Towards multi-language repository-level code generation: From-scratch to guided tasks

Jingjing Liu

doi:10.1016/j.neucom.2026.133204

Towards multi-language repository-level code generation: From-scratch to guided tasks

2026-02-27

0

PRE

AI

DOI：10.1016/j.neucom.2026.133204

Original

Original request for help

Save

Abstract

En 中文

• We introduce a repository-level code generation benchmark with multi-language support and varying task complexities which injects requirement comments as positive noise. • We propose RepoGenesis, a GRPO-based reinforcement learning framework with 3 distinct reward signals: (i) structure reward, (ii) syntactic reward, and (iii) semantic reward. These rewards respectively guide the repository structure, enforce syntactic correctness, and ensure the semantic correctness of the generated code. • Experiment results highlight the limitations of current LLMs and demonstrate that RepoGenesis enables Qwen2.5-Coder-7B to achieves the performance comparable to Claude Sonnet 4 (>100B).

Keywords:

code generation

repository-level

multi-language

reinforcement learning

semantic correctness

Towards multi-language repository-level code generation: From-scratch to guided tasks

Abstract

Journal

Organization