pith. sign in

arxiv: 2607.00060 · v1 · pith:EFZR6FFXnew · submitted 2026-06-30 · 💻 cs.CV

Synergistic Perception-Reasoning Governance: Grounding Medical MLLMs with Verifiable Anatomical Evidence

Rui Hao , Qiankun Li , Junyuan Mao , Linghao Meng , Dirui Xie , Dayu Tan , Zhigang Zeng This is my paper

Pith reviewed 2026-07-02 20:02 UTC · model grok-4.3

classification 💻 cs.CV
keywords medical MLLMshallucination mitigationROI priorstraining-free frameworkanatomical evidenceclinical VQAradiology report generationdynamic router
0
0 comments X

The pith

A training-free framework grounds medical MLLMs with ROI-derived anatomical evidence to reduce hallucinations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a holistic, training-free method that injects verifiable anatomical evidence into multimodal large language models for clinical visual question answering and radiology report generation. It recalibrates visual perception by modulating activations with ROI priors from MedSAM and anchors textual reasoning by converting anatomical coordinates into discrete semantic tokens that act as external memory. A task-aware dynamic router then chooses modality-specific interventions to balance perceptual grounding against linguistic fluency. The approach targets inference-time hallucinations where models produce fluent but image-inconsistent outputs. If effective, it would let existing models improve reliability on medical benchmarks without retraining, external databases, or multi-stage verification pipelines.

Core claim

Dual-side evidence injection via ROI-guided activation modulation for perception and anatomical coordinate mapping to semantic tokens for reasoning, selected by a task-aware dynamic router, mitigates hallucinations in medical MLLMs, yielding up to 6 percent higher close-ended accuracy and 35 percent fewer open-ended hallucinations across LLaVA, Qwen, and InternVL models on five datasets.

What carries the argument

Dual-side evidence injection that recalibrates the visual perception trajectory with ROI-guided activation modulation and anchors the textual reasoning trajectory by mapping anatomical coordinates into discrete semantic tokens, chosen by a task-aware dynamic router.

If this is right

  • Close-ended medical VQA accuracy rises by up to 6 percent on the evaluated benchmarks.
  • Open-ended hallucinations drop by approximately 35 percent without added training cost.
  • The same framework applies to multiple base models including LLaVA-1.5-7B, Qwen3-VL variants, and InternVL-3.5 variants.
  • Controlled ablations confirm that both perception recalibration and reasoning anchoring contribute to the gains.
  • The method remains training-free and generalizes across two tasks and five datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The separation of perception recalibration from reasoning anchoring could be tested as independent modules on non-medical vision-language tasks.
  • If coordinate-to-token mapping proves robust, similar discrete memory mechanisms might reduce hallucinations in other grounded reasoning domains.
  • The router's task-semantic selection offers a testable route for extending the approach to new clinical subtasks without redesigning the full pipeline.

Load-bearing premise

ROI priors extracted by MedSAM stay accurate enough across different imaging conditions and the router chooses interventions without creating new perceptual or linguistic mistakes.

What would settle it

Performance gains disappear on a held-out test set of images where MedSAM ROI extraction produces large segmentation errors under varied contrast or noise conditions.

Figures

Figures reproduced from arXiv: 2607.00060 by Dayu Tan, Dirui Xie, Junyuan Mao, Linghao Meng, Qiankun Li, Rui Hao, Zhigang Zeng.

Figure 1
Figure 1. Figure 1: The overall pipeline consists of all input, ROI prior extraction, intervention mechanism, task-aware routing, and model inference for the output generation. 2 Methods Given an image I, a textual question q, and a pretrained MLLM parametrized by θ, the standard autoregressive generation is formulated as: pθ(y | I, q) = Y T t=1 pθ(yt | y<t,Enc(I), q), (1) where Enc(I) ∈ RN×d denotes the sequence of N visual … view at source ↗
Figure 2
Figure 2. Figure 2: Evaluation of mitigation strategies. (a) Overall accuracy comparison on MM￾VisHal and CXR-VisHal datasets. (b)-(c) Absolute metric improvements of individual components and our Routing over the LLaVA-Med-1.5-7B baseline. CHAIR(↓)CheXbert(↑)RadGraph(↑) RaTEScore(↑)Recall(↑) 0 10 20 30 40 Score (%) (a) LLaVA-1.5-7B CHAIR(↓)CheXbert(↑)RadGraph(↑) RaTEScore(↑)Recall(↑) 0 10 20 30 40 (b) LLaVA-Med-1.5-7B CHAIR(… view at source ↗
Figure 3
Figure 3. Figure 3: Open-ended evaluation results on the MIMIC-CXR for report generation. Sub￾figures (a) to (f) illustrate the evaluation results across six different MLLMs. RaTEScore to 44.18%, proving structured textual injection enhances narrative quality and coherence while effectively mitigating hallucinations. Obs 2: The central challenge in open-ended evaluation remains the difficulty of balancing report completeness … view at source ↗
Figure 4
Figure 4. Figure 4: Open-ended mitigation comparison on MIMIC-CXR using LLaVA-Med-1.5-7B. SLAKE VQA-RAD IU-Xray MIMIC-CXR 60 65 70 75 80 85 90 95 Total Accuracy (%) (a) Text-side: ROI vs. Random Bbox ROI-based bbox Random bbox (1.0, 0.1) (1.5, 0.1) (2.0, 0.1) (1.0, 1.5) (1.0, 2.0) Weighting Factors (α, β) 70 75 80 85 90 95 100 Total Accuracy (%) Optimal Configuration: (1.0, 0.1) (b) Vision-side: Ablation on (α, β) SLAKE VQA-R… view at source ↗
Figure 5
Figure 5. Figure 5: Controlled comparisons for text-side and vision-side methods. (a) Performance comparison between ROI-based and random bounding boxes for Qwen3-VL-8B. (b) Ablation study of weighting factors (α, β) for InternVL3.5-8B. Baseline Text-side Vision-side Joint Generate a medical report summarizing the key findings in the given Chest X-ray, including the main anatomical structures, relevant measurements(e.g., size… view at source ↗
Figure 6
Figure 6. Figure 6: Token Activation Map (TAM) for visualization presents high-quality localiza￾tion results on MIMIC-CXR datasets under the InternVL3.5-8B. Qualitative Analysis [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
read the original abstract

Multimodal large language models (MLLMs) show strong promise for clinical VQA and radiology report generation, yet inference-time hallucinations still undermine trustworthy use: models can produce fluent conclusions that conflict with imaging evidence. Existing mitigation strategies typically rely on additional training, external retrieval/knowledge bases, or multi-stage post-hoc verification, which increases cost and pipeline complexity and often generalizes poorly across models and tasks.To address this, we propose a holistic, training-free evidence-injection framework that systematically mitigates hallucinations through dual-side evidence injection. By leveraging ROI priors acquired using MedSAM in our implementation, we recalibrate the visual perception trajectory via ROI-guided activation modulation while anchoring the textual reasoning trajectory by mapping anatomical coordinates into discrete semantic tokens as verifiable external memory. Then we introduce a task-aware dynamic router to select modality-specific interventions based on task semantics, balancing perceptual grounding and linguistic fluency. We conduct systematic evaluations on 2 tasks and 5 datasets using \texttt{LLaVA-1.5-7B}, \texttt{LLaVA-Med-1.5-7B}, \texttt{Qwen3-VL-8B/32B}, and \texttt{InternVL-3.5-8B/38B}. Controlled ablations and visualizations further validate the framework, which consistently outperforms baselines across medical benchmarks, improving close-ended accuracy by up to $\sim\mathbf{6}\%\uparrow$ and reducing open-ended hallucinations by $\sim\mathbf{35}\%\downarrow$. The code has been made available on GitHub: \href{https://github.com/Henry991115/SPRG}{\textcolor{blue}{https://github.com/Henry991115/SPRG}}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes SPRG, a training-free framework for medical MLLMs that injects verifiable anatomical evidence via MedSAM-derived ROI priors: ROI-guided activation modulation for visual perception and coordinate-to-token memory injection for textual reasoning, selected by a task-aware dynamic router. Evaluations on close-ended VQA and open-ended report generation across five datasets and multiple models (LLaVA-1.5-7B, LLaVA-Med, Qwen3-VL, InternVL) report up to ~6% accuracy gains and ~35% hallucination reduction, supported by controlled ablations and visualizations; code is released.

Significance. If the central claims hold after addressing the MedSAM validation gap, the work provides a lightweight, model-agnostic method for grounding MLLMs with anatomical evidence without retraining or external retrieval, which could meaningfully improve trustworthiness in clinical VQA and report generation. The training-free design, multi-model evaluation, and public code release are positive factors for adoption and reproducibility.

major comments (2)
  1. [§4 and §3.2] §4 (Experiments) and §3.2 (ROI-guided modulation): the headline gains of ~6% accuracy and ~35% hallucination reduction presuppose that MedSAM ROIs are sufficiently accurate on the five evaluation datasets to modulate tokens and supply non-misleading coordinates, yet no quantitative MedSAM error rates (e.g., Dice/IoU on the test splits), no ablation replacing MedSAM ROIs with ground-truth segmentations, and no ablation with controlled noise are reported; this leaves open whether the improvements are robust or conditional on imaging conditions where MedSAM succeeds.
  2. [§3.3] §3.3 (task-aware dynamic router): the router selects modality-specific interventions based on task semantics, but the manuscript provides no mechanism to gate or down-weight low-confidence MedSAM ROIs before they reach the router or the memory injection step; without such a safeguard or an ablation that measures router error propagation under ROI noise, it is unclear whether the framework can avoid introducing new perceptual or semantic errors.
minor comments (2)
  1. [Abstract and §4] The abstract and §4 mention "controlled ablations" but do not list which components (router, modulation, memory injection) were ablated or report per-component deltas; adding an explicit ablation table would strengthen the validation.
  2. [§3] Notation for the coordinate-to-token mapping and the activation modulation function is introduced without an equation number or pseudocode; a compact formal definition would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the MedSAM validation gap and router safeguards. We address each point below and will revise the manuscript accordingly to strengthen the evidence for robustness.

read point-by-point responses

  1. Referee: [§4 and §3.2] §4 (Experiments) and §3.2 (ROI-guided modulation): the headline gains of ~6% accuracy and ~35% hallucination reduction presuppose that MedSAM ROIs are sufficiently accurate on the five evaluation datasets to modulate tokens and supply non-misleading coordinates, yet no quantitative MedSAM error rates (e.g., Dice/IoU on the test splits), no ablation replacing MedSAM ROIs with ground-truth segmentations, and no ablation with controlled noise are reported; this leaves open whether the improvements are robust or conditional on imaging conditions where MedSAM succeeds.

    Authors: We agree this is a substantive gap. The framework relies on MedSAM-derived ROIs, and without reported segmentation metrics or robustness ablations the source of the observed gains remains partially opaque. In the revised manuscript we will add: (1) Dice and IoU scores computed for MedSAM on the test splits of all five datasets, (2) an ablation that substitutes ground-truth segmentations (where available) for MedSAM ROIs, and (3) a controlled noise-injection study that perturbs ROI coordinates and masks at varying levels. These additions will directly quantify the dependence on MedSAM accuracy and demonstrate whether the dual-side injection remains beneficial under realistic segmentation error. revision: yes

  2. Referee: [§3.3] §3.3 (task-aware dynamic router): the router selects modality-specific interventions based on task semantics, but the manuscript provides no mechanism to gate or down-weight low-confidence MedSAM ROIs before they reach the router or the memory injection step; without such a safeguard or an ablation that measures router error propagation under ROI noise, it is unclear whether the framework can avoid introducing new perceptual or semantic errors.

    Authors: The observation is correct: the current router description contains no explicit confidence gating for MedSAM outputs, leaving open the possibility of error propagation. We will revise §3.3 to incorporate a lightweight confidence gate that uses MedSAM’s internal mask-quality scores (or an auxiliary uncertainty estimate) to down-weight or bypass low-confidence ROIs before they reach either the modulation or memory-injection modules. We will also add an ablation that systematically varies ROI noise levels and reports both router decision accuracy and end-to-end task metrics, thereby quantifying error propagation and the effectiveness of the proposed safeguard. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical evaluation of an external-component framework

full rationale

The manuscript describes a training-free method that injects MedSAM-derived ROI priors via activation modulation and coordinate-to-token mapping, then routes interventions with a task-aware dynamic router. All performance numbers (∼6% accuracy lift, ∼35% hallucination reduction) are reported as outcomes of controlled experiments on five datasets with multiple base MLLMs. No equations, fitted parameters, or derivation steps appear in the abstract or described pipeline; the gains are not presented as mathematical consequences of the method's own definitions. MedSAM is treated as an external black-box prior rather than a self-derived quantity. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling are referenced. The derivation chain is therefore self-contained against external benchmarks and does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract contains no mathematical derivations, fitted constants, or new postulated entities; all components reference existing tools (MedSAM, LLaVA variants, etc.).

pith-pipeline@v0.9.1-grok · 5862 in / 1050 out tokens · 27902 ms · 2026-07-02T20:02:15.307334+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    In: AAAI 2024 Spring Symposium Series (Clinical Foundation Models)

    Chen, Z., Varma, M., Delbrouck, J.-B., Paschali, M., Blankemeier, L., Van Veen, D., et al.: CheXagent: Towards a foundation model for chest X-ray interpretation. In: AAAI 2024 Spring Symposium Series (Clinical Foundation Models). OpenRe- view.net (2024).https://openreview.net/forum?id=P3LOmrZWGR

    2024

  2. [2]

    In: Proceedings of the 3rd Machine Learning for Health Symposium (ML4H 2023)

    Moor, M., Huang, Q., Wu, S., Yasunaga, M., Dalmia, Y., Leskovec, J., Zakka, C., Reis, E.P., Rajpurkar, P.: Med-Flamingo: A multimodal medical few-shot learner. In: Proceedings of the 3rd Machine Learning for Health Symposium (ML4H 2023). Proceedings of Machine Learning Research, vol. 225, pp. 353–367. PMLR (2023)

    2023

  3. [3]

    In: Proceedings of the 23rd Workshop on Biomedical Natural Language Processing (BioNLP 2024), pp

    Thawakar, O.C., Shaker, A.M., Mullappilly, S.S., Cholakkal, H., Anwer, R.M., Khan, S., Laaksonen, J., Khan, F.: XrayGPT: Chest radiographs summarization using large medical vision-language models. In: Proceedings of the 23rd Workshop on Biomedical Natural Language Processing (BioNLP 2024), pp. 440–448. Associ- ation for Computational Linguistics (2024)

    2024

  4. [4]

    Nature Communications16, 7866 (2025) 10 Rui Hao et al

    Wu, C., Zhang, X., Zhang, Y., Hui, H., Wang, Y., Xie, W.: Towards generalist foun- dation model for radiology by leveraging web-scale 2D&3D medical data. Nature Communications16, 7866 (2025) 10 Rui Hao et al

    2025

  5. [5]

    In: International Conference on Learning Representations (ICLR) (Poster) (2024)

    Lee, S., Kim, W.J., Chang, J., Ye, J.C.: LLM-CXR: Instruction-finetuned LLM for CXR image understanding and generation. In: International Conference on Learning Representations (ICLR) (Poster) (2024). OpenReview.net.https:// openreview.net/forum?id=BqHaLnans2

    2024

  6. [6]

    In: Proceedings of the 2025 International Conference on Multimedia Retrieval (ICMR), pp

    Jing, L., Zhang, Y., Du, X.: Tutorial proposal: Hallucinations in large language models and large vision-language models. In: Proceedings of the 2025 International Conference on Multimedia Retrieval (ICMR), pp. 2138–2139 (2025)

    2025

  7. [7]

    Artificial Intelligence Review57(9) (2024)

    Lin, Z., Guan, S., Zhang, W., Zhang, H., Li, Y., Zhang, H.: Towards trustwor- thy LLMs: A review on debiasing and dehallucinating in large language models. Artificial Intelligence Review57(9) (2024)

    2024

  8. [8]

    International Journal of Computer Vision134(1) (2025)

    Tu, C., Ye, P., Zhou, D., Bai, L., Yu, G., Chen, T., Ouyang, W.: Attention real- location: Towards zero-cost and controllable hallucination mitigation of MLLMs. International Journal of Computer Vision134(1) (2025)

    2025

  9. [9]

    arXiv:2503.02157 (2025)

    Chang, A., Huang, L., Bhatia, P., Kass-Hout, T., Ma, F., Xiao, C.: MedHEval: Benchmarking hallucinations and mitigation strategies in medical large vision- language models. arXiv:2503.02157 (2025)

    work page arXiv 2025

  10. [10]

    In: Findings of the Association for Computational Linguistics: ACL 2025, pp

    Zhu, Z., Zhang, Y., Zhuang, X., Zhang, F., Wan, Z., Chen, Y., Long, Q., Zheng, Y., Wu, X.: Can we trust AI doctors? A survey of medical hallucination in large language and large vision-language models. In: Findings of the Association for Computational Linguistics: ACL 2025, pp. 6748–6769. Association for Computa- tional Linguistics (2025)

    2025

  11. [11]

    npj Digital Medicine8, 274 (2025)

    Asgari, E., Montaña-Brown, N., Dubois, M., et al.: A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine8, 274 (2025)

    2025

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13872–13882 (2024)

    2024

  13. [13]

    DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models

    Chuang, Y.-S., Xie, Y., Luo, H., Kim, Y., Glass, J., He, P.: DoLa: Decoding by contrasting layers improves factuality in large language models. arXiv:2309.03883 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Huang, Q., Dong, X., Zhang, P., Wang, B., He, C., Wang, J., Lin, D., Zhang, W., Yu, N.: OPERA: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13418–13427 (2024)

    2024

  15. [15]

    arXiv:2405.17820 (2024)

    Woo, S., Kim, D., Jang, J., Choi, Y., Kim, C.: Don’t miss the forest for the trees: Attentional vision calibration for large vision language models. arXiv:2405.17820 (2024)

    work page arXiv 2024

  16. [16]

    In: Proceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), pp

    Gong, X., Ming, T., Wang, X., Wei, Z.: DAMRO: Dive into the attention mecha- nism of LVLM to reduce object hallucination. In: Proceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), pp. 7696– 7712 (2024)

    2024

  17. [17]

    In: European Conference on Com- puter Vision (ECCV), pp

    Liu, S., Zheng, K., Chen, W.: Paying more attention to image: A training-free method for alleviating hallucination in LVLMs. In: European Conference on Com- puter Vision (ECCV), pp. 125–140. Springer (2025)

    2025

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Favero, A., Zancato, L., Trager, M., Choudhary, S., Perera, P., Achille, A., Swami- nathan, A., Soatto, S.: Multi-modal hallucination control by visual information grounding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14303–14312 (2024) Grounding Medical MLLMs with Verifiable Anatomical Evidence 11

    2024

  19. [19]

    IEEE, pp

    Liu,B.,Zhan,L.-M.,Xu,L.,Ma,L.,Yang,Y.,Wu,X.-M.:SLAKE:Asemantically- labeledknowledge-enhanceddatasetformedicalvisualquestionanswering.In:2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). IEEE, pp. 1650–1654 (2021)

    2021

  20. [20]

    Scientific Data 5(1), 1–10 (2018)

    Lau, J.J., Gayen, S., Ben Abacha, A., Demner-Fushman, D.: A dataset of clinically generated visual questions and answers about radiology images. Scientific Data 5(1), 1–10 (2018)

    2018

  21. [21]

    Journal of the American Medical Informatics Association23(2), 304–310 (2016)

    Demner-Fushman, D., Kohli, M.D., Rosenman, M.B., Shooshan, S.E., Rodriguez, L., Antani, S., Thoma, G.R., McDonald, C.J.: Preparing a collection of radiol- ogy examinations for distribution and retrieval. Journal of the American Medical Informatics Association23(2), 304–310 (2016)

    2016

  22. [22]

    Scientific Data6(1), 317 (2019)

    Johnson, A.E.W., Pollard, T.J., Berkowitz, S.J., Greenbaum, N.R., Lungren, M.P., Deng, C.-Y., Mark, R.G., Horng, S.: MIMIC-CXR, a de-identified publicly avail- able database of chest radiographs with free-text reports. Scientific Data6(1), 317 (2019)

    2019

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 26296–26306 (2024)

    2024

  24. [24]

    In: Advances in Neural Information Processing Systems, vol

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J.-W., Naumann, T., Poon, H., Gao, J.: LLaVA-Med: Training a large language-and-vision assistant for biomedicine in one day. In: Advances in Neural Information Processing Systems, vol. 36, pp. 28541–28564. Curran Associates, Inc. (2023)

    2023

  25. [25]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-VL technical report. arXiv preprint arXiv:2511.21631 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: InternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025