https://arxiv.org/abs/2306.05685

24th December 2023

Key Highlights

• The paper addresses the challenges in evaluating large language models (LLMs) for chat assistants, particularly for capturing human preferences that traditional benchmarks may miss.

Proposed Benchmarks

1. MT-Bench:
   • Comprises 80 high-quality multi-turn questions.
   • Evaluates a chatbot's ability in:
     • Writing
     • Roleplay
     • Information extraction
     • Reasoning
     • Math
     • Coding
     • Knowledge in STEM and humanities.
2. Chatbot Arena:
   • A crowdsourced platform for head-to-head chatbot comparisons.
   • Users pose the same question to two anonymous chatbots and vote for the preferred response.
   • Captures real-world user preferences.

Insights on LLM-as-a-Judge

• Strong LLMs (e.g., GPT-4) are proposed as judges to assess other models on open-ended questions.
• Identified biases in the LLM-as-a-judge method:
  • Position bias: Preference for the position of an answer.
  • Verbosity bias: Preference for longer responses.
  • Self-enhancement bias: Favoring its own responses.
  • Limited reasoning ability: Inaccuracies in complex judgments.

Bias Mitigation

• Solutions to reduce biases were proposed.
• Results show that strong LLM judges agree with human preferences over 80% of the time, matching human-human agreement levels.

Hybrid Evaluation Framework

• Combines:
  • Traditional capability-based benchmarks.
  • Preference-based benchmarks using LLM-as-a-judge.
• Enables:
  • Swift and automated evaluations.
  • Assessment of both technical capabilities and alignment with human preferences.

Public Resources

• Released resources include:
  • MT-Bench questions.
  • 3,000 expert votes.
  • 30,000 conversations with user preferences.
  • All datasets are made publicly available for research.

Conclusion

• The paper demonstrates that LLM-as-a-judge is a scalable, explainable, and effective method for approximating human preferences.