ConceptCoder: Code Vulnerability Detection via Concept Learning Outperforms GPT-5.2 and Claude Opus 4.5

arXiv·high signal

ConceptCoder introduces concept-based fine-tuning that simulates human code inspection — models first learn human-understandable semantic code concepts, then reason over them for vulnerability detection. This is the first work to formally define code concepts for LLMs. Evaluation shows F1 improvement from 66.32 to 72.15 averaged across 9 open-source LLMs, outperforming prompted GPT-5.2 and Claude-Opus-4.5 on vulnerability detection. Concepts from just four vulnerability types generalize to datasets covering 134 CWEs.