NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
The Confident Liar: Diagnosing Multi-Agent Debate with Log-Probabilities and LLM-as-Judge
TL;DR
Multi-agent debate systems often neglect the quality of intermediate reasoning, focusing solely on final answers. This study introduces a framework that uses log-probability distributions and an LLM-as-judge to evaluate reasoning quality in debates.
✦ Why It Matters
Engineers can leverage confidence signals to enhance the evaluation of multi-agent debate systems and improve reasoning quality.
Key Takeaways
- Implement log-probability analysis to assess agent confidence.
- Utilize LLMs for scoring reasoning quality in debates.
- Focus on the Constructor agent for better evaluation outcomes.
- Explore confidence signals to detect reasoning failures effectively.
Full Summary
Multi-agent debate systems are designed to enhance reasoning through structured argumentation, yet they typically assess only the correctness of final answers. This research develops a framework that pairs a two-agent debate architecture—a Constructor and an Auditor—with a large language model (LLM) acting as a judge to score reasoning quality.
The study analyzes token-level log-probability distributions, which indicate the agents' confidence in their reasoning, alongside rubric scores assigned by the LLM. Experiments reveal a four-phase confidence trajectory and significant asymmetry in confidence alignment with reasoning quality, with the Constructor showing a stronger correlation (AUROC 0.804) compared to the Auditor (0.634).
These findings suggest that confidence signals can effectively identify critical reasoning failures, particularly in the Constructor agent. The implications extend to improving evaluation methods across various domains, including mathematical reasoning and factual question answering.