SynthFix: Adaptive Neuro-Symbolic Code Vulnerability Repair

SynthFix is a neuro-symbolic code vulnerability repair framework that combines supervised repair learning with compiler feedback. During training, a lightweight router adaptively selects between SFT and RFT, and the reward integrates static structure, lint/compile checks, security scanning and public execution tests. In evaluations on pyrepair, CodeFlaws and SVEN with five 1.3B-7B code LLMs, SynthFix outperforms SFT-only and RFT-only baselines, with functional correctness improving by up to 54% relative and security pass rate improving by 14%.