analiza aiqbench
This article introduces DHRCL, a reinforcement learning framework designed specifically for post-training code-oriented large language models by decomposing feedback into hierarchical signals like syntax validity and unit-test pass rates. By implementing an automatic curriculum that adjusts stage duration based on validation trends rather than fixed thresholds, the method effectively balances consolidation of established patterns with refinement during functional optimization. The approach demonstrates consistent performance improvements across various Qwen model backbones when compared against standard binary or sparse reward baselines for code generation tasks.
DHRCL:Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning
arXiv:2607.26457v1 Announce Type: new Abstract: Reinforcement learning is a natural post-training paradigm for code-oriented large language models because generated programs can be evaluated through parsing, execution, unit tests, and structural analysis.However, existing methods often rely on sparse outcome rewards or statically combine heterogeneous dense signals, even though syntax validity, executability, functional correctness, and structural organization describe different and progressively dependent programming capabilities. We propose DHRCL, a reinforcement learning framework with Dense Hierarchical Rewards and Curriculum Learning. DHRCL decomposes feedback into syntax validation, execution success, unit-test pass rate, and AST-based structural similarity, and organizes these signals through a three-stage Syntax, Execution, Pass & Structural curriculum. Stage duration is determined automatically from recent validation trends rather than manually specified capability thresholds. We further introduce stage-aware probability-based token credit redistribution. The mechanism follows a consolidation-to-refinement principle: it emphasizes established token patterns during syntax-oriented optimization, applies uniform propagation for non-local execution feedback, and allocates more credit or blame to less-established token decisions during final functional optimization. Under a unified Qwen3-8B and KodCode protocol, the experiments compare DHRCL with binary, pass-rate, reward-model-based, and verifiable dense-reward baselines. We further evaluate DHRCL across Qwen3-4B, Qwen3-8B, and Qwen3-14B backbones, showing that its advantage remains consistent as model capacity increases.
Abstract:Reinforcement learning is a natural post-training paradigm for code-oriented large language models because generated programs can be evaluated through parsing, execution, unit tests, and structural this http URL, existing methods often rely on sparse outcome rewards or statically combine heterogeneous dense signals, even though syntax validity, executability, functional correctness, and structural organization describe different and progressively dependent programming capabilities. We propose DHRCL, a reinforcement learning framework with Dense Hierarchical Rewards and Curriculum Learning. DHRCL decomposes feedback into syntax validation, execution success, unit-test pass rate, and AST-based structural similarity, and organizes these signals through a three-stage Syntax, Execution, Pass & Structural curriculum. Stage duration is determined automatically from recent validation trends rather than manually specified capability thresholds. We further introduce stage-aware probability-based token credit redistribution. The mechanism follows a consolidation-to-refinement principle: it emphasizes established token patterns during syntax-oriented optimization, applies uniform propagation for non-local execution feedback, and allocates more credit or blame to less-established token decisions during final functional optimization. Under a unified Qwen3-8B and KodCode protocol, the experiments compare DHRCL with binary, pass-rate, reward-model-based, and verifiable dense-reward baselines. We further evaluate DHRCL across Qwen3-4B, Qwen3-8B, and Qwen3-14B backbones, showing that its advantage remains consistent as model capacity increases.
Machine Learning (cs.LG)
arXiv:2607.26457 [cs.LG]
(or arXiv:2607.26457v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.26457 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 292 cuvinte
Subiecte conexe
TD-Grokking: Learning from Zero-Reward Problems by Training-Time Decomposition
arXiv · 10 iun. 2026
PIRAMID: Progress and Plans
LessWrong · 27 iul. 2026
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · 5 oct. 2026
Robust Code RL via Faulty-Code-Driven Test case Synthesis and Dense Reward Shaping
arXiv · 27 aug. 2026