Reinforcement learning from human feedback (RLHF) is fundamentally limited by the capacity of humans to correctly evaluate model output. To improve human evaluation ability and overcome that limitation this work trains "critic" models that help humans to more accurately evaluate model-written code. These critics are themselves LLMs trained with RLHF to write natural language feedback highlighting problems in code from real-world assistant tasks. On code containing naturally occurring LLM errors model-written critiques are preferred over human critiques in 63% of cases, and human evaluation finds that models catch more bugs than human contractors paid for code review. We further confirm that our fine-tuned LLM critics can successfully identify hundreds of errors in ChatGPT training data rated as "flawless", even though the majority of those tasks are non-code tasks and thus out-of-distribution for the critic model. Critics can have limitations of their own, including hallucinated bugs that could mislead humans into making mistakes they might have otherwise avoided, but human-machine teams of critics and contractors catch similar numbers of bugs to LLM critics while hallucinating less than LLMs alone.

Recently, there has been considerable attention towards leveraging large language models (LLMs) to enhance decision-making processes. However, aligning the natural language text instructions generated by LLMs with the vectorized operations required for execution presents a significant challenge, often necessitating task-specific details. To circumvent the need for such task-specific granularity, inspired by preference-based policy learning approaches, we investigate the utilization of multimodal LLMs to provide automated preference feedback solely from image inputs to guide decision-making. In this study, we train a multimodal LLM, termed CriticGPT, capable of understanding trajectory videos in robot manipulation tasks, serving as a critic to offer analysis and preference feedback. Subsequently, we validate the effectiveness of preference labels generated by CriticGPT from a reward modeling perspective. Experimental evaluation of the algorithm's preference accuracy demonstrates its effective generalization ability to new tasks. Furthermore, performance on Meta-World tasks reveals that CriticGPT's reward model efficiently guides policy learning, surpassing rewards based on state-of-the-art pre-trained representation models.