Training expert LLMs in domains with scarce data is difficult, often relying on multiple-choice questions (MCQs). However, standard outcome-based reinforcement learning (RL) on MCQs is risky. While it may improve accuracy, we observe it often degrades reasoning quality such as logical consistency. Existing solutions to supervise reasoning, such as large-scale Process Reward Models (PRMs), are prohibitively expensive. To address this, we propose CLARity, a cost-effective RL framework that enhances reasoning quality using only a small, general-purpose LLM. CLARity integrates a consistency-aware reward mechanism with a 2-stage refine-then-monitor training pipeline to enhance reasoning consistency, and a dynamic data reformulation strategy to to better exploit limited data. Experiments demonstrate that CLARity improves response consistency by 16.5% and accuracy by 7.5% over baselines. Human evaluations further confirm holistic improvements in coherence and professionalism. Thus, CLARity offers a generalizable solution that enables smaller models to effectively guide expert models by reasoning consistency.Our code is open sourced at: https://github.com/Infinite-set/CLARity

Medical diagnostics is a high-stakes and complex domain that is critical to patient care. However, current evaluations of large language models (LLMs) are fundamentally misaligned with real-world clinical practice. Most of them rely on static benchmarks derived from public medical exam items, which tend to overestimate model performance and ignore the difference between textbook cases and the ambiguous, varying conditions in the real world. Recent efforts toward dynamic evaluation offer a promising alternative, but their improvements are limited to superficial perturbations and a narrow focus on accuracy. To address these gaps, we propose DyReMe, a dynamic benchmark for medical diagnostics that better reflects real clinical practice. Unlike static exam-style questions, DyReMe generates fresh, consultation-like cases that introduce distractors such as differential diagnoses and common misdiagnosis factors. It also varies expression styles to mimic diverse real-world query habits. Beyond accuracy, DyReMe evaluates LLMs on three additional clinically relevant dimensions: veracity, helpfulness, and consistency. Our experiments demonstrate that this dynamic approach yields more challenging and realistic assessments, revealing significant misalignments between the performance of state-of-the-art LLMs and real clinical practice. These findings highlight the urgent need for evaluation frameworks that better reflect the demands of trustworthy medical diagnostics.

The growing ubiquity of conversational AI highlights the need for frameworks that capture not only users' instrumental goals but also the situated, adaptive, and social practices through which they achieve them. Existing taxonomies of conversational behavior either overgeneralize, remain domain-specific, or reduce interactions to narrow dialogue functions. To address this gap, we introduce the Taxonomy of User Needs and Actions (TUNA), an empirically grounded framework developed through iterative qualitative analysis of 1193 human-AI conversations, supplemented by theoretical review and validation across diverse contexts. TUNA organizes user actions into a three-level hierarchy encompassing behaviors associated with information seeking, synthesis, procedural guidance, content creation, social interaction, and meta-conversation. By centering user agency and appropriation practices, TUNA enables multi-scale evaluation, supports policy harmonization across products, and provides a backbone for layering domain-specific taxonomies. This work contributes a systematic vocabulary for describing AI use, advancing both scholarly understanding and practical design of safer, more responsive, and more accountable conversational systems.

We study a distributed learning problem in which learning agents are embedded in a directed acyclic graph (DAG). There is a fixed and arbitrary distribution over feature/label pairs, and each agent or vertex in the graph is able to directly observe only a subset of the features -- potentially a different subset for every agent. The agents learn sequentially in some order consistent with a topological sort of the DAG, committing to a model mapping observations to predictions of the real-valued label. Each agent observes the predictions of their parents in the DAG, and trains their model using both the features of the instance that they directly observe, and the predictions of their parents as additional features. We ask when this process is sufficient to achieve \emph{information aggregation}, in the sense that some agent in the DAG is able to learn a model whose error is competitive with the best model that could have been learned (in some hypothesis class) with direct access to \emph{all} features, despite the fact that no single agent in the network has such access. We give upper and lower bounds for this problem for both linear and general hypothesis classes. Our results identify the \emph{depth} of the DAG as the key parameter: information aggregation can occur over sufficiently long paths in the DAG, assuming that all of the relevant features are well represented along the path, and there are distributions over which information aggregation cannot occur even in the linear case, and even in arbitrarily large DAGs that do not have sufficient depth (such as a hub-and-spokes topology in which the spoke vertices collectively see all the features). We complement our theoretical results with a comprehensive set of experiments.

Large Language Models (LLMs) often generate responses with inherent biases, undermining their reliability in real-world applications. Existing evaluation methods often overlook biases in long-form responses and the intrinsic variability of LLM outputs. To address these challenges, we propose FiSCo (Fine-grained Semantic Comparison), a novel statistical framework to evaluate group-level fairness in LLMs by detecting subtle semantic differences in long-form responses across demographic groups. Unlike prior work focusing on sentiment or token-level comparisons, FiSCo goes beyond surface-level analysis by operating at the claim level, leveraging entailment checks to assess the consistency of meaning across responses. We decompose model outputs into semantically distinct claims and apply statistical hypothesis testing to compare inter- and intra-group similarities, enabling robust detection of subtle biases. We formalize a new group counterfactual fairness definition and validate FiSCo on both synthetic and human-annotated datasets spanning gender, race, and age. Experiments show that FiSCo more reliably identifies nuanced biases while reducing the impact of stochastic LLM variability, outperforming various evaluation metrics.

Data contamination poses a significant threat to the reliable evaluation of Large Language Models (LLMs). This issue arises when benchmark samples may inadvertently appear in training sets, compromising the validity of reported performance. While detection methods have been developed for the pre-training and Supervised Fine-Tuning stages, a critical research gap exists for the increasingly significant phase of Reinforcement Learning (RL) post-training. As RL post-training becomes pivotal for advancing LLM reasoning, the absence of specialized contamination detection methods in this paradigm presents a critical vulnerability. To address this, we conduct the first systematic study of data detection within RL post-training scenario and propose Self-Critique. Our method is motivated by a key observation: after RL phase, the output entropy distribution of LLMs tends to collapse into highly specific and sparse modes. Self-Critique probes for the underlying policy collapse, i.e., the model's convergence to a narrow reasoning path, which causes this entropy reduction. To facilitate this research, we also introduce RL-MIA, a benchmark constructed to simulate this specific contamination scenario. Extensive experiments show that Self-Critique significantly outperforms baseline methods across multiple models and contamination tasks, achieving an AUC improvement of up to 30%. Whereas existing methods are close to a random guess for RL-phase contamination, our method makes detection possible.

Learning from demonstrations enables experts to teach robots complex tasks using interfaces such as kinesthetic teaching, joystick control, and sim-to-real transfer. However, these interfaces often constrain the expert's ability to demonstrate optimal behavior due to indirect control, setup restrictions, and hardware safety. For example, a joystick can move a robotic arm only in a 2D plane, even though the robot operates in a higher-dimensional space. As a result, the demonstrations collected by constrained experts lead to suboptimal performance of the learned policies. This raises a key question: Can a robot learn a better policy than the one demonstrated by a constrained expert? We address this by allowing the agent to go beyond direct imitation of expert actions and explore shorter and more efficient trajectories. We use the demonstrations to infer a state-only reward signal that measures task progress, and self-label reward for unknown states using temporal interpolation. Our approach outperforms common imitation learning in both sample efficiency and task completion time. On a real WidowX robotic arm, it completes the task in 12 seconds, 10x faster than behavioral cloning, as shown in real-robot videos on https://sites.google.com/view/constrainedexpert .

As Artificial Intelligence (AI) technologies continue to advance, protecting human autonomy and promoting ethical decision-making are essential to fostering trust and accountability. Human agency (the capacity of individuals to make informed decisions) should be actively preserved and reinforced by AI systems. This paper examines strategies for designing AI systems that uphold fundamental rights, strengthen human agency, and embed effective human oversight mechanisms. It discusses key oversight models, including Human-in-Command (HIC), Human-in-the-Loop (HITL), and Human-on-the-Loop (HOTL), and proposes a risk-based framework to guide the implementation of these mechanisms. By linking the level of AI model risk to the appropriate form of human oversight, the paper underscores the critical role of human involvement in the responsible deployment of AI, balancing technological innovation with the protection of individual values and rights. In doing so, it aims to ensure that AI technologies are used responsibly, safeguarding individual autonomy while maximizing societal benefits.

The performance of Large Language Models (LLMs) is highly sensitive to the prompts they are given. Drawing inspiration from the field of prompt optimization, this study investigates the potential for enhancing Automated Essay Scoring (AES) by refining the scoring rubrics used by LLMs. Specifically, our approach prompts models to iteratively refine rubrics by reflecting on models' own scoring rationales and observed discrepancies with human scores on sample essays. Experiments on the TOEFL11 and ASAP datasets using GPT-4.1, Gemini-2.5-Pro, and Qwen-3-Next-80B-A3B-Instruct show Quadratic Weighted Kappa (QWK) improvements of up to 0.19 and 0.47, respectively. Notably, even with a simple initial rubric, our approach achieves comparable or better QWK than using detailed human-authored rubrics. Our findings highlight the importance of iterative rubric refinement in LLM-based AES to enhance alignment with human evaluations.

State-of-the-art (SOTA) LLMs have progressed from struggling on proof-based Olympiad problems to solving most of the IMO 2025 problems, with leading systems reportedly handling 5 of 6 problems. Given this progress, we assess how well these models can grade proofs: detecting errors, judging their severity, and assigning fair scores beyond binary correctness. We study proof-analysis capabilities using a corpus of 90 Gemini 2.5 Pro-generated solutions that we grade on a 1-4 scale with detailed error annotations, and on MathArena solution sets for IMO/USAMO 2025 scored on a 0-7 scale. Our analysis shows that models can reliably flag incorrect (including subtly incorrect) solutions but exhibit calibration gaps in how partial credit is assigned. To address this, we introduce agentic workflows that extract and analyze reference solutions and automatically derive problem-specific rubrics for a multi-step grading process. We instantiate and compare different design choices for the grading workflows, and evaluate their trade-offs. Across our annotated corpus and MathArena, our proposed workflows achieve higher agreement with human grades and more consistent handling of partial credit across metrics. We release all code, data, and prompts/logs to facilitate future research.