IF-MAP ranks which agents to improve in multi-agent systems
Original title: IF-MAP: Influence-Function-Inspired Agent Prioritization in Multi-Agent Systems
Prompt optimization of selected agents gained 3.81 points versus 1.85 for uniform selection across six workflows.
IF-MAP is an influence-function-inspired method that scores which agent in an LLM multi-agent system would yield the largest end-to-end gain if improved. It measures loss under controlled prompt perturbations, fits a regularized local response model, and needs no model parameters or gradients. On six workflows spanning code generation, question answering and mathematical reasoning, optimizing selected agents’ prompts beat uniform selection. Upgrading selected agents across five target models also beat uniform upgrades while using fewer tokens than improving every agent.
Unconfirmed: Reported gains come from the authors’ evaluation and are not independently replicated here.
- Macro-average prompt-optimization gain was 3.81 points versus 1.85 for uniform selection.
- Selected-agent model upgrades gained 7.62 to 16.42 points versus 3.81 to 8.62.
- Qwen3-8B and DeepSeek-V3.2 token use fell 74.2% and 35.2% versus optimizing every agent.
- Target-model tokens fell 63.6% to 64.9% versus upgrading every agent.
From the original article
1 min read at huggingface.co
Improving every agent in an LLM-based multi-agent system can be prohibitively costly, making it important to identify which agent should be prioritized to achieve the greatest system-level gain with limited resources. Failure attribution can inform this decision, but an agent's responsibility for failures does not directly reveal the system-level gains achievable by improving it.
By Jie Mao, Pingchen Lu, Zikun Qu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.