Multi-Agent Reasoning with Consistency Verification Improves Uncertainty Calibration in Medical MCQA¶

🕒 Published (v1): 2026-03-25 16:22 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

MARC (Multi-Agent Reasoning with Consistency Verification) is a three-component multi-agent framework that combines domain-specialist LLM agents, inference-time consistency verification, and confidence-weighted fusion to improve uncertainty calibration in medical MCQA. On high-disagreement subsets of MedQA-USMLE, the full system achieves ECE = 0.091 (−74.4% over single-specialist baseline) and AUROC = 0.630 (+0.056) without requiring any labeled calibration data. A key finding is that calibration (ECE) and discrimination (AUROC) respond to fundamentally different architectural components.

Problem¶

LLMs deployed in clinical settings are systematically overconfident: raw verbalized confidence scores (mean ≈ 0.90) diverge sharply from observed accuracy (43–57%), making confidence useless as a deferral signal. Existing recalibration methods (temperature scaling, deep ensembles, Bayesian dropout) require either labeled calibration sets or multiple model copies; prior multi-agent medical QA work improves accuracy but ignores calibration entirely.

Method¶

MARC is a three-component pipeline applied at inference time to each question:

Specialist Agents: Four domain-specific agents (respiratory, cardiology, neurology, gastroenterology) each backed by Qwen2.5-7B-Instruct with a specialty-specific system prompt and a five-step CoT protocol. Greedy decoding ($T=0$) for main calls.
Two-Phase Consistency Verification (Wu et al., 2024): For each specialist's answer $a_k$ and reasoning $r_k$, four verification questions $\{q_j\}$ are extracted from $r_k$. Each $q_j$ is answered (a) without reference to $r_k$ and (b) with $r_k$ in context. Pairwise inconsistency is measured via Jaccard similarity ($\tau = 0.4$); the inconsistency score is $I_k = \frac{1}{N_k}\sum_{j=1}^{N_k}\mathbf{1}[\text{sim}(a_j^{\text{ind}}, a_j^{\text{ref}}) < \tau]$. The Specialist Confidence Score (S-score) is: $$S_k = C_k^{(0)} \cdot (1 - I_k)$$
S-Score Weighted Fusion: For each candidate answer $a$, the aggregate score is $\text{score}(a) = |\{k: a_k = a\}| \times \bar{S}_a$ where $\bar{S}_a$ is the mean S-score of supporting specialists. Final calibrated confidence blends vote fraction $v$ with mean and minimum S-score of winning supporters: $\hat{C} = v \cdot \bar{S}_{\hat{a}} + (1-v) \cdot \min_{k: a_k=\hat{a}} S_k$.

Evaluation uses high-disagreement-only subsets of MedQA-USMLE and MedMCQA (100 and 250 questions each); standard post-hoc baselines are excluded because they require labeled calibration sets unavailable in the inference-only target setting.

Key Contributions¶

MARC framework: First system to jointly improve ECE (calibration) and AUROC (discrimination) over a single-agent LLM baseline using purely inference-time signals with no labeled calibration data.
Decomposition finding: Two-Phase Verification is the primary driver of ECE reduction; multi-agent reasoning is the primary driver of AUROC improvement — addressing orthogonal failure modes.
S-Score multiplicative formula: Shown superior to additive and pure-consistency alternatives; rewards well-calibrated specialists and maximally penalizes overconfident-inconsistent ones.
Empirical MedQA vs. MedMCQA characterization: Clinical reasoning questions (MedQA, 109-word vignettes) respond to verification-based calibration differently than factual recall questions (MedMCQA, 15-word stems) for 7B-class models.

Results¶

MedQA-250 (primary):
Full system (C4): ECE = 0.091 (−74.4%), AUROC = 0.630 (+0.056), Acc = 59.2% (+4.8 pp) vs. single-specialist baseline
Two-Phase Verification alone (C2): ECE −50%, AUROC −0.018 (worse), Acc unchanged
Multi-agent without verification (C3): ECE −5%, AUROC +0.013, Acc +2.8 pp
MedQA-100: C4 achieves ECE = 0.098 (−73%), AUROC = 0.645 (+0.108), Acc = 59.0% (+7.0 pp)
MedMCQA-250: ECE −63%, AUROC +0.058, but Acc = 44.0% (+1.2 pp) — calibration gains persist despite weak accuracy improvement constrained by factual recall demands
Calibration gains of 49–74% ECE reduction hold across all four evaluation settings
Two-Phase Verification alone produces negative AUROC on 3/4 datasets; multi-agent fusion in C4 recovers and extends discrimination
Wall-clock: 833 min for C4 on MedQA-250 vs. 116 min for C1 (single-specialist), on a single RTX 5090

Limitations¶

Evaluated only on high-disagreement filtered subsets (100–250 questions); results explicitly cannot be extrapolated to full benchmark distributions
Only 7B-parameter models tested; performance constrained by knowledge recall capacity, particularly on MedMCQA
Four specialists cover a limited set of specialties; coverage of MedMCQA's 21 specialties is incomplete
Two-Phase Verification alone degrades discrimination (AUROC) on 3/4 datasets — consistency is an imperfect proxy for correctness
Clinical utility of the resulting calibration (sufficient for deferral decisions?) is not demonstrated, acknowledged as future work
7× inference-time cost increase (C1→C4) on large sets; not addressed as a deployment constraint

Relevance to Harnesses / Meta-Harnesses¶

MARC is a textbook example of a multi-agent meta-harness: it orchestrates a fleet of specialist sub-agents, then applies a verification sub-harness to each agent's output before aggregating results via a confidence-weighted fusion controller. The Two-Phase Verification component functions as an inline quality-gate layer — analogous to the "verify before merge" pattern in agentic harness design — that converts each sub-agent's raw output into a calibrated trust signal used by the top-level controller. The ablation results are directly relevant to harness designers: they show that consistency-checking (a common harness-layer technique) and ensemble aggregation address orthogonal failure modes and should be composed rather than substituted. The inference-time, label-free nature of the approach means it can be dropped into existing multi-agent harnesses without retraining any component.