Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for post-training large reasoning models (LRMs) using policy-gradient methods such as GRPO. To stabilize training, these methods typically center trajectory rewards by subtracting the empirical mean for each prompt. Statistically, this centering acts as a control variate (or baseline), reducing the variance of the policy-gradient estimator. Typically, the mean reward is estimated using per-prompt empirical averages for each prompt in a batch. Drawing inspiration from Stein's paradox, we propose using shrinkage estimators that combine per-prompt and across-prompt means to improve the overall per-prompt mean estimation accuracy -- particularly in the low-generation regime typical of RLVR. Theoretically, we construct a shrinkage-based baseline that provably yields lower-variance policy-gradient estimators across algorithms. Our proposed baseline serves as a drop-in replacement for existing per-prompt mean baselines, requiring no additional hyper-parameters or computation. Empirically, shrinkage baselines consistently outperform standard empirical-mean baselines, leading to lower-variance gradient updates and improved training stability.

Masked diffusion models (MDMs) offer a compelling alternative to autoregressive models (ARMs) for discrete text generation because they enable parallel token sampling, rather than sequential, left-to-right generation. This means potentially much faster inference. However, effective parallel sampling faces two competing requirements: (i) simultaneously updated tokens must be conditionally independent, and (ii) updates should prioritise high-confidence predictions. These goals conflict because high-confidence predictions often cluster and depend on each other, opportunities for parallel updates. We present PUNT, a model-agnostic sampler that reconciles this trade-off. Our method identifies token dependencies and removes lower-confidence tokens from conflicting groups. This produces sets of indices for unmasking that satisfy both independence and confidence criteria. Our approach ensures improved parallel unmasking through approximate conditional independence testing. Our experiments show that PUNT delivers a superior trade-off between accuracy and compute when compared to other strong training-free baselines, especially for generation of longer sequences. On the IFEval benchmark, it achieves up to 16\% higher accuracy over baseline methods, including sequential generation (one-by-one). These gains hold across different values of hyperparameters, mitigating the need for brittle hyperparameter tuning. Moreover, we observe that PUNT induces an emergent hierarchical generation strategy, where the model first establishes high-level paragraph structure before local refinement, suggesting a planning-like generation process that contributes to strong alignment performance.

Understanding uncertainty in large language models remains a fundamental challenge, particularly in creative tasks where multiple valid outputs exist. We present a geometric framework using credal sets - convex hulls of probability distributions - to quantify and decompose uncertainty in neural text generation, calibrated against human creative variation. Analyzing 500 creative writing prompts from the WritingPrompts dataset with 10 unique human continuations each, we evaluate four language models across five decoding strategies, generating 100,000 stories. Our credal set analysis reveals substantial gaps in capturing human creative variation, with the best model-human calibration reaching only 0.434 (Gemma-2B with temperature 0.7). We decompose total uncertainty into epistemic and aleatoric components, finding that the choice of decoding strategy contributes 39.4% to 72.0% of total epistemic uncertainty. Model scale shows weak correlation with calibration quality and no significant difference exists between base and instruction-tuned models in calibration quality. Our geometric framework provides actionable insights for improving generation systems for human-AI creative alignment. We release our complete experimental framework.

Can LLM Agents Solve Collaborative T asks? Abstract -- The ability to coordinate actions across multiple agents is critical for solving complex, real-world problems. Large Language Models (LLMs) have shown strong capabilities in communication, planning, and reasoning, raising the question of whether they can also support effective collaboration in multi-agent settings. In this work, we investigate the use of LLM agents to solve a structured victim rescue task that requires division of labor, prioritization, and cooperative planning. Agents operate in a fully known graph-based environment and must allocate resources to victims with varying needs and urgency levels. We systematically evaluate their performance using a suite of coordination-sensitive metrics, including task success rate, redundant actions, room conflicts, and urgency-weighted efficiency. This study offers new insights into the strengths and failure modes of LLMs in physically grounded multi-agent collaboration tasks, contributing to future benchmarks and architectural improvements.

This paper evaluates the capabilities of 28 large language models (LLMs) to reason with 20 defeasible reasoning patterns involving generic generalizations (e.g., 'Birds fly', 'Ravens are black') central to non-monotonic logic. Generics are of special interest to linguists, philosophers, logicians, and cognitive scientists because of their complex exception-permitting behaviour and their centrality to default reasoning, cognition, and concept acquisition. We find that while several frontier models handle many default reasoning problems well, performance varies widely across models and prompting styles. Few-shot prompting modestly improves performance for some models, but chain-of-thought (CoT) prompting often leads to serious performance degradation (mean accuracy drop -11.14%, SD 15.74% in models performing above 75% accuracy in zero-shot condition, temperature 0). Most models either struggle to distinguish between defeasible and deductive inference or misinterpret generics as universal statements. These findings underscore both the promise and limits of current LLMs for default reasoning.

Large Language Models (LLMs) enable new possibilities for qualitative research at scale, including coding and data annotation. While multi-agent systems (MAS) can emulate human coding workflows, their benefits over single-agent coding remain poorly understood. We conducted an experimental study of how agent persona and temperature shape consensus-building and coding accuracy of dialog segments based on a codebook with 8 codes. Our open-source MAS mirrors deductive human coding through structured agent discussion and consensus arbitration. Using six open-source LLMs (with 3 to 32 billion parameters) and 18 experimental configurations, we analyze over 77,000 coding decisions against a gold-standard dataset of human-annotated transcripts from online math tutoring sessions. Temperature significantly impacted whether and when consensus was reached across all six LLMs. MAS with multiple personas (including neutral, assertive, or empathetic), significantly delayed consensus in four out of six LLMs compared to uniform personas. In three of those LLMs, higher temperatures significantly diminished the effects of multiple personas on consensus. However, neither temperature nor persona pairing lead to robust improvements in coding accuracy. Single agents matched or outperformed MAS consensus in most conditions. Only one model (OpenHermesV2:7B) and code category showed above-chance gains from MAS deliberation when temperature was 0.5 or lower and especially when the agents included at least one assertive persona. Qualitative analysis of MAS collaboration for these configurations suggests that MAS may nonetheless aid in narrowing ambiguous code applications that could improve codebooks and human-AI coding. We contribute new insight into the limits of LLM-based qualitative methods, challenging the notion that diverse MAS personas lead to better outcomes. We open-source our MAS and experimentation code.

To fully leverage the potential of artificial intelligence (AI) systems in a trustworthy manner, it is desirable to couple multiple AI and non-AI systems together seamlessly for constraining and ensuring correctness of the output. This paper introduces a novel parallel discrete event simulation (PDES) based methodology to combine multiple AI and non-AI agents in a causal, rule-based way. Our approach tightly integrates the concept of passage of time, with each agent considered as an entity in the PDES framework and responding to prior requests from other agents. Such coupling mechanism enables the agents to work in a co-operative environment towards a common goal while many tasks run in parallel throughout the simulation. It further enables setting up boundaries to the outputs of the AI agents by applying necessary dynamic constraints using non-AI agents while allowing for scalability through deployment of hundreds of such agents in a larger compute cluster. Distributing smaller AI agents can enable extremely scalable simulations in the future, addressing local memory bottlenecks for model parameter storage. Within a PDES involving both AI and non-AI agents, we break down the problem at hand into structured steps, when necessary, providing a set of multiple choices to the AI agents, and then progressively solve these steps towards a final goal. At each step, the non-AI agents act as unbiased auditors, verifying each action by the AI agents so that certain rules of engagement are followed. We evaluate our approach by solving four problems from four different domains and comparing the results with those from AI models alone. Our results show greater accuracy in solving problems from various domains where the AI models struggle to solve the problems solely by themselves. Results show that overall accuracy of our approach is 68% where as the accuracy of vanilla models is less than 23%.

The question of how to make decisions that maximise the well-being of all persons is very relevant to design language models that are beneficial to humanity and free from harm. We introduce the Greatest Good Benchmark to evaluate the moral judgments of LLMs using utilitarian dilemmas. Our analysis across 15 diverse LLMs reveals consistently encoded moral preferences that diverge from established moral theories and lay population moral standards. Most LLMs have a marked preference for impartial beneficence and rejection of instrumental harm. These findings showcase the 'artificial moral compass' of LLMs, offering insights into their moral alignment.