Localized Multi-Agent Debate Boosts Reasoning Accuracy in AI
A novel protocol for inference, termed Localized Multi-Agent Debate (LMAD), has been unveiled to enhance the precision of multi-agent question-answering systems by concentrating discussions on the initial conflicting segments of reasoning traces. This approach models agent traces as typed nodes, pinpointing the first disagreement and confining debates to these localized areas. Additionally, a guarded resolution mechanism allows for the management of subsequent conflicts without revisiting previously accepted steps. In tests across four multi-hop question-answering benchmarks with ten backbones from four model families, LMAD demonstrated the highest macro-averaged judge accuracy, surpassing the top conventional baseline by as much as 7.20 percentage points. This research contributes significantly to artificial intelligence, particularly in refining the reasoning abilities of large language models.
Key facts
- Localized Multi-Agent Debate (LMAD) is an inference-time protocol for multi-agent debate.
- LMAD represents agent traces as typed nodes and locates the earliest conflict.
- Debate is restricted to local segments corresponding to the earliest conflict.
- Guarded resolution extends a shared committed state to address later conflicts.
- Evaluated on four multi-hop question-answering benchmarks.
- Used ten backbones from four model families.
- Achieved highest macro-averaged judge accuracy across all ten backbones.
- Outperformed strongest conventional baseline by up to 7.20 percentage points.
Entities
Institutions
- arXiv