well as a limited scope of safety definition, resulting in significant gaps in moderation capabilities. To bridge these gaps, we release POLYGUARD, a new state-of-the-art multilingual safety model for safeguarding LLM generations, and the corresponding training and evaluation datasets. POLYGUARD is trained on POLYGUARDMIX, the largest multilingual safety training corpus to date containing 1.91M samples across 17 languages (e.g., Chinese, Czech, English, Hindi). We also introduce POLYGUARDPROMPTS, a high quality multilingual benchmark with 29K samples for the evaluation of safety guardrails. Created by combining naturally occurring multilingual human-LLM interactions and human-verified machine translations of an English-only safety dataset (WildGuardMix; Han et al., 2024), our datasets contain prompt-output pairs with labels of prompt harmfulness, response harmfulness, and response refusal. Through extensive evaluations across multiple safety and toxicity benchmarks, we demonstrate that POLYGUARD outperforms existing state-of-the-art open-weight and commercial safety classifiers by 5.5%. Our contributions advance efforts toward safer multilingual LLMs for all global users.

As NLP evaluation shifts from static benchmarks to multi-turn interactive settings, LLM-based simulators have become widely used as user proxies, serving two roles: generating user turns and providing evaluation signals. Yet, these simulations are frequently assumed to be faithful to real human behaviors, often without rigorous verification. We formalize the Sim2Real gap in user simulation and present the first study running the full τ-bench protocol with real humans (451 participants, 165 tasks), benchmarking 31 LLM simulators across proprietary, open-source, and specialized families using the User-Sim Index (USI), a metric we introduce to quantify how well LLM simulators resemble real user interactive behaviors and feedback. Behaviorally, LLM simulators are excessively cooperative, stylistically uniform, and lack realistic frustration or ambiguity, creating an "easy mode" that inflates agent success rates above the human baseline. In evaluations, real humans provide nuanced judgments across eight quality dimensions while simulated users produce uniformly more positive feedback; rule-based rewards are failing to capture rich feedback signals generated by human users. Overall, higher general model capability does not necessarily yield more faithful user simulation. These findings highlight the importance of human validation when using LLM-based user simulators in the agent development cycle and motivate improved models for user simulation.

Improvements in language models are often driven by improving the quality of the data we train them on, which can be limiting when strong supervision is scarce. In this work, we show that paired preference data consisting of individually weak data points can enable gains beyond the strength of each individual data point. We formulate the delta learning hypothesis to explain this phenomenon, positing that the relative quality delta between points suffices to drive learning via preference tuning--even when supervised finetuning on the weak data hurts. We validate our hypothesis in controlled experiments and at scale, where we post-train 8B models on preference data generated by pairing a small 3B model's responses with outputs from an even smaller 1.5B model to create a meaningful delta. Strikingly, on a standard 11-benchmark evaluation suite (MATH, MMLU, etc.), our simple recipe matches the performance of Tulu 3, a state-of-the-art open model tuned from the same base model while relying on much stronger supervisors (e.g., GPT-4o). Thus, delta learning enables simpler and cheaper open recipes for state-of-the-art post-training. To better understand delta learning, we prove in logistic regression that the performance gap between two weak teacher models provides useful signal for improving a stronger student. Overall, our work shows that models can learn surprisingly well from paired data that might typically be considered weak.

The ideal LLM content moderation system would be both structurally interpretable (so its decisions can be explained to users) and steerable (to reflect a community's values or align to safety standards). However, current systems fall short on both of these dimensions. To address this gap, we present SafetyAnalyst, a novel LLM safety moderation framework. Given a prompt, SafetyAnalyst creates a structured "harm-benefit tree," which identifies 1) the actions that could be taken if a compliant response were provided, 2) the harmful and beneficial effects of those actions (along with their likelihood, severity, and immediacy), and 3) the stakeholders that would be impacted by those effects. It then aggregates this structured representation into a harmfulness score based on a parameterized set of safety preferences, which can be transparently aligned to particular values. Using extensive harm-benefit features generated by SOTA LLMs on 19k prompts, we fine-tuned an open-weight LM to specialize in generating harm-benefit trees through symbolic knowledge distillation. On a comprehensive set of prompt safety benchmarks, we show that our system (average F1=0.75) outperforms existing LLM safety moderation systems (average F1<0.72) on prompt harmfulness classification, while offering the additional advantages of interpretability and steerability.

Large language models have led to remarkable progress on many NLP tasks, and researchers are turning to ever-larger text corpora to train them. Some of the largest corpora available are made by scraping significant portions of the internet, and are frequently introduced with only minimal documentation. In this work we provide some of the first documentation for the Colossal Clean Crawled Corpus (C4; Raffel et al., 2020), a dataset created by applying a set of filters to a single snapshot of Common Crawl. We begin by investigating where the data came from, and find a significant amount of text from unexpected sources like patents and US military websites. Then we explore the content of the text itself, and find machine-generated text (e.g., from machine translation systems) and evaluation examples from other benchmark NLP datasets. To understand the impact of the filters applied to create this dataset, we evaluate the text that was removed, and show that blocklist filtering disproportionately removes text from and about minority individuals. Finally, we conclude with some recommendations for how to created and document web-scale datasets from a scrape of the internet.

AI agents are increasingly being deployed to automate tasks, often based on underspecified user instructions. Making unwarranted assumptions to compensate for the missing information and failing to ask clarifying questions can lead to suboptimal outcomes, safety risks due to tool misuse, and wasted computational resources. In this work, we study the ability of LLM agents to handle underspecified instructions in interactive code generation settings by evaluating proprietary and open-weight models on their performance across three key steps: (a) detecting underspecificity, (b) asking targeted clarification questions, and (c) leveraging the interaction to improve performance in underspecified scenarios. We introduce Ambig-SWE, an underspecified variant of SWE-Bench Verified, specifically designed to evaluate agent behavior under ambiguity and interaction. Our findings reveal that models struggle to distinguish between well-specified and underspecified instructions. However, when models interact for underspecified inputs, they effectively obtain vital information from the user leading to significant improvements in performance, up to 74% over the non-interactive settings, underscoring the value of effective interaction. Our study highlights critical gaps in how current state-of-the-art models handle missing information in complex software engineering tasks and structures the evaluation into distinct steps to enable targeted improvements.

The escalating debate on AI's capabilities warrants developing reliable metrics to assess machine "intelligence". Recently, many anecdotal examples were used to suggest that newer large language models (LLMs) like ChatGPT and GPT-4 exhibit Neural Theory-of-Mind (N-ToM); however, prior work reached conflicting conclusions regarding those abilities. We investigate the extent of LLMs' N-ToM through an extensive evaluation on 6 tasks and find that while LLMs exhibit certain N-ToM abilities, this behavior is far from being robust. We further examine the factors impacting performance on N-ToM tasks and discover that LLMs struggle with adversarial examples, indicating reliance on shallow heuristics rather than robust ToM abilities. We caution against drawing conclusions from anecdotal examples, limited benchmark testing, and using human-designed psychological tests to evaluate models.

Humans learn social skills through both imitation and social interaction. This social learning process is largely understudied by existing research on building language agents. Motivated by this gap, we propose an interactive learning method, SOTOPIA-pi, improving the social intelligence of language agents. This method leverages behavior cloning and self-reinforcement training on filtered social interaction data according to large language model (LLM) ratings. We show that our training method allows a 7B LLM to reach the social goal completion ability of an expert model (GPT-4-based agent), while improving the safety of language agents and maintaining general QA ability on the MMLU benchmark. We also find that this training paradigm uncovers some difficulties in LLM-based evaluation of social intelligence: LLM-based evaluators overestimate the abilities of the language agents trained specifically for social interaction.

Humans intuitively navigate social interactions by simulating unspoken dynamics and reasoning about others' perspectives, even with limited information. In contrast, AI systems struggle to automatically structure and reason about these implicit social contexts. In this paper, we introduce a novel structured social world representation formalism (S3AP), designed to help AI systems reason more effectively about social dynamics. Following a POMDP-driven design, S3AP represents social interactions as structured tuples, such as state, observation, agent actions, and mental states, which can be automatically induced from free-form narratives or other inputs. We first show S3AP can help LLMs better understand social narratives across 5 social reasoning tasks (e.g., +51% improvement on FANToM's theory-of-mind reasoning with OpenAI's o1), reaching new state-of-the-art (SOTA) performance. We then induce social world models from these structured representations, demonstrating their ability to predict future social dynamics and improve agent decision-making, yielding up to +18% improvement on the SOTOPIA social interaction benchmark. Our findings highlight the promise of S3AP as a powerful, general-purpose representation for social world states, enabling the development of more socially-aware systems that better navigate social interactions.

Humans learn social norms and behaviors from verbal feedback (e.g., a parent saying "that was rude" or a friend explaining "here's why that hurt"). Yet, learning from feedback for LLMs has largely focused on domains like code and math, where RL rewards are directly verifiable and condensed into scalar values. As LLMs are increasingly used to simulate human behavior, e.g., standing in for users, patients, students, and other personas, there is a pressing need to make them more human-like, which requires embracing a fundamentally different kind of signal: feedback that is verbal, subjective, and multi-faceted. We present DITTO, a model trained by treating verbal feedback as a first-class signal in reinforcement learning. After each rollout, DITTO receives verbal feedback and generates a feedback-conditioned improved rollout; both outputs are jointly optimized with GRPO, distilling verbal guidance into the base policy without requiring feedback at test time. We also introduce SOUL (Simulation gym Of hUman-Like behavior), a unified benchmark and training data suite spanning 10 tasks across six categories: Theory of Mind, character role play, social skill, learner simulation, user simulation, and persona simulation. DITTO achieves an average 36% improvement over the base model and exceeds GPT-5.4 on 6 of 10 SOUL benchmarks, demonstrating that RL with verbal feedback is a promising direction for training LLMs to simulate human behavior.

Foundation Models that are capable of processing and generating multi-modal data have transformed AI's role in medicine. However, a key limitation of their reliability is hallucination, where inaccurate or fabricated information can impact clinical decisions and patient safety. We define medical hallucination as any instance in which a model generates misleading medical content. This paper examines the unique characteristics, causes, and implications of medical hallucinations, with a particular focus on how these errors manifest themselves in real-world clinical scenarios. Our contributions include (1) a taxonomy for understanding and addressing medical hallucinations, (2) benchmarking models using medical hallucination dataset and physician-annotated LLM responses to real medical cases, providing direct insight into the clinical impact of hallucinations, and (3) a multi-national clinician survey on their experiences with medical hallucinations. Our results reveal that inference techniques such as Chain-of-Thought (CoT) and Search Augmented Generation can effectively reduce hallucination rates. However, despite these improvements, non-trivial levels of hallucination persist. These findings underscore the ethical and practical imperative for robust detection and mitigation strategies, establishing a foundation for regulatory policies that prioritize patient safety and maintain clinical integrity as AI becomes more integrated into healthcare. The feedback from clinicians highlights the urgent need for not only technical advances but also for clearer ethical and regulatory guidelines to ensure patient safety. A repository organizing the paper resources, summaries, and additional information is available at https://github.com/mitmedialab/medical hallucination.

Large language models are increasingly deployed as human simulators for interactive evaluation and social simulation. Yet helpfulness-driven post-training pulls them toward a homogeneous, overly agreeable assistant register, creating a behavioral Sim2Real gap. We present OdysSim, the largest open systematic investigation of behavioral foundation models, i.e., models trained to simulate human behavior at scale. We propose SOUL, a taxonomy of five capability axes (CONV, SS, COG, ROLE, EVAL) that unifies 62 datasets and 23 benchmark tasks under one framework. Specifically, we curate the OdysSim corpus (21.4M interactions, 10B tokens, retrofitted with back-generated social contexts), construct the SOUL-Index benchmark, and develop an end-to-end training recipe combining midtraining, task-specific RL, and expert distillation. The resulting open 8B OSim model ranks first or tied-first on 8 of 23 tasks, outperforming any individual frontier model by this count, with the strongest gains on conversational and social tasks. Its outputs are also more human-like in length, formatting, and word choice, and it transfers zero-shot to out-of-distribution user simulation on τ-bench, nearly matching real users on reaction alignment (93.2 vs. 93.5). We further show that LLM-as-judge RL induces reward-hacking patterns, and that our detectors can mitigate them during post-training. Together, our findings suggest that behavioral foundation models require rethinking the LLM training paradigm. We release all artifacts to support future research.

Most benchmark datasets targeting commonsense reasoning focus on everyday scenarios: physical knowledge like knowing that you could fill a cup under a waterfall [Talmor et al., 2019], social knowledge like bumping into someone is awkward [Sap et al., 2019], and other generic situations. However, there is a rich space of commonsense inferences anchored to knowledge about specific entities: for example, deciding the truthfulness of a claim "Harry Potter can teach classes on how to fly on a broomstick." Can models learn to combine entity knowledge with commonsense reasoning in this fashion? We introduce CREAK, a testbed for commonsense reasoning about entity knowledge, bridging fact-checking about entities (Harry Potter is a wizard and is skilled at riding a broomstick) with commonsense inferences (if you're good at a skill you can teach others how to do it). Our dataset consists of 13k human-authored English claims about entities that are either true or false, in addition to a small contrast set. Crowdworkers can easily come up with these statements and human performance on the dataset is high (high 90s); we argue that models should be able to blend entity knowledge and commonsense reasoning to do well here. In our experiments, we focus on the closed-book setting and observe that a baseline model finetuned on existing fact verification benchmark struggles on CREAK. Training a model on CREAK improves accuracy by a substantial margin, but still falls short of human performance. Our benchmark provides a unique probe into natural language understanding models, testing both its ability to retrieve facts (e.g., who teaches at the University of Chicago?) and unstated commonsense knowledge (e.g., butlers do not yell at guests).

Large language models deployed as agents over large tool catalogs face a critical tool-retrieval bottleneck. As embedding-based retrieval approaches rely on compact encoders that may under-capture specialized tool semantics, parametric tool retrieval addresses this by encoding each tool as a virtual token appended to the LLM vocabulary, fine-tuned in two stages (memorization then retrieval SFT) to use the LLM as a retriever, achieving strong performance on standard ToolBench retrieval benchmarks. Yet these benchmarks use verbose, fully-specified queries, and their evaluation applies constrained decoding that restricts outputs to valid token paths, neither reveals whether the model actually understands its tools. We introduce ToolSense, an open-source LLM-powered diagnostic framework that takes any tool catalog as input and automatically generates three benchmarks: a Realistic Retrieval Benchmark (RRB) with queries at three ambiguity tiers, an MCQ probing benchmark, and a QA probing benchmark. Applying ToolSense to ToolBench (~47k tools) and evaluating five parametric model training configurations reveals a knowledge-retrieval dissociation: on RRB queries, several configurations collapse by ~50-64 percentage points compared to fully-specified ToolBench benchmarks, falling below the embedding-model baseline. Additionally, despite strong retrieval performance, some models score near-random on factual probes, suggesting a knowledge-retrieval dissociation. We open-source the ToolSense framework and the ToolBench diagnostic benchmarks at https://github.com/SAP/toolsense.

Recent Vision-Language Models (e.g., ColPali) enable fine-grained Visual Document Retrieval (VDR) but incur prohibitive multi-vector index storage overhead. Existing training-free pruning methods either rely on heuristic layer choices or degrade sharply under aggressive compression, leading prior work to argue that effective high-compression pruning requires query-dependent training. We challenge this view with Structural Anchor Pruning (SAP), a self-calibrating, training-free, and query-agnostic index-time pruning framework with three components: (i) Score Retention (SR), a white-box per-layer compression diagnostic; (ii) SR-guided window selection, a procedure that automatically locates the structural pruning region for any backbone with no per-model hyperparameters; and (iii) a visual in-degree centrality scorer that identifies anchor patches within the selected window. On the ViDoRe v1/v2 benchmarks across three architectures spanning 18, 28, and 36 backbone layers, SAP retains over 90\% of NDCG@5 while pruning more than 90\% of visual tokens, without any per-model parameter tuning. Our layer-resolved SR analysis reveals an Alignment-Aggregation Divergence: the document's visual structure is preserved as a stable ``Structural Plateau'' within the backbone, but the final layers reshape this representation into a sparse, query-aligned form that is no longer suitable for pruning. This is the mechanistic reason SAP succeeds where final-layer methods fail.

Few-shot adaptation for Vision-Language Models (VLMs) presents a dilemma: balancing in-distribution accuracy with out-of-distribution generalization. Recent research has utilized low-level concepts such as visual attributes to enhance generalization. However, this study reveals that VLMs overly rely on a small subset of attributes on decision-making, which co-occur with the category but are not inherently part of it, termed spuriously correlated attributes. This biased nature of VLMs results in poor generalization. To address this, 1) we first propose Spurious Attribute Probing (SAP), identifying and filtering out these problematic attributes to significantly enhance the generalization of existing attribute-based methods; 2) We introduce Spurious Attribute Shielding (SAS), a plug-and-play module that mitigates the influence of these attributes on prediction, seamlessly integrating into various Parameter-Efficient Fine-Tuning (PEFT) methods. In experiments, SAP and SAS significantly enhance accuracy on distribution shifts across 11 datasets and 3 generalization tasks without compromising downstream performance, establishing a new state-of-the-art benchmark.