Paper summary
Paper summary: FlowRL: Matching Reward Distributions for LLM Reasoning
FlowRL replaces reward maximisation with matching the whole reward distribution (via a learnable partition function and reverse KL), to keep diverse reasoning paths alive; it reports +10.0% over GRPO on maths.
- Reward-maximising RL tends to collapse onto the most common successful path; matching the reward distribution keeps rarer valid paths.
- The scalar reward is turned into a target distribution, and the policy is pulled toward it.