Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Vision-Language Models (VLMs) show great promise as diagnostic aids for breast ultrasound, but they easily fall into a trap called “sycophancy.” This means if a clinician makes a biased suggestion, the model often abandons its correct initial judgement just to agree with the human. Such a flaw can dangerously delay necessary biopsy referrals. To address this issue, existing text-only defense methods lack both clinical evaluation and post-hoc correction tailored for medical VLMs.We introduce Pressure-Tuned Auditor (PTA), a two-agent framework designed to tackle this challenge. First, we establish the first clinically motivated sycophancy benchmark, pairing breast ultrasound cases with clean or misleading prompts generated from 16 realistic templates. Second, we implement multimodal Pressure-Tune training, which combines structured rationale supervision with visual grounding to pinpoint localizable ultrasound evidence. Third, we design a two-agent audit pipeline where a Pressure-Tuned Agent 2 independently verifies and corrects Agent 1’s outputs. Compared with three competitive baselines under identical fine-tuning budgets, our approach drastically reduces the sycophancy rate from 20.8% to just 0.9%, while maintaining a 98.5±1.5% cancer detection accuracy with zero harmful false alarms. Furthermore, visual grounding significantly improves lesion localization (IoU 0.59 vs.0.12), offering transparent and reviewable spatial evidence.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4746_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

https://www.kaggle.com/datasets/aryashah2k/breast-ultrasound-images-dataset

BibTex

@InProceedings{ZhaHon_PressureTuned_MICCAI2026,
        author = { Zhang, Hongze AND Fu, Yuxuan AND Tan, Xiaoyu AND Qiu, Xihe},
        title = { { Pressure-Tuned Auditor for Sycophancy-Resistant Breast Ultrasound VLM Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors proposed the Pressure-Tuned Auditor (PTA), a framework with two VLM agents that 1) one unmodified Agent 1 to provide diagnosis; 2) another pressure-tuned Agent 2 to verify and correct the diagnosis to avoid sycophancy. The authors also establish a clinically grounded sycophancy benchmark.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    • The paper is well-written, with a clear and structured presentation of the proposed method that is easy to follow.

    • Beyond the methodology, the authors contribute a new sycophancy benchmark specifically for breast ultrasound, providing a valuable resource for the medical AI community.

    • The experimental results successfully demonstrate the effectiveness of the proposed approach in the targeted scenarios.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    • All used prompts and template should be given for evaluation and reproducibility.

    • Do baselines like Syndata and SPT utilize visual grounding during LoRA training? To ensure a fair comparison, please clarify if the baselines used the same level of visual grounding during training.

    • The “prompt-only auditor” is a crucial ablation baseline to quantify the pressure-tuned’s actual gain. Although mentioned as future work, its absence makes it difficult to fully evaluate the core contribution.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The proposed is promising and the contribution of the new sycophancy benchmark for breast ultrasound is valuable. However, my recommendation is a Weak Accept because the current experimental discussion lacks the depth required to fully demonstrate the method’s robustness and generalizability. While the work is promising, more thorough analysis is needed to justify the broader impact of the findings.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I believe this paper demonstrates sufficient novelty, despite a few minor issues.



Review #2

  • Please describe the contribution of the paper

    This paper introduces a method to enhance Vision-Language Models used in breast ultrasound in order to make an anti-sycophancy framework. It proposes a two-agent framework that addresses this problem. Their results show reduction in sycophancy as well as achieving good accuracy. The authors have also claimed that this is the first clinically-grounded sycophancy benchmark for medical VLMs.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Contribution of visual evidence for judgment.
    • Extending Pressure-Tune from a text-only LLMs to the multimodal setting.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The way visual grounding provides spatial evidence for radiologist verification is not clearly and sufficiently explained.
    • Challenges in handling unseen comments in clinical reports.
  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
    • Further elaboration is needed on how visual grounding provides spatial evidence to support radiologist verification.
    • Provide some examples to show the challenges associated with handling unseen comments in clinical reports.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Limited novelty and a design that appears to be closely tailored to a specific application setting, which may limit its general applicability.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes Pressure-Tuned Auditor (PTA), a two-agent framework to mitigate sycophancy in vision–language models (VLMs) for breast ultrasound diagnosis, where models may change correct predictions to agree with misleading clinician prompts. The authors introduce a clinically grounded 139-case paired sycophancy benchmark with 16 clinical templates using paired clean and misleading prompts. They extend Pressure-Tune training to the multimodal setting with adversarial rationale supervision and visual grounding and design an auditing agent that verifies and corrects predictions from a base VLM. Experiments on a breast ultrasound cohort show that PTA reduces sycophancy from 20.4% to 0.9%, achieves 99.3% cancer classification accuracy, and improves lesion localization quality, suggesting improved robustness and interpretability for clinical VLM deployment.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Clinically meaningful problem: The paper addresses sycophancy in medical vision–language models (VLMs), a realistic failure mode where models may change correct predictions under misleading clinician prompts. This is an important and underexplored safety issue for ultrasound-based decision support. Also, Table 2 demonstrates that all four off-the-shelf VLMs exhibit substantial sycophancy (15.1%–51.5%), including models with high clean accuracy, effectively motivating the problem. 2.Clinically grounded sycophancy benchmark: The paired clean vs misleading prompt evaluation using 16 radiology-style templates provides a practical and interpretable way to measure diagnostic flips. This is stronger than generic text-only sycophancy benchmarks and is a useful contribution for evaluating medical VLM robustness. 3.Practical two-agent auditing framework: The proposed auditor architecture enables post-hoc correction of predictions under prompt pressure. This is a realistic design choice for deployment settings where clinician interaction may introduce bias. 4.Visual grounding improves interpretability: Adding grounding supervision improves lesion localization (IoU 0.65 vs 0.19), which strengthens the clinical usability of the auditing framework.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.PTA does not outperform SPT on sycophancy reduction: The paper’s headline claim is sycophancy reduction, yet SPT strictly dominates PTA on every sycophancy-related measure: 0.0% vs. 0.9% pipeline SR, 100% vs. 95.5% catch rate, 0 vs. 1 missed flip, and identical false alarms. PTA only leads on cancer accuracy, but the Wilson CIs overlap (SPT: 93.85–99.26%, PTA: 96.04–99.87%), making that advantage statistically inconclusive. The paper never directly acknowledges that its proposed method is outperformed on its primary metric by a baseline. 2.Ablation conclusion in Table 3 is not well-supported: The paper claims “PT is sufficient for sycophancy resistance” and presents Full (PT+VG)’s zero false alarms as an improvement over No VG. However, the entire comparison rests on a single-case difference - 1 missed flip vs. 1 false alarm across only 22 sycophancy events. A difference of this magnitude on a sample this small is statistically meaningless and cannot support a definitive recommendation. The paper should acknowledge this uncertainty and provide a proper tradeoff analysis, including discussion of the clinical implications: a missed sycophantic flip (malignant case downgraded, biopsy deferred) and a false alarm (correct diagnosis reviewed unnecessarily) carry very different consequences, and the choice between configurations should be informed by this asymmetry rather than treated as equivalent. 3.Missing inference-only two-agent baseline: Without a two-agent prompting baseline (self-critique without fine-tuning), it is impossible to separate the contribution of the two-agent architecture from the Pressure-Tune training. This is a critical missing ablation given that the two-agent setup alone may account for much of the observed gain. 4.Missing prompt-only auditor baseline: The paper does not compare against inference-only two-agent prompting (e.g., self-critique/debate), making it difficult to isolate the benefit of Pressure-Tune training from the auditing architecture itself. 5.Limited evaluation scale and generalization evidence: Results are reported on a 139-case single-center test set. Multi-center validation would strengthen claims about clinical robustness. 6.Moderate methodological novelty: The framework integrates existing components (Pressure-Tune, LoRA, multi-agent auditing, grounding) rather than introducing a fundamentally new learning formulation. 7.Limited BI-RADS grading performance: Although cancer classification accuracy is strong, fine-grained BI-RADS grading remains weaker, particularly within 4a/4b/4c categories, which are clinically relevant distinctions.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses a clinically important problem, and the benchmark is a genuine contribution. However, the experimental results do not support the claims as presented. Most critically, Table 1 shows that the SPT baseline achieves a lower pipeline sycophancy rate (0.0%) than PTA (0.9%), correcting all flips without false alarms. While PTA improves cancer classification accuracy, the robustness advantage over prior work is therefore not clearly established and should be interpreted more cautiously.. Furthermore, the absence of an inference-only two-agent prompting baseline makes it difficult to isolate the contribution of Pressure-Tune training from the two-agent auditing architecture. Finally, the evaluation is limited to a 139-case single-center cohort with templated misleading prompts and overlapping confidence intervals between methods, which does not provide statistically conclusive evidence of superiority over strong baselines. In addition, the methodological novelty is moderate, as the framework primarily combines existing components rather than introducing a fundamentally new learning formulation, which further weakens the case for acceptance, given the inconclusive robustness improvement over prior approaches.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The paper addresses sycophancy in VLMs for breast ultrasound diagnosis, a clinically realistic and underexplored safety problem. Its main contributions are a clinically grounded sycophancy benchmark with 139 paired cases and 16 misleading templates, a multimodal extension of Pressure-Tune, and a two-agent audit pipeline that returns a diagnosis, reason, and optional bounding-box evidence.

    My initial rejection was mainly based on Table 1, where SPT achieved 0.0% pipeline SR while PTA achieved 0.9% SR. However, the rebuttal shows that this difference is only one case. The reported SR confidence intervals overlap, and McNemar’s test comparing PTA and SPT on the 22 paired flips gives p = 1.00.Therefore, the results do not provide statistical evidence that SPT is superior to PTA on sycophancy resistance. A fairer interpretation is that PTA achieves comparable near-zero sycophancy while adding a deployable audit interface with structured rationale and visual grounding, which SPT does not provide.

    The rebuttal also clarified that the Prompt Defense baseline was an inference-only two-agent setup without fine-tuning. This partially addresses my concern about a missing prompt-only auditor baseline, although the manuscript must describe this clearly. The authors also appropriately reframed VG as mainly supporting lesion localization, not SR reduction. This is reasonable because the IoU improvement from 0.19 to 0.65 provides stronger evidence for the visual grounding contribution than the small Table 3 SR difference.

    Some concerns remain. First, Agent 1’s 5.8% clean BI-RADS accuracy is unexplained and should be addressed. Second, the Prompt Defense baseline needs a precise description in the paper. Third, the 16 templates and benchmarks should be released for reproducibility. Finally, the evaluation is still limited to a 139-case single-center cohort, and the paper should more clearly acknowledge this limitation and the asymmetric clinical cost of missed malignant downgrades versus false alarms.

    Overall, I recommend Weak Accept. The paper does not establish clear superiority over SPT, but it introduces an important benchmark and a clinically useful audit framework that achieves near-zero sycophancy while adding structured reasoning and lesion localization. Acceptance should be contingent on careful claim reframing and the clarifications above in the camera-ready.



Review #4

  • Please describe the contribution of the paper

    This paper identifies and addresses sycophancy in medical Vision-Language Models (VLMs) for breast ultrasound diagnosis — a clinically grounded problem where a model reverts a correct diagnosis upon receiving a biased suggestion from a clinician, potentially delaying biopsy referrals. The authors argue that this failure mode, largely overlooked in prior medical AI literature, poses a concrete risk in real-world clinical workflows governed by BI-RADS classification. To tackle this, the paper proposes the Pressure-Tuned Auditor (PTA), a two-agent framework in which a pressure-tuned Agent 2 independently audits and corrects Agent 1’s outputs. The authors claim three main contributions: (1) the first clinically-motivated sycophancy benchmark for medical VLMs, (2) a multimodal Pressure-Tune training strategy combining adversarial CoT supervision with visual grounding, and (3) a two-agent audit pipeline that the authors report achieves substantial reduction in sycophancy while maintaining high diagnostic accuracy.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The most notable strength of this paper lies in its clinical relevance. The problem of sycophancy in medical VLMs is not merely a theoretical concern — the authors effectively motivate the issue within a real-world clinical context, where a sycophantic diagnostic flip could directly result in a missed malignancy and delayed biopsy. This framing distinguishes the work from prior sycophancy studies that rely on generic trivia or opinion-based benchmarks. Additionally, the construction of a clinically-motivated sycophancy benchmark, paired with both clean and misleading prompts across multiple off-the-shelf VLMs, is a meaningful contribution to the community. Demonstrating that sycophancy is pervasive across diverse VLMs strengthens the motivation for the proposed framework.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    While the paper presents a clinically relevant problem and a seemingly functional framework, several significant weaknesses undermine the validity and generalizability of the proposed method.

    1.Dataset Scale and Statistical Reliability The evaluation is conducted on only 139 test cases, of which merely 22 exhibit sycophantic flips. Drawing conclusions about method superiority from such a small number of flip cases is statistically unreliable. For instance, the difference between PTA (21/22 caught) and SPT (22/22 caught) amounts to a single case, yet the authors use this to argue the advantage of their approach.

    2.Missing Critical Baseline As acknowledged by the authors themselves, the paper does not include an inference-only two-agent prompting baseline, such as self-critique or debate without fine-tuning. This omission makes it impossible to disentangle whether the performance gains stem from the two-agent architecture itself or from the Pressure-Tune training. Including such a comparison would have provided a more comprehensive analysis of each component’s contribution, and its absence is a missed opportunity to further strengthen the validity of the proposed framework.

    3.Discrepancy Between Training and Inference During training, Agent 2 is supervised on full Chain-of-Thought (CoT) responses. However, at inference, only a compact audit format is output rather than the full CoT. The authors do not provide sufficient justification or empirical evidence for why this discrepancy does not negatively impact performance. It remains unclear whether the model truly internalizes the structured reasoning, or whether the truncated inference output simply masks potential reasoning failures.

    4.Incomplete Notation in Section 2.2 (minor) In Section 2.2, the ground-truth label is defined as yi∈{B,M}, where B and M are used without explicit definition at the point of introduction. While the meaning of B (Benign) and M (Malignant) can be inferred from context, such notational ambiguity is undesirable in a technical paper and should be formally defined upon first use.

  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The primary factors leading to the rejection recommendation are as follows. The most critical concern is the limited scale and statistical reliability of the evaluation. With only 22 sycophantic flip cases in the test set, the quantitative comparisons between methods lack statistical grounding. Method superiority argued on the basis of a single-case difference is not a convincing basis for the claims made in this paper. Furthermore, the absence of an inference-only two-agent prompting baseline leaves a fundamental question unanswered: whether the observed improvements are attributable to the two-agent architecture or to the Pressure-Tune training itself. This unresolved ambiguity directly undermines the paper’s central contribution. While the clinical framing of the problem is commendable and the direction of research is promising, the current experimental design is insufficient to support the claims made. The single-institution dataset further limits the generalizability of the findings. These issues, taken together, suggest that the paper requires substantial additional experimentation before it can be considered for acceptance.

  • Reviewer confidence

    Not confident (1)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I appreciate the authors’ transparent rebuttal, which effectively addresses my main concerns. 1.Statistical Reliability: I appreciate the honest clarification using the McNemar test. I agree that while the sycophancy rate is statistically tied with the baseline, PTA offers a practical advantage through higher cancer accuracy and visual grounding via bounding boxes. 2.Missing Baseline: Clarifying that “Prompt Defense” is actually the inference-only two-agent baseline fully resolves my ablation concern. Please ensure this is explicitly clarified in the camera-ready version. With all my primary concerns resolved, I am happy to raise my score to Accept.



Author Feedback

We thank R1, R2, R3, R4.R1/R2 recognize the benchmark and clinical motivation; R3 also acknowledges “the benchmark is a genuine contribution” on an “important and underexplored” problem.

[R3, MR] Statistical reliability of PTA vs SPT: R3 acknowledges the Cancer-Acc Wilson CIs overlap (PTA [96.04%, 99.87%] vs SPT [93.85%, 99.26%]) and treats this as evidence PTA’s advantage is statistically inconclusive. By the symmetric standard, the same applies to SR: PTA [0.16%, 5.06%] vs SPT [0.00%, 3.43%]—fully overlapping; McNemar’s exact test on the 22 paired flips (b=0, c=1) yields p=1.00.Three implications: (i) both reach near-zero sycophancy (0.9% and 0.0%); (ii) the 0.0% vs 0.9% gap is a point estimate, not a confirmed advantage; (iii) a 1-case difference does not establish PTA is unambiguously outperformed on the main metric. The trade-off matters: PTA reaches comparably near-zero SR alongside higher cancer accuracy (138/139 vs 136/139, Table 1) and adds a deployable audit interface with bounding-box evidence (SPT does not). Both share Qwen2.5-VL-7B; PTA’s broader LoRA scope (attention+MLP, vs SPT’s top-5%) is needed for audit/bbox output, and is statistically equivalent on sycophancy.

[R1, R3, R4, MR] Inference-only two-agent baseline: Our submitted Prompt Defense baseline (Table 1) is in fact an inference-only two-agent configuration: Agent 1 (Qwen2.5-VL-7B, no FT) diagnoses under the misleading prompt; Agent 2 (Qwen2.5-VL-7B, no FT, with an anti-sycophancy system prompt) audits Agent 1’s output—no fine-tuning anywhere. The manuscript description (“prepends a system instruction”) understates this structure; we will expand it in the camera-ready. Prompt-only auditing reaches SR 3.7% with 11 false alarms, showing prompt engineering alone is insufficient; fine-tuning is needed both to reduce false alarms and to enable structured audit outputs—feature-grounded 4-step rationales and bounding-box evidence—that prompt-only cannot deliver.

[R3] PT/VG positioning and Table 3 ablation: We agree the 1 missed flip vs 1 false alarm difference is statistically inconclusive on 22 events, and will revise Sec 3.4.PT is the core mechanism for sycophancy resistance; VG is for lesion localization, not SR reduction. Full and No-VG are equivalent on SR within test-set precision. VG’s contribution rests on a robust effect: IoU 0.19→0.65 (Fig. 3, 3.4x), giving radiologists verifiable spatial correspondence between the audit decision and imaging evidence.

[R1] Baseline fairness: SynData and SPT do not use BUSI grounding and were trained under the same conditions as PTA—same Qwen2.5-VL-7B, same PT dialogues, same fine-tuning budget—differing only in fine-tuning scope. Templates, prompts, and code will be released with the camera-ready.

[R2] Handling unseen clinical comments: For free-form comments beyond our 16 templates (e.g., “patient is young, likely fibroadenoma”—demographics + clinical assumption), Agent 2 anchors decisions on imaging features rather than the comment, as illustrated by the audit Reason cited in our [R4] response.

[R4] Training-inference CoT gap: The compact JSON inherits 4-step CoT—not decision-only. Agent 2 emits a Reason like: “I note your benign view, but irregular margins, posterior shadowing, coarse calcifications support malignancy”—showing feature, evidence, conflict, decision. All per-case audits will be released upon acceptance. B/M will be defined in Sec2.2. [R2] VG spatial evidence: Agent 2 outputs a bounding box on the lesion it reasons over; radiologists overlay it on the BUS image to verify the cited features, binding the audit to localizable visual evidence.

[R3, R4] Cohort and clinical reports: Generating medically plausible misleading prompts requires public BUS datasets with detailed lesion-level annotations (e.g., echogenicity, margin); none provide such reports. Our 139 cases (from 665) include these annotations by sonographers with 5+ years of clinical experience and are pathology-confirmed.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The recommendations are mixed, with two favoring acceptance and one recommending rejection; hence, sending for rebuttal. Given several issues, especially that the proposed method does not clearly outperform a strong baseline on its own main metric and that key ablations are missing, the rebuttal could focus on these clarifications.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The authors provided convincing rebuttal, and the reviewers updated their scores accordingly. The fact that prompt engineering alone was insufficient indeed proves the value of the approach. The authors clearly outline their contributions and I hope the camera-ready version will entail these updates too.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Congratulations!



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    4/4 Accept (2 upgraded from Reject). Rebuttal successfully used statistical tests (McNemar’s test p=1.00) to refute critical concerns. Sycophancy benchmark is an important contribution to medical VLM safety.



back to top