DebUnc: Improving Large Language Model Agent Communication With Uncertainty Metrics
This addresses the issue of inaccurate communication in multi-agent systems for AI researchers, though it is incremental as it builds on existing debate frameworks.
The paper tackles the problem of misleading confident-sounding responses in multi-agent LLM debates by proposing DebUnc, a framework that uses uncertainty metrics to assess agent confidence, resulting in improved performance across benchmarks as uncertainty estimation becomes more reliable.
Multi-agent debates have been introduced to improve the accuracy of Large Language Models (LLMs) by having multiple agents discuss solutions to a problem over several rounds of debate. However, models often generate incorrect yet confident-sounding responses, which can mislead others. This issue arises partly because agents do not consider how confident their peers are. To address this, we propose DebUnc, a debate framework that uses uncertainty metrics to assess agent confidence. Confidence is then conveyed through a modified attention mechanism that adjusts token weights, or through textual prompts. Evaluations across benchmarks show that attention-based methods are particularly effective and that performance continues to improve as uncertainty estimation becomes more reliable. The code is available at https://github.com/lukeyoffe/debunc.