Definition
Reinforcement learning for reasoning trains models to produce useful chains of thought by rewarding correct final answers (or verified intermediate steps) and letting the model figure out the reasoning that gets there. Most of the 2024–2026 jump in math and code performance has roots here.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- LUFFY: Learning to Reason Under Off-Policy Guidance
- RLVR is Not RL: Understanding RLVR via Reweighted Sampling
- Demystifying Long Chain-of-Thought Reasoning in LLMs
- Process Reward Models to Align Embodied Agent Learning
- CURL: Contrastive Unsupervised Representations for Reinforcement Learning
- Let's Verify Step by Step
- Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Model Parameters
- STaR: Bootstrapping Reasoning With Reasoning
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning