Bootstrapping Niche Multilingual Code Translation via Reinforcement Learning with Execution-Based Verifiable Supervision

A reinforcement learning pipeline optimizes niche, many-to-many code translation by using execution-based feedback to ensure behavioral preservation. The method expands Python seed programs into a multilingual pool to train a reward model, which then drives Group Relative Policy Optimization (GRPO) across 600 directed language pairs.