IF-MAP: Influence-Function-Inspired Agent Prioritization in Multi-Agent Systems
Optimizing IF-MAP’s selected agents raised performance by 3.81 percentage points on average, versus 1.85 with uniform agent selection.
Across six multi-agent workflows, the method ranks agents by their potential to improve the whole system, rather than their responsibility for failures. It probes controlled prompt changes without requiring model parameters or gradients. Its priorities also guided model upgrades, yielding average gains of 7.62–16.42 percentage points across five target models, versus 3.81–8.62 with uniform selection. Including probing overhead, selective upgrades used 63.6–64.9% fewer target-model tokens than upgrading every agent.
HF Daily Papers' note
Across six multi-agent workflows, the method ranks agents by their potential to improve the whole system, rather than their responsibility for failures. It probes controlled prompt changes without requiring model parameters or gradients. Its priorities also guided model upgrades, yielding average gains of 7.62–16.42 percentage points across five target models, versus 3.81–8.62 with uniform selection. Including probing overhead, selective upgrades used 63.6–64.9% fewer target-model tokens than upgrading every agent.
HF Daily Papers' note
score 4