Megadose The AI curve, compressed.

More Agents, More Samples, or More Thinking? Untangling Coordination, Ensembling, and Test-Time Compute in Multi-Agent LLMs

· HF Daily Papers ·
Multi-agent systems averaged 4.2 percentage points lower accuracy than baselines matched on sample count or resource budget.

The paper synthesizes nine empirical evaluations covering 516 unique multi-agent configurations. In 74% of comparable contrasts, multi-agent accuracy equaled or fell below the baseline. It argues that comparisons against a single greedy pass confound coordination with extra samples and inference compute. The author concludes that positive returns require partitioned evidence, external verification, or error diversity across models when communication itself adds no information.

HF Daily Papers' note

score 5

Categories: Research