EvoCoder: Evolving Code Generation through Population-Based Search and Peer-Guided Adaptation

ACL ARR 2026 January Submission8854 Authors

Keywords: large language models, code generation, evolutionary search, surrogate feedback, program repair

Abstract: Code generation for constraint-heavy problems remains brittle for large language models (LLMs): limited and noisy evaluation signals make iterative refinement prone to early lock-in. We propose EvoCoder, an evolutionary framework that decouples natural-language solution strategies (genotypes) from executable programs (phenotypes) to enable population-based multi-trajectory search. EvoCoder operates in three stages: (1) Genotypic variation and expression, sampling a strategy pool and selecting a diverse subset via k-DPP before code generation; (2) Surrogate feedback environment construction, synthesizing boundary-focused tests and improving feedback reliability through dual-channel verification; and (3) Selection and peer-guided Lamarckian adaptation, where Semantic Spectrum Analysis localizes failure-driving divergences and transfers effective peer logic to repair the elite solution. Experiments on seven benchmarks (HumanEval/MBPP variants, APPS, and CodeContest) show consistent gains over strong baselines across proprietary and open-source backbones, achieving up to 98.2% Pass@1 on HumanEval, 93.7% on MBPP, 44.8% on CodeContest, and 57.3% on APPS. On ColBench's Backend Programming track, EvoCoder also generalizes better than baselines.

Paper Type: Long

Research Area: Code Models

Research Area Keywords: code generation, program repair, evaluation of code models, code reasoning

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 8854