⚖️ Awesome LLM Judges ⚖️

This repo curates recent research on LLM Judges for automated evaluation.

[!TIP]
⚖️ Check out Verdict — our in-house library for hassle-free implementations of the papers below!

📚 Table of Contents

• 🌱 Starter
• 🎭 Ensemble
  • 🤔 Debate
• 🎯 Finetuned Models
  • 🌀 Hallucination
  • 🏆 Generative Reward Models
• 🛡️ Safety
  • 🛑 Content Moderation
  • 🔍 Scalable Oversight
• 👨‍⚖️ Evaluating Judges
  • ⚖️ Biases
• 🤖 Agents
• ✨ Contributing

🌱 Starter

• Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
• G-EVAL: NLG Evaluation using GPT-4 with Better Human Alignment
• Benchmarking Foundation Models with Language-Model-as-an-Examiner

🎭 Multi-Judge

• Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models

🤔 Debate

• ScaleEval: Scalable Meta-Evaluation of LLMs as Evaluators via Agent Debate
• ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
• Can ChatGPT Defend its Belief in Truth? Evaluating LLM Reasoning via Debate
• Debating with More Persuasive LLMs Leads to More Truthful Answers

🎯 Finetuned Models

• Prometheus: Inducing Fine-grained Evaluation Capability in Language Models
• Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models
• JudgeLM: Fine-tuned Large Language Models are Scalable Judges

🌀 Hallucination

• HALU-J: Critique-Based Hallucination Judge
• MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents
• Lynx: An Open Source Hallucination Evaluation Model

🏆 Generative Reward Models

• Generative Verifiers: Reward Modeling as Next-Token Prediction
• Critique-out-Loud Reward Models

🛡️ Safety

🛑 Content Moderation

• A STRONGREJECT for Empty Jailbreaks (Sections C.4 & C.5)
• OR-Bench: An Over-Refusal Benchmark for Large Language Models (Sections A.3 & A.11)
• WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs
• Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations

🔍 Scalable Oversight

• On Scalable Oversight with Weak LLMs Judging Strong LLMs
• Debate Helps Supervise Unreliable Experts
• Great Models Think Alike and this Undermines AI Oversight
• LLM Critics Help Catch LLM Bugs

👨‍⚖️ Judging the Judges: Meta-Evaluation

• JudgeBench: A Benchmark for Evaluating LLM-based Judges
• RewardBench: Evaluating Reward Models for Language Modeling
• Evaluating Large Language Models at Evaluating Instruction Following
• Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge
• From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge
• Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences
• Re-evaluating Automatic LLM System Ranking for Alignment with Human Preference
• The Alternative Annotator Test for LLM-as-a-Judge: How to Statistically Justify Replacing Human Annotators with LLMs
• ReIFE: Re-evaluating Instruction-Following Evaluation

⚖️ Biases

• Large Language Models are not Fair Evaluators
• Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions
• Large Language Models are Inconsistent and Biased Evaluators
• Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges

🤖 Agents

🚧 Coming Soon -- Stay tuned!

✨ Contributing

Have a paper to add? Found a mistake? 🧐

• Open a pull request or submit an issue! Contributions are welcome. 🙌
• Questions? Reach out to [email protected].