CoChat
Sign Up
Cited by 1Artificial Intelligence in Healthcare and EducationMachine Learning in HealthcareTopic Modeling

Model confrontation and collaboration: A debate intelligence framework for enhancing medical reasoning in large language models

Open access
Sun, Xinti, Hong, Qiyang, Zhang, Mengyan, Li, Yuyan, Chen, Tingwei, Huang, Zigeng, Liang, Guihan, Tang, Wenjun +5 more·2026·Cell Reports Medicine
DOI 10.1016/j.xcrm.2025.102547PMC PMC12866169
Open this paper in CoChat — free accountOpen PDF

Abstract

Medical reasoning is fundamental to clinical decision-making, underpinning tasks such as patient communication, diagnosis, and treatment planning. Inspired by psychological findings that peer interaction promotes self-correction, we introduce model confrontation and collaboration (MCC), a debate intelligence framework that transcends static ensemble methods by integrating critique and self-reflection to iteratively refine reasoning through structured, multi-round confrontation and collaboration among diverse large language models (LLMs). In multiple-choice benchmarks, MCC achieved mean accuracy on MedQA (92.6%) and PubMedQA (84.8%) and demonstrated strong performance on medical subsets of MMLU. In long-form medical question answering, MCC outperformed all individual LLMs and the domain-specific LLM Med-PaLM 2 in both physician and layperson evaluations. In diagnostic dialog tasks, MCC further excelled in both history-taking and diagnostic accuracy, reaching a top-1 diagnosis rate of 80%. These results position MCC as a scalable, model-agnostic framework that advances medical reasoning through collaborative deliberation.

AI snapshot

Written from the paper’s full text.

TL;DR

Model confrontation and collaboration (MCC) is a debate-based multi-LLM framework that iteratively critiques and refines medical reasoning across heterogeneous models. It improved accuracy on multiple-choice benchmarks and outperformed individual LLMs and Med-PaLM 2 on long-form and diagnostic dialog tasks.

Key findings

  • On MedQA (1,273 questions), MCC achieved 92.6% ± 0.3% mean accuracy, exceeding previous SOTA Med-Gemini (91.1%) and individual GPT-o1 (84.8% ± 0.4%), Qwen-QwQ (81.8% ± 0.5%), and DeepSeek-R1 (89.7% ± 0.4%; all p < 0.001).
  • On PubMedQA (500 questions), MCC reached 84.8% ± 0.4%, above previous SOTA MDTeamGPT (83.9%) and individual LLMs; on MMLU medical subsets, MCC achieved SOTA on five of six topics, with accuracies from 91.0% to 99.7%.
  • In long-form medical QA, MCC outperformed all individual LLMs and Med-PaLM 2 across 12 physician-rated dimensions, with gains including correct retrieval 11.8% (p < 0.001), correct reasoning 8.8% (p < 0.001), and bias reduction 11.1% (p < 0.001).
  • In OSCE-style diagnostic dialogs, MCC reached top-1 primary diagnosis accuracy of 80.0% (12/15), versus 66.7% for GPT-o1 and DeepSeek-R1 and 60.0% for Qwen-QwQ; exact McNemar tests were not significant due to limited discordant pairs.

Methods

  • MCC integrates GPT-o1, Qwen-QwQ, and DeepSeek-R1 in three stages: independent initial response generation, adversarial activation when responses diverge, and consensus optimization; if no consensus within three rounds, final output is selected by majority voting.
  • Evaluation covered multiple-choice benchmarks (MedQA, PubMedQA, MMLU medical subsets), advanced/adversarial benchmarks (MedXpertQA, MetaMedQA, RABBITS), long-form QA (MultiMedQA 140, HealthBench Consensus n = 3,671 and Hard n = 1,000), and OSCE-style diagnostic dialogs using 16 simulated patient scenarios.
  • Multiple-choice results were reported as mean accuracy with 95% CIs over 10 independent runs; long-form outputs were rated by 10 physicians and 10 laypersons; diagnostic dialog analyses used exact McNemar, Wilcoxon signed-rank, Friedman, and Bonferroni-adjusted tests.

Limitations

  • The paper states three main limitations: human evaluators were recruited from a single region in China, limiting generalizability; MCC’s performance boundaries are uncertain for highly ambiguous, inconsistent, or sparse real-world clinical data, where debate may amplify uncertainty; and MCC requires increased computational cost and longer inference time than single-model baselines.
  • Evident from the text: the OSCE diagnostic evaluation used only 16 cases, and primary diagnosis comparisons were underpowered, with exact McNemar tests not reaching significance.
  • Evident from the text: long-form human evaluation used a small MultiMedQA-140 sample and manual ratings, which the authors note introduce subjectivity and potential bias.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.xcrm.2025.102547

Cites

33 of 33 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

14 of 20 cited sources checked · 5 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    These models typically rely on a single-LLM architecture, which inherently lacks mechanisms for external validation, cross-perspective critique, and self-correction.

    “However, these models typically rely on a single-LLM architecture, which inherently lacks mechanisms for external validation, cross-perspective critique, and self-correction5—key elements for ensuring reliability in complex, high-stakes medical scenarios.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [5] Luo, 2025, Learning from Peers in Reasoning Models, arXiv source

      The Introduction (section 1; page unavailable) describes limitations in reasoning models’ self-correction, and the Abstract describes peer interaction as a way to improve it. But the Introduction explicitly says these models already have self-verification and self-refinement capabilities, contrary to the claim that they inherently lack self-correction. The cited passages also do not establish that the models typically use a single-LLM architecture or lack external validation. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Premisecitation

    Current multi-LLM frameworks rely on a single model family, limiting epistemic diversity and opportunities for cross-validation.

    “However, current frameworks rely on a single model family (e.g., GPT series6), limiting epistemic diversity and opportunities for cross-validation.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [6] Chen, 2025, Enhancing diagnostic capability with multi-agents conversational large language models, npj Digit. Med. source

      The Methods (Study design; Multi-agent conversation framework) describe MAC configurations using GPT-3.5 or GPT-4 as the base model, supporting the GPT-family example. The paper does not say that reliance on one model family limits epistemic diversity or cross-validation; the Discussion instead says its multi-agent conversations explore diverse reasoning paths. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
  • Premisecitation

    Several multi-LLM frameworks (MedAgents, MDTeamGPT, MDAgents) employ role specialization to enhance accuracy.

    “Several multi-LLM frameworks, such as MedAgents,8 MDTeamGPT,9 and MDAgents,34 employ role specialization (e.g., proposer, reviewer, simulated expert, or specialty-specific expert pools) to enhance accuracy.”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [8] Tang, 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [9] Chen, 2025, MDTeamGPT: A Self-Evolving LLM-based Multi-Agent Framework for Multi-Disciplinary Team Medical Consultation, arXiv source
    • cites [34] Kim, 2025, MDAgents: an adaptive collaboration of LLMs for medical decision-making source
  • Premisecitation

    Prior studies have noted the presence of malformed and incomplete questions in the MedQA benchmark.

    “Prior studies have noted the presence of malformed and incomplete questions in the MedQA benchmark,22 consistent with our observations during evaluation.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [22] Griot, 2025, Large Language Models lack essential metacognition for reliable medical reasoning, Nat. Commun. source
  • Interpretationcitation

    Correct initial answers exhibit higher robustness to perturbation, consistent with findings that well-grounded LLM reasoning resists noise when supported by coherent internal representations.

    “Second, correct initial answers exhibit higher robustness to perturbation, consistent with findings that well-grounded LLM reasoning resists noise when supported by coherent internal representations.26,27”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [26] Jiang, 2025, Misaligning Reasoning with Answers - A Framework for Assessing LLM CoT Robustness, arXiv source
    • cites [27] Li, 2024, An Empirical Study on Large Language Models in Accuracy and Robustness under Chinese Industrial Scenarios, arXiv source

      Section 5.2 (page not provided) reports that robustness scores for initially correct answers generally exceed those for initially incorrect answers, with an exception for the spark model. The paper does not establish that coherent internal representations explain this difference or that well-grounded reasoning resists noise. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • May combine two claims. 50
  • Classified as probably not a substantial claim. 28
  • The quote only partly states this claim. 1

  • May combine two claims.

    Premisecitation

    Existing multi-LLM frameworks rely on fixed mechanisms such as consensus voting or stepwise reasoning to aggregate independent LLM outputs but lack dynamic, debate-style interaction essential for robust reasoning.

    “Prior studies rely on fixed mechanisms such as consensus voting8,9 or stepwise reasoning10 to aggregate independent LLM outputs but lack dynamic, debate-style interaction essential for robust reasoning.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [8] Tang, 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [9] Chen, 2025, MDTeamGPT: A Self-Evolving LLM-based Multi-Agent Framework for Multi-Disciplinary Team Medical Consultation, arXiv source
    • cites [10] Wang, 2025, Mixture-of-Agents Enhances Large Language Model Capabilities source

      The Introduction describes MoA as aggregating independently generated responses through successive layers of refinement, supporting the stepwise-aggregation part of the claim. But §4.2 Model Ensemble describes prior multi-agent studies in which agents discuss problems interactively, contrary to the blanket assertion that prior frameworks lack debate-style interaction. The paper also does not establish that debate is essential for robust reasoning. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LLMs tailored for healthcare have achieved state-of-the-art (SOTA) performance on a range of medical benchmarks.

    “In recent years, large language models (LLMs) tailored for healthcare have achieved state-of-the-art (SOTA) performance on a range of medical benchmarks.2,3,4”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [2] Singhal, 2025, Toward expert-level medical question answering with large language models, Nat Med source
    • cites [3] Saab, 2024, Capabilities of Gemini Models in Medicine, arXiv source
    • cites [4] Nori, 2023, Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine, arXiv source

  • May combine two claims.

    Premisecitation

    Mixture-of-Agents primarily refine immediate predecessor outputs without revisiting earlier contexts, a limitation that may restrict their applicability in dynamic clinical discussions.

    “yet approaches such as Mixture-of-Agents,10 although introducing hierarchical architectures and network-based communication, primarily refine immediate predecessor outputs without revisiting earlier contexts, a limitation that may restrict their applicability in dynamic clinical discussions.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [10] Wang, 2025, Mixture-of-Agents Enhances Large Language Model Capabilities source

      The Introduction and §2.2 describe successive layers refining responses from the preceding layer, supporting that part of the claim. But §2.2 also says each layer’s input includes the original prompt, so the paper does not support an unqualified claim that MoA never revisits earlier context. It does not assess applicability to dynamic clinical discussions. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Interpretationcitation

    Although reasoning-capable LLMs possess the capacity to self-correct, their initial responses often reflect a form of prompt-induced cognitive rigidity-known as the "Prefix Dominance Trap"-in which models fixate on superficial cues in the question and fail to activate the optimal reasoning trajectory.

    “First, although reasoning-capable LLMs possess the capacity to self-correct, their initial responses often reflect a form of prompt-induced cognitive rigidity—known as the “Prefix Dominance Trap”5—in which models fixate on superficial cues in the question and fail to activate the optimal reasoning trajectory.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [5] Luo, 2025, Learning from Peers in Reasoning Models, arXiv source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    To account for the inherent stochasticity of LLMs, results are reported from 10 independent runs per benchmark as mean accuracy with 95% confidence intervals.

    “To account for the inherent stochasticity of LLMs, we report results from 10 independent runs per benchmark as mean accuracy with 95% confidence intervals (CIs), moving beyond the single-run point estimates common in prior works.2,3,4”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [2] Singhal, 2025, Toward expert-level medical question answering with large language models, Nat Med source

      The Statistical analyses section says the paper computed binomial confidence intervals for multiple-choice accuracy, and Table 1 reports accuracy with 95% confidence intervals. It does not report mean accuracy across 10 independent runs per benchmark. Nor is it an example of reporting only single-run point estimates, as the citing sentence’s contrast suggests. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [3] Saab, 2024, Capabilities of Gemini Models in Medicine, arXiv source

      Table 1 (Section 4.1; page unavailable) reports benchmark accuracies as point estimates, supporting the description of this paper as prior work presenting point estimates. The cited paper does not establish that those estimates came from single runs. The 10-run procedure and 95% confidence intervals are the citing authors’ own method, not a method attributed to this source. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [4] Nori, 2023, Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine, arXiv source

      The citation is attached to the comparison with prior work, not to the citing authors’ own 10-run procedure. Table 1 (Results; page unavailable) reports benchmark accuracy as point estimates without confidence intervals, but the paper does not establish that those estimates came from single independent runs. Its five parallel API calls are an inference-time ensemble, not 10 benchmark runs. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    Among three debate rounds, GPT-o1 corrected 82.4% of its initial errors (from 108 to 19), Qwen-QwQ corrected 55.7% (from 149 to 66), and DeepSeek-R1 corrected 30% (from 50 to 35).

    “Among three debate rounds, GPT-o1 corrected 82.4% of its initial errors (from 108 to 19), Qwen-QwQ corrected 55.7% (from 149 to 66), and DeepSeek-R1 corrected 30% (from 50 to 35) (Figure S1).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    At the MCC initial stage, consensus was reached among three models on 1,019 questions, of which 938 (73.7%) were correctly answered and 81 (6.4%) were incorrectly answered.

    “At the MCC initial stage, consensus was reached among three models on 1,019 questions, of which 938 (73.7%) were correctly answered and 81 (6.4%) were incorrectly answered (Figure 2A).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On MedQA (1,273 questions), MCC achieved SOTA accuracy of 92.6% ± 0.3%, outperforming the previous SOTA of 91.1% by Med-Gemini and individual reasoning LLMs (GPT-o1 84.8% ± 0.4%, Qwen-QwQ 81.8% ± 0.5%, DeepSeek-R1 89.7% ± 0.4%; all p < 0.001) and Med-PaLM 2 (86.5%).

    “On the MedQA benchmark of 1,273 questions, MCC achieved SOTA accuracy of 92.6% ± 0.3%, outperforming the previous SOTA of 91.1% by Med-Gemini3 and individual reasoning LLMs (GPT-o1, 84.8% ± 0.4%; Qwen-QwQ, 81.8% ± 0.5%; and DeepSeek-R1, 89.7% ± 0.4%; all p < 0.001), as well as medical LLM Med-PaLM 22 (86.5%) (Table 1).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

  • May combine two claims.

    Findingown data

    Among all 254 non-consensus cases, DeepSeek-R1 achieved the highest initial accuracy with 204 correct responses (80.3%), followed by GPT-o1 with 146 (57.5%) and Qwen-QwQ with 105 (41.3%).

    “Among all 254 non-consensus cases, DeepSeek-R1 achieved the highest initial accuracy with 204 correct responses (80.3%), followed by GPT-o1 with 146 (57.5%) and Qwen-QwQ with 105 (41.3%) (Figure 2B).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the MMLU benchmarks, MCC achieved SOTA performance on five of six clinical topics, with accuracies from 91.0% to 99.7%, consistently outperforming Med-PaLM 2 and individual LLMs.

    “On the MMLU benchmarks, MCC achieved SOTA performance on five of six clinical topics, with accuracies from 91.0% to 99.7%, consistently outperforming Med-PaLM 2 and individual LLMs (Table 1; Data S6, S7, S8, S9, S10, and S11).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

  • May combine two claims.

    Findingown data

    On PubMedQA (500 questions), MCC achieved SOTA accuracy of 84.8% ± 0.4%, outperforming the previous SOTA of 83.9% by MDTeamGPT and individual reasoning LLMs (GPT-o1 74.5% ± 0.4%, Qwen-QwQ 81.8% ± 0.7%, DeepSeek-R1 80.9% ± 0.6%) and Med-PaLM 2 (81.8%).

    “On the PubMedQA benchmark of 500 questions, MCC achieved SOTA accuracy of 84.8% ± 0.4%, outperforming the previous SOTA of 83.9% by MDTeamGPT9 and individual reasoning LLMs (GPT-o1, 74.5% ± 0.4%; Qwen-QwQ, 81.8% ± 0.7%, and DeepSeek-R1, 80.9% ± 0.6%), as well as a medical LLM (Med-PaLM 2, 81.8%) (Table 1).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

  • May combine two claims.

    Findingown data

    In tasks like clinical knowledge, anatomy, and professional medicine, MCC resolved 100% of disagreement cases or outperformed simple majority voting by up to 30%.

    “In tasks like clinical knowledge, anatomy, and professional medicine, MCC resolved 100% of disagreement cases or outperformed simple majority voting by up to 30%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Of the 13 cases where at least one LLM initially answered correctly yet MCC reached an erroneous consensus, 4 were suppression-induced errors, in which the correct model maintained its position but failed to persuade the incorrect majority.

    “In the type of “suppression-induced errors” (4 cases), the correct model maintained its position using statements like “I still maintain…” but failed to persuade the incorrect majority, leading to a wrong final vote, indicating possible entrenchment or mutual reinforcement of erroneous views.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    In the third round, 89 questions still remained without consensus; final answers for these were determined via majority voting.

    “In the third round, 89 questions still remained without consensus; final answers for these were determined via majority voting.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In the second round, 138 more correct consensuses were achieved, along with 8 incorrect ones.

    “In the second round, 138 more correct consensuses were achieved, along with 8 incorrect ones.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the first-round debate, the models reached a correct consensus on 104 questions and an incorrect consensus on 6 questions.

    “In the first-round debate, the models reached a correct consensus on 104 questions and an incorrect consensus on 6 questions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    MCC significantly reduced answer defect rates by 3.2%-9.1% across these metrics, with statistically significant gains (p < 0.05) in seven dimensions.

    “MCC significantly reduced answer defect rates by 3.2%–9.1% across these metrics, with statistically significant gains (p < 0.05) in seven dimensions.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Physician and layperson ratings reliably measure the quality of long-form medical responses.

  • May combine two claims.

    Findingown data

    Overall accuracy in two-model MCC frameworks declined relative to the full MCC setting, in some cases even dropping below the performance of individual LLMs, primarily due to ineffective debates ending in ties.

    “The results showed that overall accuracy in two-model MCC frameworks declined relative to the full MCC setting, in some cases even dropping below the performance of individual LLMs, primarily due to ineffective debates ending in ties (Table S3).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    MCC's responses were more frequently rated as directly helpful and better aligned with user intent than those of all individual LLMs by layperson raters.

    “MCC’s responses were more frequently rated as directly helpful and better aligned with user intent than those of all individual LLMs (Figure 3D).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Physician and layperson ratings reliably measure the quality of long-form medical responses.

  • May combine two claims.

    Findingown data

    MCC's PICR was significantly higher than that of all baseline LLMs, with median paired improvements of +0.29, +0.15, and +0.11 compared to GPT-o1, Qwen-QwQ, and DeepSeek-R1, respectively (all Bonferroni-adjusted p < 0.01).

    “Formal paired analyses indicated that MCC’s PICR was significantly higher than that of all baseline LLMs, with median paired improvements of +0.29, +0.15, and +0.11 compared to GPT-o1, Qwen-QwQ, and DeepSeek-R1, respectively (all Bonferroni-adjusted p < 0.01).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

  • May combine two claims.

    Findingown data

    For PRQR, MCC achieved values above 80% in 14 cases, whereas GPT-o1, Qwen-QwQ, and DeepSeek-R1 exceeded this threshold in 10, 13, and 14 cases, respectively.

    “For PRQR, MCC achieved values above 80% in 14 cases, whereas GPT-o1, Qwen-QwQ, and DeepSeek-R1 exceeded this threshold in 10, 13, and 14 cases, respectively (Figure 4A).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

  • May combine two claims.

    Findingown data

    On MetaMedQA, MCC achieved an overall accuracy of 84.1%, outperforming individual LLMs (GPT-o1: 76.1%; Qwen-QwQ: 72.6%; DeepSeek-R1: 81.7%) and the best-performing model reported in the original study (GPT-4o: 78.8%).

    “Specifically, MCC achieved an overall accuracy of 84.1%, outperforming individual LLMs (GPT-o1: 76.1%; Qwen-QwQ: 72.6%; DeepSeek-R1: 81.7%) and best-performing model reported in the original study (GPT-4o: 78.8%).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

  • May combine two claims.

    Findingown data

    MCC incurred +$6.5 in total ($25.5 vs. $19.0) and +$0.005 per case on average ($0.020 vs. $0.015).

    “In monetary cost, MCC incurred +$6.5 in total ($25.5 vs. $19.0) and +$0.005 per case on average ($0.020 vs. $0.015).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    MCC significantly outperformed GPT-o1 on PRQR (+0.08; Bonferroni-adjusted p < 0.001), while the improvements over Qwen-QwQ and DeepSeek-R1 were smaller and not statistically significant.

    “MCC significantly outperformed GPT-o1 (+0.08; Bonferroni-adjusted p < 0.001), while the improvements over Qwen-QwQ and DeepSeek-R1 were smaller and not statistically significant.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

  • May combine two claims.

    Findingown data

    In the RABBITS MedMCQA subsets, GPT-o1 (from 89.7% to 86.8%, −2.9%) and Qwen-QwQ (from 89.1% to 85.9%, −3.2%) exhibited declines, while DeepSeek-R1 (from 91.1% to 89.7%, −1.4%) showed a moderate drop and MCC (from 93.4% to 92.8%, −0.6%) remained comparatively stable.

    “In the MedMCQA subsets, GPT-o1 (from 89.7% to 86.8%, −2.9%) and Qwen-QwQ (from 89.1% to 85.9%, −3.2%) exhibited declines, while DeepSeek-R1 (from 91.1% to 89.7%, −1.4%) showed a moderate drop and MCC (from 93.4% to 92.8%, −0.6%) remained comparatively stable.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Compared with majority voting, MCC required +6.9 h in total time (23.7 h vs. 16.8 h) and +19.5 s per case on average (67.0 s vs. 47.5 s).

    “As shown in Table S2, compared with majority voting, MCC required +6.9 h in total time (23.7 h vs. 16.8 h) and +19.5 s per case on average (67.0 s vs. 47.5 s), largely dominated by the slowest constituent model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Across these 12 dimensions, MCC achieved 1.6%-11.2% lower flawed-content rates, with statistically significant improvements in 10 (permutation p < 0.05), most prominently in comprehension (7.6%, p < 0.0001) and recall (11.2%, p < 0.0001).

    “Across these 12 dimensions, MCC achieved 1.6%–11.2% lower flawed-content rates, with statistically significant improvements in 10 (permutation p < 0.05), most prominently in comprehension (7.6%, p < 0.0001) and recall (11.2%, p < 0.0001) (Figure 3B).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Physician and layperson ratings reliably measure the quality of long-form medical responses.

  • May combine two claims.

    Findingown data

    On PubMedQA, MCC attains 84.8% ± 0.4% accuracy, exceeding MDTeamGPT (83.9%) by +0.9 percentage points.

    “On PubMedQA, MCC attains 84.8% ± 0.4%, exceeding MDTeamGPT (83.9%) by +0.9 percentage points.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For PICR, MCC achieved scores above 80% in 14 of 16 cases; GPT-o1 did not exceed the 80% threshold in any case, while Qwen-QwQ and DeepSeek-R1 reached this level in only 4 and 8 cases, respectively.

    “For PICR, MCC achieved scores above 80% in 14 of 16 cases. In contrast, GPT-o1 did not exceed the 80% threshold in any case, while Qwen-QwQ and DeepSeek-R1 reached this level in only 4 and 8 cases, respectively (Figure 4A).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

  • May combine two claims.

    Findingown data

    MCC achieved notable gains in correct retrieval (11.8%, p < 0.001), correct reasoning (8.8%, p < 0.001), and possibility of bias reduction (11.1%, p < 0.001).

    “MCC consistently outperformed all individual LLMs across all 12 physician-rated criteria (Figure 3A), achieving notable gains in correct retrieval (11.8%, p < 0.001), correct reasoning (8.8%, p < 0.001), and possibility of bias reduction (11.1%, p < 0.001).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Physician and layperson ratings reliably measure the quality of long-form medical responses.

  • May combine two claims.

    Findingown data

    In the RABBITS MedQA subsets, accuracy declined modestly for GPT-o1 (from 90.2% to 87.8%, −2.4%), Qwen-QwQ (from 88.6% to 86.8%, −1.8%), and DeepSeek-R1 (from 91.8% to 90.7%, −1.1%), while MCC showed minimal impact (from 94.2% to 93.7%, −0.5%).

    “In the MedQA subsets, accuracy declined modestly for GPT-o1 (from 90.2% to 87.8%, −2.4%), Qwen-QwQ (from 88.6% to 86.8%, −1.8%), and DeepSeek-R1 (from 91.8% to 90.7%, −1.1%), while MCC showed minimal impact (from 94.2% to 93.7%, −0.5%).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On MedQA, MCC achieves 92.6% ± 0.3% accuracy, surpassing MDTeamGPT (90.1%) by +2.5 percentage points.

    “Specifically, on MedQA, MCC achieves 92.6% ± 0.3%, surpassing the previous best method MDTeamGPT (90.1%) by +2.5 percentage points.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On MedXpertQA, MCC achieved an accuracy of 40.2% on the 1,861 reasoning questions and 40.1% on the 589 understanding questions, outperforming all individual LLM baselines.

    “On this benchmark, MCC achieved an accuracy of 40.2% on the 1,861 reasoning questions and 40.1% on the 589 understanding questions, outperforming all individual LLM baselines (GPT-o1, Qwen-QwQ, and DeepSeek-R1).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

  • May combine two claims.

    Findingown data

    Paired analysis confirmed that MCC significantly outperformed GPT-o1 (+0.25; Bonferroni-adjusted p = 0.01) and Qwen-QwQ (+0.33; Bonferroni-adjusted p = 0.02) on SDA, while differences vs. DeepSeek-R1 were not significant.

    “Paired analysis confirmed that MCC significantly outperformed GPT-o1 (+0.25; Bonferroni-adjusted p = 0.01) and Qwen-QwQ (+0.33; Bonferroni-adjusted p = 0.02), while differences vs. DeepSeek-R1 were not significant.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

  • May combine two claims.

    Findingown data

    Exact McNemar tests did not reach significance for PDA (p = 0.50, 0.25, and 0.50 for GPT-o1, Qwen-QwQ, and DeepSeek-R1, respectively), as only 2-3 discordant pairs were observed per comparison, indicating limited statistical power.

    “However, exact McNemar tests did not reach significance (p = 0.50, 0.25, and 0.50 for GPT-o1, Qwen-QwQ, and DeepSeek-R1, respectively), as only 2–3 discordant pairs were observed per comparison, indicating limited statistical power.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the HealthBench Consensus subset (n = 3,671), MCC achieved a score of 0.921, outperforming baseline LLMs (GPT-o1: 0.873; Qwen-QwQ: 0.866; DeepSeek-R1: 0.887).

    “On the HealthBench Consensus subset (n = 3,671), MCC achieved a score of 0.921, outperforming baseline LLMs (GPT-o1: 0.873; Qwen-QwQ: 0.866; DeepSeek-R1: 0.887) (Table S10).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

  • May combine two claims.

    Findingown data

    The largest improvements were observed in reading comprehension (9.1%, p < 0.0001), knowledge recall (7.5%, p < 0.0001), consensus alignment (6.3%, p = 0.0006), and demographic bias reduction (6.4%, p = 0.0010).

    “The largest improvements were observed in reading comprehension (9.1%, p < 0.0001), knowledge recall (7.5%, p < 0.0001), consensus alignment (6.3%, p = 0.0006), and demographic bias reduction (6.4%, p = 0.0010) (Figure 3C).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Physician and layperson ratings reliably measure the quality of long-form medical responses.

  • May combine two claims.

    Findingreasoning

    In MCC debates of multiple-choice questions on the MedQA benchmark, most of the choice transitions are from incorrect to correct, while reverse transitions-from correct to incorrect-are rare.

    “In MCC debates of multiple-choice questions, an interesting observation in MedQA benchmark is that most of the choice transitions are from incorrect to correct, while reverse transitions—from correct to incorrect—are rare.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The three-replica configuration showed limited improvement, only slightly surpassing individual model performance, indicating little complementarity among instances of the same model.

    “The three-replica configuration also showed limited improvement, only slightly surpassing individual model performance, indicating little complementarity among instances of the same model (Table S3).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Among 15 cases with a clearly defined primary diagnosis, MCC achieved a PDA of 80.0% (12/15), outperforming GPT-o1 and DeepSeek-R1 (both 66.7%, 10/15) and Qwen-QwQ (60.0%, 9/15).

    “Among 15 cases with a clearly defined primary diagnosis, MCC achieved a PDA of 80.0% (12/15), outperforming GPT-o1 and DeepSeek-R1 (both 66.7%, 10/15) and Qwen-QwQ (60.0%, 9/15) (Figure 4A).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    The actual cost and time of running MCC vary depending on the number of debate rounds and whether consensus is reached.

    “Notably, the actual cost and time of running MCC vary depending on the number of debate rounds and whether consensus is reached.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given

    MCC requires increased computational cost and longer inference time compared with single-model baselines, which may affect its scalability and practical deployability in clinical settings.

    “Third, MCC requires increased computational cost and longer inference time compared with single-model baselines, which may affect its scalability and practical deployability in clinical settings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Per-round inference time, token usage, and cost estimates can be reasonably approximated using the per-case means.

    “Per-round inference time, token usage, and cost estimates can be reasonably approximated using the per-case means.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    Computational costs (runtime, token usage, financial expenditure) vary with dataset size and the proportion of disagreement-triggered cases requiring multi-round deliberation.

    “The computational costs, expressed in terms of runtime, token usage, and financial expenditure, vary with dataset size and the proportion of disagreement-triggered cases requiring multi-round deliberation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Under layperson-clinician dual assessments and expert rubrics, MCC generated higher-quality, more logically rigorous, and safer responses, indicating its potential relevance to doctor-patient communication and patient health education.

    “Under layperson-clinician dual assessments and expert rubrics, MCC generated higher-quality, more logically rigorous, and safer responses, indicating its potential relevance to doctor-patient communication and patient health education.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    MCC is a scalable, model-agnostic framework that advances medical reasoning through collaborative deliberation.

    “These results position MCC as a scalable, model-agnostic framework that advances medical reasoning through collaborative deliberation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Multi-LLM debate mirrors MDT discussions and collective diagnostic reasoning, where independent expert perspectives challenge and cross-validate each other to reduce diagnostic errors and enhance reliability.

    “Clinically, multi-LLM debate mirrors both multidisciplinary team (MDT) discussions and collective diagnostic reasoning, where independent expert perspectives challenge and cross-validate each other to reduce diagnostic errors and enhance reliability—providing a direct clinical rationale for adopting multi-LLM reasoning.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Ten physician raters assessed responses across 12 independent rating dimensions.

    “We first recruited 10 physician raters to assess responses across 12 independent rating dimensions (Table S7).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Ten layperson raters evaluated model responses in terms of helpfulness and relevance.

    “Complementing the physician assessment, we engaged 10 layperson raters to evaluate model responses in terms of helpfulness and relevance (Table S9).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. The quote only partly states this claim.

    Study design and conductown data

    Bonferroni adjustment was applied to paired analyses of PICR, PRQR, and SDA.

    “Formal paired analyses indicated that MCC’s PICR was significantly higher than that of all baseline LLMs, with median paired improvements of +0.29, +0.15, and +0.11 compared to GPT-o1, Qwen-QwQ, and DeepSeek-R1, respectively (all Bonferroni-adjusted p < 0.01).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The physician evaluation was extended to a blinded comparative-ranking assessment using nine predefined quality metrics.

    “We then extended the physician evaluation to a blinded comparative-ranking assessment using nine predefined quality metrics (Table S8).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Permutation tests were used to assess statistical significance in the long-form evaluation.

    “Across these 12 dimensions, MCC achieved 1.6%–11.2% lower flawed-content rates, with statistically significant improvements in 10 (permutation p < 0.05), most prominently in comprehension (7.6%, p < 0.0001) and recall (11.2%, p < 0.0001) (Figure 3B).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    MCC operates without task-specific training, external knowledge injection, or tool invocation; instead, it enhances reasoning by leveraging structured peer interactions among diverse LLMs oriented toward the reasoning process.

    “By contrast, MCC operates without task-specific training, external knowledge injection, or tool invocation; instead, it enhances reasoning by leveraging structured peer interactions among diverse LLMs oriented toward the reasoning process.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    A structured OSCE-style evaluation using 16 simulated patient scenarios covering a broad range of clinical contexts was conducted, evaluating MCC performance in two stages (history-taking and disease diagnosis).

    “We conducted a structured OSCE-style evaluation using 16 simulated patient scenarios covering a broad range of clinical contexts, which evaluated MCC performance in two stages (history-taking and disease diagnosis) (STAR Methods).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    In headache diagnosis, neurology and ophthalmology agents may pursue parallel reasoning paths without effective reconciliation.

    “For instance, in headache diagnosis, neurology and ophthalmology agents may pursue parallel reasoning paths without effective reconciliation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    MCC enables flexible, role-agnostic collaboration free from architectural constraints, integrating heterogeneous models while preserving the full dialog context to support perspective integration and reduce collective bias.

    “In contrast, MCC enables flexible, role-agnostic collaboration free from architectural constraints, integrating heterogeneous models while preserving the full dialog context to support perspective integration and reduce collective bias.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Many multi-LLM systems utilize homogeneous model families.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    MCC demonstrates significant potential for real-world healthcare applications.

    “MCC demonstrates significant potential for real-world healthcare applications.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    MCC's responses were not only broader in medical coverage but also more coherent and actionable, reflecting integration of diverse disciplinary perspectives-unlike the fragmented outputs typical of single models.

    “Notably, MCC’s responses were not only broader in medical coverage but also more coherent and actionable, reflecting integration of diverse disciplinary perspectives—unlike the fragmented outputs typical of single models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    Across 10 repeated trials, MCC consistently encountered items for which all initial responses failed, suggesting a ceiling effect that likely reflects inherent limitations in the benchmark design-such as ambiguous, underspecified, or flawed items-rather than the true upper bound of model capability.

    “Across 10 repeated trials, MCC consistently encountered items for which all initial responses failed, suggesting a ceiling effect that likely reflects inherent limitations in the benchmark design—such as ambiguous, underspecified, or flawed items—rather than the true upper bound of model capability.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    In clinical diagnostics, MCC's debate-driven mechanism serves as a reasoning aid for complex cases, providing comprehensive perspectives to support multidisciplinary treatment planning and precision decision-making.

    “In clinical diagnostics, its debate-driven mechanism serves as a reasoning aid for complex cases, providing comprehensive perspectives to support multidisciplinary treatment planning and precision decision-making.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Exact McNemar tests were used for PDA comparisons.

    “However, exact McNemar tests did not reach significance (p = 0.50, 0.25, and 0.50 for GPT-o1, Qwen-QwQ, and DeepSeek-R1, respectively), as only 2–3 discordant pairs were observed per comparison, indicating limited statistical power.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    MCC's transparent reasoning process and self-correction capabilities enhance trustworthiness and safety in medical AI applications.

    “The transparent reasoning process and self-correction capabilities enhance trustworthiness and safety in medical AI applications.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    Raters were blinded to the source of each response, and the display order of responses was randomized to prevent ordering bias.

    “Raters were blinded to the source of each response, and the display order of responses was randomized to prevent ordering bias.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationown data

    The concise format enables clinicians to efficiently review core reasoning pathways and consensus points under time constraints.

    “This concise format enables clinicians to efficiently review the core reasoning pathways and consensus points under time constraints, while the complete debate logs remain available for in-depth traceability and analysis.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The geographic homogeneity of the human evaluators (all recruited from a single region in China) may limit the generalizability of the findings to broader populations.

    “First, the geographic homogeneity of the human evaluators (all recruited from a single region in China) may limit the generalizability of the findings to broader populations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Different LLMs may express the same diagnosis in varied linguistic forms, so simple lexical matching is insufficient to determine agreement.

    “However, since different LLMs may express the same diagnosis in varied linguistic forms, simple lexical matching is insufficient to determine agreement.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    By exposing participants to these articulated reasoning steps in real time, MCC can help reinforce key diagnostic concepts and deepen learners' intuitive grasp of clinical reasoning.

    “By exposing participants to these articulated reasoning steps in real time, MCC can help reinforce key diagnostic concepts and deepen learners’ intuitive grasp of clinical reasoning.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The MCC framework supports flexible API-based integration with a wide range of commercially available LLMs, enabling access without code modification.

    “The MCC framework supports flexible API-based integration with a wide range of commercially available LLMs, enabling access without code modification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    MCC possesses zero-shot scalability.

    “Considering that MCC possesses zero-shot scalability, continued advances in the field of LLMs will enable future researchers to integrate higher-quality models to further explore its performance boundaries.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    MCC's performance boundaries remain uncertain when dealing with highly ambiguous, inconsistent, or sparse real-world clinical data, where debate may amplify rather than resolve uncertainty.

    “Second, MCC’s performance boundaries remain uncertain when dealing with highly ambiguous, inconsistent, or sparse real-world clinical data, where debate may amplify rather than resolve uncertainty.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    All raters were blinded to the source of the responses and conducted their evaluations independently.

    “All raters were blinded to the source of the responses and conducted their evaluations independently.”
    Quote in the paperSubstantial claim
71 of 71

  • The paper relies on this without support.

    The medical benchmarks used (MedQA, PubMedQA, MMLU, MedXpertQA, MetaMedQA, RABBITS, HealthBench) validly measure medical reasoning ability.

    “We integrated multiple benchmarks, including multiple-choice questions and long-form question answering, to evaluate the effectiveness of MCC (Figure 1A).”

  • The paper relies on this without support.

    The OSCE-style simulated patient scenarios validly represent real clinical diagnostic interactions.

    “We conducted a structured OSCE-style evaluation using 16 simulated patient scenarios covering a broad range of clinical contexts, which evaluated MCC performance in two stages (history-taking and disease diagnosis) (STAR Methods).”

  • The paper relies on this without support.

    Physician and layperson ratings reliably measure the quality of long-form medical responses.

    “We first recruited 10 physician raters to assess responses across 12 independent rating dimensions (Table S7).”

  • The paper relies on this without support.

    Many multi-LLM systems utilize homogeneous model families.

    “Third, many systems utilize homogeneous model families, whereas our ablation studies confirm that limited diversity in knowledge sources and reasoning approaches reduces overall performance.”

  • The paper relies on this without support.

    Overly verbose or highly technical responses can increase cognitive load and introduce psychological barriers for patients.

    “In real-world medical interactions, overly verbose or highly technical responses can increase cognitive load and introduce psychological barriers for patients.”
5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • MedQA debate-round counts exceed the unresolved cases

    Major

    The reported first-round outcomes (104 + 6), second-round outcomes (138 + 8), and 89 cases remaining after the third round total 345, although only 254 cases entered debate. The first two rounds alone also report 242 correct consensuses, exceeding the stated final total of 241 correct answers among unresolved cases.

    Reported 254 · recomputed 345

  • Case-package description conflicts with the number of defined primary diagnoses

    Minor

    The methods say each of the 16 case packages contained a gold-standard primary diagnosis, but the results say only 15 cases had a clearly defined primary diagnosis.

2 of 2

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Authors
Sun, X., Hong, Q., Zhang, M., Li, Y., Chen, T., Huang, Z., Liang, G., Tang, W., Xu, S., Ni, X., Pang, J., Wan, P., & Long, E.
Affiliations
  • Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
  • Peking University (CN)
Corresponding author
  • Wan, Peixing
  • Long, Erping
ORCID
Venue
Cell Reports Medicine
Publisher
Elsevier BV
Published
2026-01
Volume
7
Issue
1
Pages
102547
ISSN
2666-3791 (ISSN-L)
DOI
10.1016/j.xcrm.2025.102547
PMC
PMC12866169
OpenAlex
W7118208329
Cited by
1
References
16
Citations by year
2026: 1
FWCI
2.15
Type
journal article
Language
en
Open access
Licence
Locations
  • Cell Reports Medicine (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (submittedVersion, cc-by-nc-nd)
Funders
  • National Key Research and Development Program of China (2024ZD0528500)
  • National Natural Science Foundation of China (32470635, 32300483, 82090011)
  • Chinese Academy of Medical Sciences (2023-I2M-2-001, 2023-I2M-3-010)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
large language models, medical reasoning, peer interaction, self-correction, multi-turn dialogue, medical question answering, clinical decision-making, MedQA
MeSH terms
Clinical Decision-Making, Clinical Reasoning, Large Language Models, Communication, Cooperative Behavior

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review