MedMMV: Verifiable Multimodal Clinical Reasoning

• MedMMV is a controllable multimodal multi-agent framework that delivers reliable clinical reasoning by integrating clinical notes, images, and specialized query processing.
• The framework decomposes reasoning into stages using specialist agents, a structured evidence graph, and a hallucination-supervised repair loop to ensure verifiability.
• MedMMV demonstrates improved accuracy and truthfulness on medical benchmarks over traditional chain-of-thought methods, bolstering confidence in clinical decision support.

Searching arXiv for the exact MedMMV paper and closely related medical multimodal reasoning benchmarks/frameworks. MedMMV is a controllable multimodal multi-agent framework for reliable and verifiable clinical reasoning that was introduced to address a specific failure mode in medical multimodal LLMs: instability in early evidence interpretation can create branching reasoning trajectories that cascade into globally inconsistent conclusions (Liu et al., 29 Sep 2025). The framework is designed for multimodal medical question answering and clinical decision support, taking clinical notes $\mathcal T$, medical images $\mathcal I$, and a diagnostic or management question, and producing not only an answer but also a selected reasoning path and an evidence subgraph supporting that path (Liu et al., 29 Sep 2025). Its core design combines diversified short rollouts, a structured evidence graph built by specialist agents, supervision by a Hallucination Detector, and final path selection with a Combined Uncertainty scorer, with the stated goal of constraining stochasticity and hallucination while yielding auditable decision flows (Liu et al., 29 Sep 2025).

1. Conceptual basis and motivating failure mode

MedMMV is grounded in the claim that medical multimodal reasoning systems often fail not simply by producing an incorrect final answer, but by becoming unstable at the earliest stages of evidence interpretation (Liu et al., 29 Sep 2025). The paper argues that small stochastic differences in early reasoning can push a model into different branches across repeated runs, and that these unstable branches are more likely to accumulate hallucinated or unsupported claims (Liu et al., 29 Sep 2025). In this framing, instability and hallucination are linked process failures rather than separate problems.

To study this, the paper introduces two audit quantities. The first is the random guess measure, or RGM, which measures dispersion across repeated generations for the same multiple-choice question. For question $q$ with options $\mathcal O_q$, and sampled choices $c_{q,i}$, the empirical option prior before round $i$ is defined as

$$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$

and the question-level instability is

$$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$

Higher RGM indicates greater early instability (Liu et al., 29 Sep 2025).

The second quantity is the cross-modal hallucination rate, or CMHR. Each generation is rated by a panel of GPT-4o, Claude-Sonnet-4, and Gemini on truthfulness $T_{i,m}$, informativeness $I_{i,m}$, and image-text consistency $\mathcal I$0 when an image is present. With judge-averaged values $\mathcal I$1, $\mathcal I$2, and $\mathcal I$3, reliability is defined as

$\mathcal I$4

and cross-modal hallucination rate is

$\mathcal I$5

This makes reliability a product of truthfulness, informativeness, and multimodal consistency (Liu et al., 29 Sep 2025).

The pilot audit reported in the paper uses 100 cases with 15 generations per case and finds that high prior entropy predicts higher next-step CMHR, more branch switching correlates with higher mean CMHR, and a majority-path switch increases immediate hallucination risk from about $\mathcal I$6 to about $\mathcal I$7, a $\mathcal I$8 jump (Liu et al., 29 Sep 2025). A mediation model,

$\mathcal I$9

is used to argue that early instability affects later hallucination through switching behavior, with the reported indirect effect $q$0, Sobel $q$1, and $q$2 (Liu et al., 29 Sep 2025). This suggests that MedMMV is best understood as a process-control framework for medical multimodal reasoning rather than only an answer-improvement method.

2. System architecture and multi-agent organization

MedMMV is described as a controllable multimodal multi-agent framework because it decomposes clinical reasoning into several interacting modules with distinct responsibilities (Liu et al., 29 Sep 2025). Its inputs are clinical notes $q$3, medical images $q$4 when available, and a query; its outputs are a final answer $q$5, a selected reasoning path $q$6, and an evidence subgraph $q$7 (Liu et al., 29 Sep 2025). The framework’s overall output is therefore

$q$8

The architecture has three main stages. First, it generates diversified short preliminary rollouts. Second, it performs evidence-grounded path expansion and refinement using a structured evidence graph and a hallucination-supervised repair loop. Third, it ranks surviving paths using a Combined Uncertainty score and extracts the final answer from the highest-scoring path (Liu et al., 29 Sep 2025).

The evidence graph is built by three specialist agents. TextDoctor parses $q$9 to extract structured facts such as symptoms, history, examination findings, labs, and body-system information. ImageDoctor parses $\mathcal O_q$0 into objective imaging findings and is explicitly instructed to avoid diagnosis, instead focusing on organ or region, distribution, count, size, density, margins, laterality, and associated features. WebSearch retrieves clinically established relations among extracted findings from external medical sources (Liu et al., 29 Sep 2025). These outputs are combined into a structured evidence graph $\mathcal O_q$1, whose nodes represent atomic facts and whose edges encode relations and provenance. The paper does not provide a full graph schema, but the graph functions as the framework’s factual substrate (Liu et al., 29 Sep 2025).

The full system therefore includes a base MLLM rollout generator, TextDoctor, ImageDoctor, WebSearch, an evidence-graph builder, a FactCheck module, a Hallucination Detector or HD Supervisor, an AutoRepair module, a Combined Uncertainty scorer, and a diagnosis extractor (Liu et al., 29 Sep 2025). This division of labor is central to the paper’s claim of verifiability: reasoning is not treated as one monolithic generation, but as a multi-step process where intermediate outputs can be checked against explicit structured evidence.

3. Reasoning process: rollouts, evidence grounding, and hallucination supervision

The first stage of MedMMV generates $\mathcal O_q$2 independent short preliminary rollouts

$\mathcal O_q$3

from a base multimodal LLM (Liu et al., 29 Sep 2025). The paper’s rationale is that early evidence interpretation is where instability originates, so the system should sample multiple plausible directions before committing to a single long reasoning chain (Liu et al., 29 Sep 2025). These rollouts are deliberately short and diversified, although the paper does not define an explicit diversity objective or specify $\mathcal O_q$4 in the main text.

Each preliminary path is then refined in parallel against the evidence graph $\mathcal O_q$5. Claims made in a path are checked against the graph, unsupported or contradictory statements are flagged, and the path is revised. This loop continues until convergence or the revision budget is exhausted. The implementation allows up to three self-revision loops per trajectory (Liu et al., 29 Sep 2025). The paper describes the refined-path output as

$\mathcal O_q$6

where $\mathcal O_q$7 is the evidence subgraph supporting that path (Liu et al., 29 Sep 2025).

The Hallucination Detector, also called the Hallucination and Consistency Detector or HD Supervisor, is central to this stage. It receives the original clinical context, the current reasoning path or module output, and fact-check results against the evidence graph, and it identifies fabricated information, unsupported claims, and contradictions (Liu et al., 29 Sep 2025). Its prompt defines an explicit recommendation space that includes ACCEPT AS IS, ACCEPT WITH CAUTION, REVISE REDUCE FABRICATION, REVISE STRENGTHEN EVIDENCE, REVISE REMOVE CONTRADICTIONS, REVISE COMPREHENSIVE, REJECT HIGH HALLUCINATION, and REJECT UNRELIABLE (Liu et al., 29 Sep 2025). A representative output includes a hallucination flag, confidence, recommendation, severity, and improvement instructions.

Operationally, the Hallucination Detector sits inside a repair loop: $\mathcal O_q$8 If feedback is empty, the path is accepted; otherwise, AutoRepair revises the path and the cycle repeats (Liu et al., 29 Sep 2025). The paper does not define a learned detection objective for this component. Its role is supervisory and prompt-based rather than being a separately optimized detector model (Liu et al., 29 Sep 2025).

This stage is what makes MedMMV “verifiable” in the paper’s sense. Each candidate reasoning path is grounded in a specific evidence subgraph, and unsupported reasoning is not merely penalized in aggregate but explicitly revised or rejected before final answer extraction (Liu et al., 29 Sep 2025).

4. Combined Uncertainty scoring and final answer selection

After refinement, MedMMV ranks the surviving paths using a Combined Uncertainty, or CU, score. For a refined path $\mathcal O_q$9, the score is

$c_{q,i}$0

The appendix states that all three weights are set to 1 in experiments, so in practice the score is the sum of evidence alignment and coherence minus repair burden (Liu et al., 29 Sep 2025).

The three components are defined qualitatively. Evidence alignment $c_{q,i}$1 is the proportion of claims in the path that are verified against the supporting evidence graph. Reasoning coherence $c_{q,i}$2 is an MLLM-evaluator score for logical flow and internal consistency. Repair penalty $c_{q,i}$3 is the number of repair cycles required in Stage 2 (Liu et al., 29 Sep 2025). The CU score therefore prefers paths that are well supported, logically coherent, and cheap to repair.

Path selection is performed by

$c_{q,i}$4

and the final answer is extracted as

$c_{q,i}$5

This makes final answer generation a consequence of path selection rather than a fresh unconstrained generation step (Liu et al., 29 Sep 2025).

The appendix reports representative GPT-5 statistics for the CU components: mean $c_{q,i}$6, mean $c_{q,i}$7, mean $c_{q,i}$8, and mean CU $c_{q,i}$9 (Liu et al., 29 Sep 2025). The paper refers to this as uncertainty-aware selection, although the score is a composite heuristic rather than a Bayesian uncertainty estimate.

5. Empirical performance and physician evaluation

MedMMV is evaluated on six public medical benchmarks: multimodal VQA datasets MedXpertQA-MM, MedFrameQA, and PathVQA, and text-only QA datasets MedXpertQA-Text, MedMCQA, and MedQA (Liu et al., 29 Sep 2025). The experimental comparison includes direct chain-of-thought baselines with GPT-5, GPT-4o, GPT-oss-120B, Claude-Sonnet-4, Qwen2.5-VL-7B, and Qwen2.5-VL-72B, as well as agent baselines MDAgents, ReConcile, ColaCare, and MedAgent, instantiated with GPT-5 for fairness (Liu et al., 29 Sep 2025).

The paper’s headline result is that MedMMV improves accuracy by up to $i$0, specifically on MedFrameQA with GPT-5, where accuracy rises from $i$1 to $i$2 (Liu et al., 29 Sep 2025). Other GPT-5 backbone improvements include:

• MedXpert-MM: $i$3
• PathVQA: $i$4
• MedXpert-Text: $i$5
• MedQA: $i$6
• MedMCQA: $i$7 (Liu et al., 29 Sep 2025)

The paper places particular emphasis on reliability metrics beyond accuracy. It reports TRUE, INFO, and a normalized product

$i$8

For GPT-5 on MedXpert-MM, MedMMV changes the scores from $i$9 under direct CoT to $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$0, a $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$1 improvement in joint quality without reducing informativeness substantially (Liu et al., 29 Sep 2025). On MedXpert-Text, the corresponding shift is $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$2 in $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$3 (Liu et al., 29 Sep 2025). The paper characterizes this pattern as improved truthfulness without sacrificing information content.

A blind physician evaluation is used to support this claim. The study involves 27 licensed physicians across nine medical categories, with 2–3 physicians rating each response in a head-to-head comparison between MedMMV and CoT (Liu et al., 29 Sep 2025). On MedXpert-MM, physician-rated averages are:

• MedMMV: $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$4
• CoT: $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$5

This indicates a $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$6 gain in truthfulness and a $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$7 gain in joint quality, even though informativeness is slightly lower for MedMMV in this particular human study (Liu et al., 29 Sep 2025). The paper interprets this as evidence that MedMMV produces more clinically truthful reasoning rather than merely longer or more detailed responses.

6. Ablations, limitations, and relation to adjacent medical multimodal systems

The ablation study identifies the Combined Uncertainty scorer and the hallucination-supervised repair loop as the most important components. Replacing CU scoring with random path selection causes the largest drop, including about $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$8 accuracy loss on MedXpert-MM and about $$p_{q,i-1}(o) = \frac{1}{i-1}\sum_{t=1}^{i-1}\mathbf{1}[c_{q,t}=o], \quad o\in\mathcal{O}_q,$$9 reduction in $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$0 (Liu et al., 29 Sep 2025). Removing the self-feedback hallucination detector yields about $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$1 accuracy loss on multimodal data and about $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$2 on text-only data, while removing path expansion yields about $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$3 loss on multimodal and about $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$4 on text (Liu et al., 29 Sep 2025). Specialist agents also matter, but the paper presents them as less critical than control-and-selection mechanisms.

The framework is computationally heavier than direct CoT. The paper reports that GPT-5 CoT costs about $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$5 per item (Liu et al., 29 Sep 2025). Average output tokens rise from 1,636 to 9,144, and search cost is estimated at about $$\mathrm{RGM}_q=\frac{1}{k_q-1}\sum_{i=2}^{k_q} H_2\!\big(\mathbf p_{q,i-1}\big), \qquad \mathrm{RGM}^{\text{early}}_q=\frac{1}{L}\sum_{i=2}^{L+1} H_2\!\big(\mathbf p_{q,i-1}\big).$$6 extra per item (Liu et al., 29 Sep 2025). The authors present this as a practical tradeoff for high-stakes applications, but also acknowledge that it likely limits use in time-critical settings.

Several limitations are explicit. The framework depends on the quality of the evidence graph built by TextDoctor, ImageDoctor, and WebSearch; if those components err, the graph can propagate mistakes (Liu et al., 29 Sep 2025). Benchmark evaluation uses static public datasets rather than fully interactive or longitudinal clinical workflows (Liu et al., 29 Sep 2025). The paper also leaves several elements only partially formalized, including the full graph schema, a learned hallucination-detection objective, and a more explicit convergence criterion for refinement (Liu et al., 29 Sep 2025). Minor reporting inconsistencies are also present, particularly in the appendix’s backbone description for human evaluation and in the ablation table (Liu et al., 29 Sep 2025).

In the broader medical multimodal literature, MedMMV belongs to a line of work that shifts evaluation and model design away from static perception tasks toward process-level clinical reasoning. A closely related benchmark development is 3MDBench, which evaluates multimodal dialogue-based telemedicine consultations rather than single-turn VQA (Sviridov et al., 26 Mar 2025). Another relevant direction is MMedPO, which addresses factuality problems caused by modality misalignment in medical LVLMs by constructing clinically meaningful preference pairs and weighting them by clinical relevance (Zhu et al., 2024). MedMMV differs from both by focusing on instability-controlled reasoning trajectories, evidence graphs, and auditable multi-agent deliberation, but these papers collectively suggest a broader movement toward clinically grounded multimodal reasoning rather than isolated medical image answering (Liu et al., 29 Sep 2025).

Taken together, MedMMV is best understood as methodological infrastructure for trustworthy multimodal clinical reasoning. Its central claim is that reliability in medical AI depends not only on answer accuracy but on controlling early branching instability, grounding intermediate steps in structured evidence, and selecting final reasoning paths through an auditable and verifiable process (Liu et al., 29 Sep 2025).