Research map / Multi-agent systems and games / LLM agents

MedLA: A Logic-Driven Multi-Agent Framework for Complex Medical Reasoning with Large Language Models

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MedLA, a multi-agent framework where each agent builds syllogism-based logical trees for medical MCQs and revises them through multi-round, graph-guided peer discussion, evaluated on MedDDx, MedRAG-QA and MedXpertQA.

Problem

Multi-agent medical reasoning systems rely on fixed roles or positional debate and cannot localize fine-grained logical or premise-level conflicts in complex clinical questions.

Approach

Represent each agent's reasoning as an explicit syllogism-based logical tree (major premise, minor premise, conclusion), with a credibility agent flagging weak nodes, and iteratively compare and revise these trees across agents without fine-tuning or retrieval.

Key results

• On MedDDx, MedLA+LLaMA3.1(8B) reaches 44.3 average accuracy versus 36.9 for LLaMA3.1-8B and 37.7 for MDAgents (+7.4 over the base model per the paper's table).
• On MedRAG-QA multi-choice tasks, MedLA+LLaMA3.1(8B) averages 69.9 versus 64.2 for LLaMA3.1-8B and 60.8 for MDAgents.
• On MedXpertQA with DeepSeek-R1, MedLA scores 36.0 versus 21.3 for the baseline (60-question random sample).

Surprising findings

• Gains from MedLA grow with difficulty tier on MedDDx (+4.6 basic, +6.4 intermediate, +11.1 expert points).
• Majority voting with the same backbone already gives sizable gains, but the logic-tree scaffold adds further improvement over CoT-only outputs.

Limitations

• Logic trees are produced as text blocks via prompting since LLMs cannot emit structured trees directly, so structure is not guaranteed.
• Substantially higher latency than single-pass baselines (17 subagent steps).
• Reported baseline numbers are taken from prior work (MIRAGE/MedDDx) rather than re-run, and the MedXpertQA evaluation uses only 60 sampled questions.

Experimental setup

Datasets: MedDDx MMLU-Med MedQA-US BioASQ-Y/N MedXpertQA STaRK-Prime

Benchmarks: MedDDx MMLU-Med MedQA-US BioASQ-Y/N MedXpertQA

Metrics: accuracy standard deviation over three runs

Baselines: QAGNN JointLK Dragon Majority Voting DyLAN MedAgents MDAgents Mistral-7B MedAlpaca-7B PMC-LLaMA-7B LLaMA3-8B LLaMA3.1-8B

Builds on: chain-of-thought MedAgents MDAgents DyLAN MedRAG

Closest papers by meaning

• MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.88)
• MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.88)
• KGARevion: An AI Agent for Knowledge-Intensive Biomedical QA (ICLR 2025 · similarity 0.87)
• Reinforcement Learning for Tool-Calling Agents in Fast Healthcare Interoperability Resources (FHIR) (ICML 2026 · similarity 0.87)
• Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning (CVPR 2026 · similarity 0.86)
• MedChain: Bridging the Gap Between LLM Agents and Clinical Practice with Interactive Sequence (NeurIPS 2025 · similarity 0.86)