NeurIPS 2026 paper: RL-trained skills transfer poorly to composed reasoning tasks

StanfordAILab · x · 2026-11-17

A NeurIPS 2026 paper, "Compositional Reasoning in Language Models under Reinforcement Learning Post-Training," finds that training LLMs on individual RL-learned skills transfers poorly to composed tasks, while transfer in the other direction is much stronger — a caution for skill-by-skill RL post-training pipelines.

Related event: Study Finds RL Skills Struggle to Compose into Complex Reasoning(2 posts)→

Original post →

More from Research

Research channel →