Training code generation models to debug their own outputs

Amazon researchers published a NeurIPS 2024 paper on training LLMs to self-debug code using supervised fine-tuning (SFT) and reinforcement learning (RL), moving beyond few-shot prompting approaches. They generated synthetic debugging training data via LLMs to address data scarcity, and found that fine-tuned models produced better initial code generations in addition to stronger debugging. The work directly advances agentic coding capabilities — models that can generate, test, and iteratively fix their own outputs.