Skip to content
conftrace_
2026 EMNLP EMNLP 2026

MEDitor: A Plug-and-Play Medical Multi-Agent Routing Protocol

AbstractReliable medical reasoning with large language models (LLMs) is hindered by a practical trade-off between accuracy and deployment cost: larger models often perform strongly but are expensive to host, while compact open-source medical LLMs can fail on difficult cases. We propose MEDitor, a training-free, plug-and-play multi-agent routing protocol that improves accuracy and traceability by selectively escalating disagreement cases. MEDitor first queries two lightweight medical experts and terminates early upon agreement. For disagreements, it escalates to an evidence-grounded coordinator that constructs option-aligned evidence, followed by a candidate-constrained judge that selects between the competing candidates or abstains. A deterministic fallback guarantees reproducibility. Across five English-language medical multiple-choice QA (MCQA) benchmarks, MEDitor achieves 65.94% average accuracy, competitive with representative ∼70B open-source models on the same evaluation suite. These results show that lightweight expert orchestration can provide a parameter-efficient and auditable route to competitive medical MCQA accuracy by reserving evidence-grounded arbitration for contested cases.

More from EMNLP 2026

All EMNLP 2026 papers →