Definition
LLM-as-judge uses one language model to score another’s outputs, replacing slow and expensive human evaluation for many tasks. It’s indispensable at scale and has well-known biases: judges tend to prefer longer answers, their own family of models, and reasoning that looks confident.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- Self-Refine: Iterative Refinement with Self-Feedback
- Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations
- Large Language Models are not Robust Multiple Choice Selectors
- Faithful Chain-of-Thought Reasoning
- Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
- Agent-as-a-Judge: Evaluate Agents with Agents
- Self-Consistency Improves Chain of Thought Reasoning in Language Models
- ReAct: Synergizing Reasoning and Acting in Language Models
- Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges
- Self-Rewarding Language Models
- SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
- Chain-of-Verification Reduces Hallucination in Large Language Models
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
- RewardBench: Evaluating Reward Models for Language Modeling