Research

Stable RL with verifiable rewards requires both reliable baselines and diverse training prompts. Group-based methods improve baseline estimates by sampling more responses per prompt, but this leaves fewer distinct prompts under a fixed rollout budget. POISE estimates baselines with a lightweight probe of the actor's hidden states and entropy, without a separate LLM-scale critic. Predicting each rollout's baseline from an independent rollout keeps the policy gradient unbiased with two rollouts per prompt. This allows greater prompt diversity and improves average performance and training stability over the evaluated baselines in multi-domain training.

Prompting an LLM to express a value can produce different behavior from the model expressing that value without instructions. We examine whether the two rely on the same internal mechanisms. We identify shared neurons that causally contribute to both forms of value expression, alongside components with distinct roles. Components specific to intrinsic expression promote response diversity, whereas those specific to prompted expression strengthen instruction compliance, even on tasks unrelated to values.

A role-playing agent should reflect how a character's values, motivations, and relationships change throughout a story. Yet existing benchmarks largely treat characters as fixed personas or test only what they know at a given point. ArcANE tests whether an agent's responses fit the character's development by presenting the same situation at different narrative stages, including situations outside the original story. Across six models, providing a record of the character's development up to the queried stage improves this fidelity over the evaluated narrative-context baselines.

Predicting whether a message will persuade someone requires knowing which parts of their history matter for that discussion. Retrieving similar posts or building a generic profile may miss this information. We train a query generator to retrieve useful records and a profiler to summarize them, using their effect on a fixed model's view-change predictions as supervision. No ground-truth profiles are needed. On ChangeMyView, this improves prediction across multiple predictor models. Our analyses show that useful records and profile dimensions vary with both the discussion and the predictor.

RL improves vision-language-action (VLA) models, but which parameter changes drive these gains remains unclear. Across the flow-based VLAs we study, low-rank RL updates concentrate in timestep modules. These modules depend only on denoising time, not observations or instructions, yet module-replacement experiments show that their updates capture a large share of RL gains. We link the low-rank structure to training on discrete denoising timesteps. Targeting these modules with LoRA outperforms standard configurations, which typically omit them. Steering along their learned shift directions further improves RL-trained policies without additional RL training.

Negating a question should exclude the original answer, yet LLMs often repeat it, especially when they were confident in that answer. In the models we analyze, negation suppresses original-answer retrieval and promotes favored candidates within the answer category. However, these preferences do not reliably exclude the original answer, and the correction can be too weak to overturn it. We therefore propose ALiT, which requires larger corrections for more confident original predictions. ALiT reduces repetition while preserving the answer category and avoiding the sharp diversity loss observed with supervised fine-tuning.

When a language model changes its answer under negation, does it alter factual retrieval or only the final prediction? We study Qwen3 and Gemma 3 on prompts they answer correctly before negation and whose predictions change after negation. Negation interacts with relation information early, while later components shift the prediction away from the original answer. This suggests that negation modifies factual retrieval itself, rather than acting only on a completed prediction.

Biomedical benchmarks often focus on multiple-choice questions, but long-form answers can mix accurate and inaccurate claims that a single correctness judgment cannot capture. We introduce DAHL, a benchmark of 8,573 questions generated from research articles across 29 biomedical categories. Its automated evaluation pipeline splits each answer into individual claims and scores their factual accuracy, allowing evaluation of which parts of a long-form answer are correct.