CriticGPT: train critics to catch LLM bugs
OpenAI’s LLM Critics Help Catch LLM Bugs—specialized critique models that find errors in LLM-generated code more reliably.
Source: LLM Critics Help Catch LLM Bugs
CriticGPT (the paper’s public title is “LLM Critics Help Catch LLM Bugs”) trains critique models to find flaws in LLM-written code. The meta-point: generation and criticism are different skills, and you should train for criticism explicitly.
This work sits behind a lot of later “dedicated reviewer” thinking—including industrial verifier posts that cite CriticGPT as an ancestor not yet ready for repo-scale deployment.
Why it matters for code review
Do not assume your best coder model is your best reviewer. CriticGPT is the clean research argument for a specialized critique objective—and for measuring review on its own terms (bugs caught, false alarms) rather than on coding Elo alone.