Challenge-Aware Knowledge Recruitment with Meta-Agent Supervision for Medical Multi-Agent Reasoning
Summary
- Existing multi-agent systems often select experts and retrieve knowledge based solely on initial questions rather than evolving reasoning gaps.
- Researchers propose challenge-aware adaptive recruitment, where an external challenger inspects reasoning traces to identify factual or expertise gaps.
- On the MEDAGENTSBENCH hard subset, challenge-aware in-between retrieval improves accuracy across 8 of 9 datasets compared to upfront retrieval.
- Challenger-triggered specialist recruitment raises GPT-4.1-mini MedQA accuracy from 46% to 54% while showing positive generalization.
- Naive combinations of retrieval and specialist recruitment degrade performance, demonstrating the need for coordinated intervention channels.
Why it matters
The study highlights that the timing and motivation of external resource recruitment are crucial architectural design choices for improving multi-agent medical reasoning systems.