DHRCL Boosts Code LLM Training with Hierarchical Rewards.

Training large language models (LLMs) for code generation often uses reinforcement learning, where generated programs are evaluated. However, existing methods frequently rely on sparse outcome rewards or combine dense signals statically, overlooking the progressive dependencies between programming capabilities like syntax, executability, and functional correctness. Researchers propose DHRCL, a novel reinforcement learning framework that employs Dense Hierarchical Rewards and Curriculum Learning. DHRCL breaks down feedback into distinct levels: syntax validation, execution success, unit-test pass rates, and abstract syntax tree (AST)-based structural similarity. These signals are then structured into a three-stage curriculum: Syntax, Execution, and Pass & Structural. A key innovation is the automatic determination of stage duration based on recent validation trends, eliminating manual thresholding. DHRCL also introduces stage-aware probability-based token credit redistribution, which adapts how credit or blame is assigned to tokens based on the current curriculum stage, emphasizing established patterns early and refining less-established decisions later. Experiments across various Qwen3 backbones demonstrate DHRCL's consistent advantage over other reward baselines.