Reinforcement learning has become a standard tool for post-training language models and teaching them to reason, follow instructions, and align with human preferences. Nevertheless, the ingredients that produce this reasoning capability are still only partly understood: reward models, the construction of verifiers, and long chains of thought optimized end to end against sparse reward signals.

This workshop investigates formal characterizations of reasoning and the role of reinforcement learning in promoting reasoning capability. Does post-training introduce new capabilities, or does it sharpen a distribution the base model already contains? How should credit be assigned across tokens when the only feedback is whether the final answer was correct? What happens when the verifier is imperfect, the reward is subject to hacking, or the task admits answers that are difficult to verify? Potential topics include:

RL from human feedback, preference optimization, and reward modeling

RL with verifiable rewards; automated verification for mathematics, code, and formal reasoning

Credit assignment and exploration over long reasoning traces

Theoretical accounts of post-training (e.g., sample complexity and generalization)

Reward hacking, specification gaming, and the robustness of learned verifiers

Multi-turn, agentic, and tool-using settings; RL for language-model agents

Evaluation of reasoning; data and benchmarks for reasoning

The one-day event consists of invited talks and a poster session, and aims to foster connections among researchers in reinforcement learning, natural language processing, learning theory, optimization, statistics, and related areas.