Multi-Agent Reasoning with Consistency Verification Improves Uncertainty Calibration in Medical MCQA: A Deep Dive

Forum topic · 小凯 · 2026-03-26

Summary

A detailed explainer of a paper (arXiv:2603.24481) proposing a multi-agent framework that improves uncertainty calibration of large language models in medical multiple-choice QA. Four specialist agents (pulmonology, cardiology, neurology, gastroenterology) built on Qwen2.5-7B-Instruct independently analyze each question, generate chains of reasoning, and undergo two-phase self-verification to produce an S-Score reflecting internal consistency of their reasoning. Answers are fused via S-Score-weighted aggregation rather than simple majority voting. On MedQA-USMLE and MedMCQA, the method reduces Expected Calibration Error (ECE) by 49-74% (e.g., 0.356 to 0.091 on MedQA-250) while modestly improving accuracy (58.4% to 62.3%) and selective-prediction AUROC (0.574 to 0.630). Ablations show two-phase verification drives most of the calibration gains, while multi-agent diversity drives accuracy gains. The article explains why calibration matters for safe human-AI collaboration in medicine, using accessible analogies such as weather forecasting and clinical consultations.

Key points

Problem addressed: Medical LLMs are often miscalibrated—overconfident even when wrong. This undermines trust and makes it impossible for clinicians to know when to rely on AI versus intervene themselves. Calibration (alignment between stated confidence and actual accuracy) can matter more than raw accuracy in high-stakes settings.
Proposed approach (arXiv:2603.24481, March 2026):
Four specialist agents (pulmonology, cardiology, neurology, gastroenterology), each a Qwen2.5-7B-Instruct model with a different system prompt, mimic a clinical consultation panel.
Each agent generates a chain of reasoning, then acts as a verifier of its own reasoning (two-phase verification), checking logical soundness and completeness.
Reasoning chains are regenerated under variations; agreement across runs (self-consistency) yields an S-Score (Specialist Confidence Score).
Final answers use S-Score-weighted fusion: Score(option) = Σ (S-Score of agents that chose this option); the final system confidence is proportional to the winning option's score share. This can flag cases where a weak majority consensus is opposed by a highly confident dissenter.

Results

| Benchmark | Single-agent ECE | Multi-agent ECE | ECE reduction | |-----------|------------------|-----------------|---------------| | MedQA-100 | 0.356 | 0.153 | 57.0% | | MedQA-250 | 0.356 | 0.091 | 74.4% | | MedMCQA-100 | 0.321 | 0.163 | 49.2% | | MedMCQA-250 | 0.321 | 0.149 | 53.6% |

Accuracy improves from 58.4% (single-agent baseline) to 62.3% (full system).
Selective-prediction AUROC (identifying cases needing human review) rises from 0.574 to 0.630.
Ablations: multi-agent setup alone improves calibration modestly (ECE 0.356 → 0.245); two-phase verification is the main calibration driver (→ 0.091); S-Score weighting adds further accuracy gains.
Calibration improvements persist on the high-disagreement subset of hard cases, meaning the system honestly reports uncertainty exactly where human intervention matters most.

Why it matters

The framework teaches AI to "know what it does not know" (echoing Confucius's maxim cited in the original post). In medicine, a confidently wrong AI is more dangerous than one that admits uncertainty. By simulating consultation dynamics and self-verification, the system provides a usable deferral signal for clinicians. Suggested future directions include more specialties (radiology, pathology, genetics), adversarial/cross verification, and tighter human-in-the-loop collaboration.

References cited in the source

Martinez, J. R. B. (2026). *Multi-Agent Reasoning with Consistency Verification Improves Uncertainty Calibration in Medical MCQA*. arXiv:2603.24481.
Guo, C., et al. (2017). On Calibration of Modern Neural Networks. ICML.
Wang, X., et al. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR.
Jin, D., et al. (2021). MedQA: A Large-Scale Open Domain QA Dataset from Medical Exams. ACL.
Model: Qwen2.5-7B-Instruct (Alibaba Cloud, 2024). Benchmarks: MedQA-USMLE, MedMCQA.