Reproduce “Confidence Calibration and Rationalization for LLMs via Multi-Agent Deliberation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ruixin Yang, Dheeraj Rajagopal, Shirley Anugrah Hayati, Bin Hu, Dongyeop Kang

A multi-agent deliberation framework calibrates and rationalizes Large Language Model confidence by simulating a two-stage group discussion among tool-augmented LLM agents. This approach yields superior Expected Calibration Error (ECE) on complex reasoning tasks like GSM8K and DateUnderstanding, while maintaining comparable performance on other generative QA tasks.

Key insights

  • Structured multi-agent deliberation, including peer feedback and rationalization, significantly improves LLM confidence calibration beyond simple ensemble methods.
  • Leveraging diverse 'expert' LLM agents and dynamically allocating them can contribute to better initial stances for deliberation.
  • The multi-agent deliberation process inherently supports the generation of human-readable rationales for confidence adjustments, enhancing interpretability and trustworthiness of LLM outputs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.