Tag
This paper investigates offline reinforcement learning for post-training code-generating LLMs, showing that it can improve zero-shot code generation performance using existing datasets without online sampling.
Proposes CodeBlock, a structure-aware sparse supervision framework for supervised fine-tuning of code LLMs. It selects high-quality instruction-response pairs and partitions code responses into syntactically coherent coding items, applying loss only to selected items to achieve stronger pass@1 rates using only 1.9% of supervised response tokens.
RepoExec is a new benchmark for evaluating repository-level code generation focused on executability, functional correctness via automated test generation, and cross-file dependency integration. Experiments reveal that pretrained LLMs beat instruction-tuned ones on correctness, while fine-tuned models better leverage provided code dependencies.