LLM Critics Help Catch LLM Bugs
FreeLLM Critics: AI trained to catch LLM bugs
FreeFree tier
Inputs: code
About LLM Critics Help Catch LLM Bugs
LLM Critics are models trained using reinforcement learning from human feedback (RLHF) to write natural language critiques that help humans evaluate model-written code more accurately. In evaluations, model-written critiques are preferred over human critiques in 63% of cases, and LLM Critics catch more bugs than human contractors paid for code review. The critics also successfully identify hundreds of errors in ChatGPT training data rated as 'flawless', even for non-code tasks. However, they can hallucinate bugs, potentially misleading human evaluators. Combining critics with human contractors reduces hallucinations while maintaining bug detection performance.
Key Features
Trained with reinforcement learning from human feedback (RLHF) to produce natural language critiques
Model-written critiques preferred over human critiques in 63% of cases
Detects more real-world bugs in code than human contractors
Capable of identifying errors in non-code tasks (out-of-distribution)
Can be combined with human contractors to reduce hallucinations
Pros & Cons
Pros
- Outperforms human contractors in bug detection rate
- Works on both code and non-code tasks
- Can improve the quality of RLHF training data
- Reduces reliance on expensive human coding experts
Cons
- May hallucinate bugs that could mislead human evaluators
- Primarily designed for code evaluation; performance on other domains may vary
- Requires careful integration with human oversight to mitigate hallucination risks
Best For
Improving code review accuracy for AI-generated codeEnhancing human evaluation ability in RLHF pipelinesCatching bugs in LLM training data and assistant outputsReducing hallucinations when used in human-machine teams
FAQ
What are LLM Critics?
LLM Critics are language models trained with reinforcement learning from human feedback to write natural language feedback highlighting problems in code from real-world assistant tasks.
How do LLM Critics compare to human evaluators?
Model-written critiques are preferred over human critiques in 63% of cases, and LLM Critics catch more bugs than human contractors paid for code review.
Can LLM Critics handle non-code tasks?
Yes, they successfully identify hundreds of errors in ChatGPT training data rated as 'flawless', even though the majority of those tasks are non-code tasks, demonstrating out-of-distribution generalization.
What are the limitations of LLM Critics?
They can hallucinate bugs that could mislead humans into making mistakes they might have otherwise avoided. Combining critics with human contractors reduces hallucinations while catching similar numbers of bugs.