From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning

Researchers introduced SCRL, a curriculum reinforcement learning method for LLM reasoning.

SCRL turns reference reasoning chains into verifiable subproblems and normalizes rewards at the subproblem level to improve credit assignment.

On seven math benchmarks, it outperformed strong curriculum-learning baselines and boosted accuracy and pass rates on difficult exams.

The approach tackles a core training bottleneck: hard problems often yield too few correct final answers to learn from effectively.