acceptodds
Under review as a conference paper at ICLR 2027

Steerable but Not Readable: Fixation Heads in Medical Vision-Language Models

Abstract

Medical vision-language models (VLMs) write fluent findings but do not show which image region supports each sentence, so a clinician cannot check a claim against the scan. The usual remedy reads an attention or attribution map over the scan as a record of what the model used. Generalist VLMs contain a few attention heads that follow the region being described, yet it is unknown whether medical fine-tuning keeps these heads and whether their attention can serve as provenance. We test these fixation heads in medical VLMs in both directions, reading their attention as evidence and writing a region into it. Our evaluation across three model families and six corpora shows that (i) every model has fixation heads and supervised medical tunes inherit them from their base, (ii) the heads are steerable, as aiming them at a region changes what the model says, and (iii) they are not readable, as their attention barely follows the question, leaves the organ as the model answers and locates findings no better than random boxes, as do standard attribution maps. Anatomy-conditioned generation turns the write into provenance without training, writing one anatomical region into the heads per sentence so that the sentence depends on it, and adds clinical findings to chest reports. MedGaze, a low-rank adapter on the heads' query rows, learns the write, moving the heads onto the anatomy with no region at test and answering organ questions more accurately than training-free decoders. Our findings emphasize the need to write provenance during decoding rather than read it from attention maps.

Primary area:
interpretability and explainable AI
Keywords:
Attention Heads, Medical Vision Language Models, Visual Grounding, Attention Steering, Interpretability, Radiology report generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_fXVrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether fixation heads in medical vision-language models are reliable explanations of image use, or controllable components of generation. It follows their inheritance through medical tuning, evaluates several attention-based read-outs, and then intervenes on the same heads. The proposed ACG generates responses under explicit regional steering, and MedGaze learns anatomical attention through a decoding-only adapter on selected query rows. Experiments span 25 models and six corpora, with different subsets for individual evaluations. Results show that native attention is unreliable under the tested read-out protocols, steering changes regional descriptions and improves finding F1 in a controlled report-tour setting, and MedGaze improves organ-question accuracy and anatomical attention without test-time masks.

Critical strengths
  1. Steerability and readability are separated in the same components. Sections 5.2–5.3 intervene on and explain the same fixation heads. Steering changes generated descriptions, while native attention localizes poorly and detects its own errors poorly, and the temporal analysis shows attention shifting substantially between reasoning and answering.

  2. The controls are more than decorative. Paired base-to-tune comparisons, held-out montage transfer and the random-head controls each close off a different alternative explanation, and the answer-only adapter separates anatomical alignment from ordinary supervised task learning. The identical-prompt report tours are the cleanest of these, since they rule out the reading that the model was simply asked about anatomy.

  3. The inheritance analysis provides insight into what medical adaptation preserves. Section 5.1 and Figure 8 examine corresponding heads in base models and medical tunes, rather than comparing only downstream accuracy. All four supervised Qwen tunes retain their base’s top fixation head, and fixation matrices remain strongly correlated across the paired checkpoints. This provides a concrete account of how spatial-routing behavior survives adaptation and offers a useful starting point for studying how different tuning procedures modify model internals.

  4. Table 6 evaluates several attention- and gradient-based methods against random boxes matched to each method’s proposal sizes. This matters because large proposals can obtain favorable overlap without identifying informative evidence. Comparing each method with its own geometric baseline makes the localization audit more convincing than comparing heatmaps visually or using a single unmatched random baseline. The result is a useful evaluation procedure beyond the fixation-head method itself.

Critical weaknesses
  1. Table 9 shows that organ attention rises from 0.092 to 0.333 while visual reliance moves from 0.35 to 0.23, and the change is not significant after correction. The counterfactual variant is more convincing about regional dependence, but it optimizes a different objective, so it cannot be used to rescue the default model's interpretation. I am not convinced the default MedGaze learns content-dependent provenance. The same gap opens on the ACG side, where the masking experiments establish dependence on a steered region more directly than they support individual clinical claims. Learned attention alignment and evidence-faithful generation are different contributions, and the paper currently claims both.

  2. The localization evaluation lets candidate regions do part of the work. Table 7 reports SLAKE organ masks at a hit rate of 0.919 against a stated chance of 0.880, and those masks share their source with the evaluation boxes. A 0.039 margin is narrow enough that the definition of the chance level carries the result, and I could not find that definition in the table or its caption. The report-tour evaluation also uses dataset-provided anatomical boxes. Without a source-matched selection baseline, and without separating annotation-assisted from automatic-region results, a high localization score remains ambiguous between good selection and good candidates.

  3. The aggregate VQA improvement masks substantial differences between question types. Table 10 reports a much larger accuracy gain on yes/no questions than on organ-choice questions, while Appendix D.5 states that approximately 85% of the yes/no answers are positive. The improved MCC indicates that the gain is not simply a constant-answer effect, but response calibration and learned answer priors remain plausible contributors.

Questions for authors
  1. What supports the provenance reading of default MedGaze? Beyond anatomical attention alignment, what evidence shows the default model learns content-dependent provenance?

  2. How does steering-based selection compare against a source-matched baseline? Run random selection and a simple prior over the same candidate masks ACG uses, then report IoU and results stratified by candidate count, especially for images with multiple eligible regions.

  3. Which reported gains survive an automatic segmenter?

3 comments
AuthorsOpenReview ↗

Response (1/1): Provenance via ACG, source-matched selection and the yes/no prior

We thank the reviewer for the careful reading and for the positive assessment of our method and findings. Here are the responses to the minor concerns:

W1 / Q1. Provenance and default MedGaze. We agree that learned alignment and evidence-faithful generation are different contributions, and the revision assigns them explicitly (Sec 3.4, Sec 5.4, and Sec 5.5).

  • ACG carries the provenance claim: a region per sentence, a causal test that the sentence depends on it, and now a per-pair dependence score (positive median in 8 of 8 models).
  • Default MedGaze carries the alignment claim, and new results show what alignment buys in behaviour. With a corner cue that predicts the answer in training (14 models), MedGaze shrinks the learned shortcut in all 11 models that adopt it (30.7 to 12.7 points), while a control without the attention term keeps it (new Figure 7a). It also protects reliance under adaptation: 0.23 nats against 0.18 for answer-only, 12 of 14 models, frozen 0.35.
  • Provenance language for MedGaze is reserved for the reliance variant, whose answers depend on the organ (above frozen in 13 of 14 models).

W2 / Q2. Selection vs candidates. Chance is the mean over images of one over the number of candidate regions, which is the expected hit rate of random selection; the revised Table 7 caption defines it. About 70% of SLAKE images have a single eligible organ, which compresses the overall margin; on images with several candidates the fixation heads reach 61.7 against 39.4 chance. SLAKE's masks are the corpus annotation (the source of the evaluation boxes) and the chest masks are automatic; Table 7 now marks which is which. Candidates are whole organ masks, so a hit is the case IoU = 1 with the target. The source-matched comparison: same candidates and selection rule, written into the fixation heads or into 100 random heads (mean of 3 draws), 8 models, hit rate in %:

imagescandidatesfixation headsrandom headsfix. > rand.largestrandom selection
SLAKE267.854.67/883.650.0
SLAKE361.845.87/8 (p=0.016p=0.016)42.133.3
SLAKE4+48.338.45/882.221.2
SLAKEmore than one61.748.77/8 (p=0.016p=0.016)74.639.4
VinDr-CXR, automatic11 to 1417.28.17/8 (p=0.016p=0.016)10.37.2

The write is head-specific: the fixation heads beat random heads at every candidate count and random selection everywhere, and on automatic chest candidates they also beat the size prior (17.2 vs 10.3). The size prior is blind to the sentence. It scores high on SLAKE only because the asked organ is the largest candidate in 74.6% of multi-candidate images, and it would give every sentence the same organ, whereas the write follows each sentence to its own region. In ACG that region comes from the tour or the user, one per sentence.

Q3. Gains with an automatic segmenter. Every headline gain except the SLAKE blanking tests has an automatic-region or region-free version. SLAKE ACG with automatic SAM masks reaches 2.3× chance (0.210), and chest ACG uses an automatic segmenter, reaching 1.8× chance on MS-CXR and 2.0× on VinDr-CXR (25 models). With segmenter masks instead of Chest ImaGenome boxes, the steered report tour adds +6.4 F1 over its heads-off tour (Holm p=0.035p=0.035), on par with the box tour (Holm p=0.16p=0.16 for the difference). MedGaze takes no region at test.

W3. Yes/no gains and the answer prior. On a label-balanced set (424 held-out presence questions, half "yes", 14 models):

balanced acc."yes" rateMCCAUROC
frozen80.450.10.6160.889
answer-only81.661.70.6530.911
MedGaze82.061.00.6590.913

Both adapters move the "yes" rate toward the training distribution, and discrimination improves beyond that prior: MedGaze's AUROC, which does not depend on the decision threshold, rises over frozen in 13 of 14 models (p=0.0004p=0.0004), and its MCC in 12 of 14 (p=0.002p=0.002; Holm 0.012). The prior is shared by both adapters, so it does not enter the comparisons that isolate the attention term.

Changes in the revision. §Sec 3.4, Sec 5.4 and Sec 5.5 assign provenance to ACG and alignment to MedGaze; Table 7 defines chance and marks region sources; the appendix adds the source-matched selection table and the balanced yes/no table; new Figure 7 shows the shortcut result. The method, the main results and the conclusions remain unchanged. All changes are highlighted in yellow.

We are grateful for the constructive feedbacks. We hope these results and discussions address your concerns, and we would be glad to answer any further questions.

Reviewer_fXVrOpenReview ↗

Official Comment by Reviewer fXVr

I've read the response and revision. My first weakness is addressed by assigning provenance to ACG and alignment to MedGaze. The other two are also addressed in substance, so I will not repeat them.

  1. The abstract states that MedGaze matches the answer-only adapter in accuracy while outperforming training-free decoders. Table 1 suggests that the accuracy gains are largely shared with the answer-only adapter. The paper should make clearer what MedGaze adds beyond answer-only adaptation, especially when presenting accuracy gains.
  2. Section 3.4 defines provenance in terms of dependence on a region and separates it from clinical correctness. This distinction should be maintained throughout the paper, including when describing a region as supporting a sentence. The 74.9% result in Appendix G.5 measures relative dependence for reference phrases under steering; it does not establish that the region clinically supports each generated sentence.
  3. G.9 finds that masking the cells with the highest fixation-head attention affects phrase probability no more than masking an equal-area random region. Section 5.5 finds that training the heads’ query rows with an attention term reduces reliance on a learned shortcut. These results concern different settings, but their relationship deserves discussion. One possibility is that steering or training creates a dependence that was not present in the unmodified model, rather than revealing an existing one. The paper should clarify whether this is the intended interpretation.

The work remains substantial in scale and includes meaningful controls. The shortcut experiment, in particular, provides a strong basis for the paper’s central story. The paper would benefit from a clearer account of how the control channel, provenance, and accuracy results contribute to that story, and which benefits are specific to MedGaze.

AuthorsOpenReview ↗

Follow-up Response (1/1): What MedGaze adds, dependence vs correctness, and how the results fit together

Thank you for reading the response and the revision, and for the updated assessment. We are glad that the three weaknesses are addressed and that the shortcut experiment gives the central story a strong basis. Your three suggestions are now in the revised PDF, highlighted in green.

1. What MedGaze adds beyond answer-only adaptation. The accuracy gain over the frozen model is shared with the answer-only adapter, and §5.5 now says so where the gain is presented. At that accuracy, four benefits are specific to MedGaze:

  • Attention. The heads move onto the organ: 0.33 against 0.09 for the answer-only adapter, in 14 of 14 models.
  • Robustness. The learned shortcut shrinks from 30.7 to 12.7 points, with higher accuracy on uncued images (84.5% vs 78.5%), while a control without the attention term keeps the shortcut.
  • Visual reliance. It stays higher under adaptation: 0.23 vs 0.18 nats, 12 of 14 models (Holm p=0.02p=0.02).
  • Reports. They gain findings: 27.2 vs 25.4 finding F1, 9 of 10 models (p=0.010p=0.010).

In the revision: Section 5.5 (new sentence) and Appendix G.10.

2. Dependence and clinical correctness. We agree, and this is the paper's definition: a region is the provenance of a sentence when the sentence depends on it, and clinical correctness is a separate check by a radiologist or a labeller (§3.4). Appendix G.5 now states this for the 74.9% result: it scores dependence under steering on reference phrases, and the clinical correctness of generated sentences is measured separately, by the finding F1 of the reports (§5.6).

In the revision: Appendix G.5 (added a new sentence) and Appendix G.10.

3. How G.9 and the shortcut result relate. Yes, that is the intended interpretation, and it is the paper's central claim: the write creates a dependence that the unmodified model does not have, rather than revealing an existing one. Three results show it.

  • Unmodified model. The cells the heads attend are used no more than a random region (G.9), and blanking a tour box moves report F1 only from 21.9 to 17.3.
  • Region written by steering. With the same box written into the heads, blanking it lowers F1 from 30.6 to 16.7, which is 9.2 points more than without steering (p=0.012p=0.012, 8 of 9 models).
  • Region learned by training. The attention term acts on the same channel. It lowers the heads' attention on the predictive cue from 0.068 to 0.016 and the shortcut gap from 30.7 to 12.7 points, while the control without it keeps both.

G.9 and the shortcut result are therefore two sides of one finding. The heads' attention does not report what the model uses, and setting it, by steering or by training, changes what the model uses. This is the sense of the title: the heads are steerable but not readable.

In the revision: Appendix G.10, which gives this account in one place: how the control, provenance and accuracy results fit together, and which benefits are specific to MedGaze.

Thank you again for your close reading and your thoughtful engagement throughout the discussion. Your suggestions have made the paper's account of its own results clearer and its claims more precise, and we sincerely appreciate the time and care you have given to this work. We hope that these responses and the revision address your suggestions to your full satisfaction.

Regards.

Reviewer_o3nBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies ''fixation heads'', attention heads in the LLM of a VLM that follow the image region named in the query, in medical VLMs. Following the discovery score and steering intervention of Gandikota & Bau (2026), it evaluates 25 VLMs from three families (including 14 medical tunes paired with their bases) on six corpora and reports that (i) every model has fixation heads and supervised medical tunes inherit them; (ii) biasing their attention logits towards a region changes what the model says; and (iii) their attention cannot be read as evidence, as it barely follows the question, leaves the organ at the answer token, and proposes boxes no better than random, like standard attribution maps. It then proposes anatomy-conditioned generation (ACG), which steers the heads to one anatomical region per response to attach a region to each sentence, and MedGaze, a LoRA on the fixation heads' query rows trained with an attention loss on SLAKE organ boxes, which raises the heads' attention on the organ (0.09 → 0.33) and organ VQA accuracy (75.1 → 84.8%) with no region at test.

Critical strengths

The study is large and carefully controlled: 25 models with paired base–tune comparisons, random-head and all-head controls, size-matched random boxes, an other-patient counterfactual, and Holm-corrected tests with one model as one observation. The distinction between reading attention as evidence and writing a region into it is a useful framing for medical interpretability, and the paper reports results that do not favour its methods (e.g., the answer-only adapter, the reliance variant).

Critical weaknesses

(1) The methodological novelty is limited.

The discovery score, the top-100 head set, the random-head and all-head controls and the attention-mask steering follow Gandikota & Bau (2026), as the paper states, and ACG, which generates one response per steered region, is a direct use of this steering. Supervising attention with boxes and adapting selected heads are also existing ideas (Selvaraju et al., 2019; Luo et al., 2025, both cited). The new contributions are therefore the medical analysis (inheritance across tunes and the reading tests) and the specific MedGaze recipe, although for medical MLLMs [1] already showed that they often fail to ground their predictions in clinically relevant regions. The paper should discuss how its reading tests and its writing methods relate to and extend these findings.

(2) MedGaze's accuracy gain does not come from moving the heads.

  • An answer-only adapter on the same heads and split is as accurate as MedGaze (84.7 vs. 83.9, MCC 0.61 vs. 0.58, Table 1A; 82.1 vs. 81.8 on the second split, Table 1C). The attention term moves the heads but adds no accuracy.
  • MedGaze is trained and evaluated only on SLAKE organ questions (held-out images from the same corpus). Its only quantitative evaluation beyond SLAKE, report F1 on Chest ImaGenome, is not significant after correction (+2.8 points, Holm p = 0.16), and it is not tested on other VQA benchmarks already used in the paper (e.g., VQA-RAD, PathVQA) or other question types, so it is unclear whether the gain, or the attention shift, transfers beyond the training distribution.
  • MedGaze is trained on the SLAKE training split and tested on held-out SLAKE images, whereas VCD, PAI and CRG are training-free, so ''more accurate than training-free decoders'' (abstract) mainly reflects in-domain fine-tuning.

(3) The ''not readable'' conclusion is broader than the evidence. The 80× contrast between writing and asking is by construction: a ±4 logit bias on 100 heads puts about 0.97 of their attention in the region (Table 4), so comparing it with the effect of a question says little about readability.

Reference

[1] How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images. (Liu et al., ICLR 2026)

Questions for authors

Please refer to the Weaknesses.

7 comments
AuthorsOpenReview ↗

Response (1/2): Distinction of the proposed method and what the attention term adds

We thank the reviewer for a close reading and appreciating the usefulness of our work for interpretable medical image analysis. We address each of the concerns as follows:

1. Novelty, and relation to Liu et al. (2026) [1]. We build on Gandikota & Bau (2026) for head discovery and steering, as stated in the paper. Our paper adds:

  • Read and write tested on the same heads. The heads that causally steer the output are tested as a read-out with five tests and size-matched random boxes.
  • Inheritance. Supervised medical tunes keep their base's fixation heads; all four supervised Qwen tunes keep the base's top head.
  • ACG. A region per sentence, with a causal test that the sentence depends on it, now scored per pair.
  • MedGaze. A decoding-only adapter on the heads' query rows that leaves the KV cache untouched. Its attention term removes most of a shortcut that plain adaptation learns (point 2a).

HINT (Selvaraju et al., 2019) aligns gradient importance with human attention in VQA models. Selective LoRA (Luo et al., 2025) adapts selected heads and visual tokens without attention supervision. MedGaze supervises the attention of discovered heads inside a generative VLM, and we test what that supervision changes in the answer.

Liu et al. is complementary, and Section 2 now cites it. Their VGRefine selects heads whose attention aligns with relevant regions and masks attention with a map from the model's own attention, training-free, improving Med-VQA accuracy. Their finding that medical MLLMs often fail to ground agrees with ours. The questions differ: they use attention as a signal to refine answers; we ask whether a head's attention shows which region a given claim used. For the heads that causally steer the output it does not, so we write the region and test dependence causally. An attention-derived mask can help accuracy as a prior without being provenance for a sentence, so the two results are consistent.

These distinctions clarify where our method adds value: it targets how visually grounded a medical VLM’s responses are. We believe this analysis could also provide a starting point for building interleaved reasoning models with finer-grained multimodal alignment at lower cost.

2a. What the attention term adds (NEW Exp, 14 models). The two adapters are equally accurate (Table 1). We tested what the attention term changes in behaviour, using RRR's decoy protocol. All adapters are trained on the same SLAKE items, seed and steps, with a corner marker on 90% of "yes" and 10% of "no" training items, placed away from the asked organ. The shortcut gap is accuracy with the marker on "yes" answers minus accuracy with it on "no" answers.

cue p=0.9p=0.9answer-onlyno attention termMedGaze
gap, 11 models that adopt the cue30.730.012.7
gap, all 14 models24.424.310.1
accuracy on uncued images, after cued training78.578.584.5

MedGaze shrinks the gap in all 11 adopting models (p=0.001p=0.001). A control with every MedGaze term except the attention term keeps the shortcut (30.0 vs 30.7), so the attention term accounts for 96% of the reduction (99% over all 14). On anti-cue items MedGaze is significantly more accurate than answer-only in 9 of the 11. The loss moves the heads off the cue; the answer then largely ignores a cue that every other head and MLP still sees, which the control shows does not happen without the attention term. On chest reports, answer-only gains nothing over frozen (+1.0, p=0.77p=0.77), while MedGaze is above it in 9 of 10 models (p=0.010p=0.010). MedGaze's report gain over frozen keeps its sign under all three labellers (+2.8 to +3.5 points, 7 or 8 of 10 models).

References: [1] How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images. (Liu et al., ICLR 2026)

AuthorsOpenReview ↗

Response (2/2): Transfer, training-free decoders and the reading tests

Continuation of Response (1/2): Distinction of the proposed method and what the attention term adds.

2b. Beyond the training distribution. MedGaze's training mixes organ questions with a box, which carry the attention term, with other closed SLAKE questions, which carry the answer loss only. To test beyond the training images and corpus, we scored both adapters with no region on unseen SLAKE images and, zero-shot, on VQA-RAD:

13 modelsanswer-only vs frozenMedGaze vs frozenMedGaze vs answer-only
SLAKE, all closed types, new images+2.3 (11/13, p=0.001p=0.001)+2.3 (11/13, p=0.008p=0.008)p=0.89p=0.89
VQA-RAD yes/no (251 questions)median −0.4, n.s. (p=0.49p=0.49)median −0.4, n.s. (p=0.30p=0.30)n.s. (p=0.31p=0.31)

On unseen SLAKE images the gain holds pooled over closed questions of every type (modality, position, abnormality and others), equally for both adapters. On VQA-RAD, a corpus never seen in training, neither adapter changes accuracy significantly, and MedGaze matches answer-only (p=0.31p=0.31). MedGemma is reported separately: frozen, it rarely answers VQA-RAD in yes/no format (32% accuracy), and either adapter adds about 35 points.

VQA-RAD has no organ boxes, so the attention shift cannot be scored there. We used VQA-RAD as the radiology transfer set because PathVQA's pathology slides have no anatomical organs to supervise. Within scope, the gain over frozen also holds on a second image split (Table 1C) and with heads ranked on training images only (7 of 7 models).

2c. Training-free decoders. We agree to the reviewers comment. So, the like-for-like comparison is the answer-only adapter, which Table 1 includes. The decoder rows (VCD, PAI, CRG with the mask) and fixation steering change accuracy by at most +2.4 points, against +13.4 for MedGaze. CRG is the closest row to VGRefine, since it redirects attention with the annotated organ mask itself. The abstract now states the answer-only comparison and the shortcut result.

3. "Not readable" and the 80× contrast. The 80× ratio shows the scale of the write at b=4b=4, and its size grows with the bias; the conclusion does not rest on it. "Readable" means the map marks the region the answer used, better than a size-matched random region. The conclusion rests on the paper's five reading tests, none of which applies a bias:

  1. attention barely follows the question;
  2. attention leaves the organ as the model answers (8.5% to 4.3% against 6.0% chance, in all 7 models of the trace study);
  3. in the unsteered model, masking the radiologist's box leaves the phrase's likelihood unchanged, while another patient's image lowers it (17 of 17);
  4. box proposals are no better than size-matched random boxes (0.154 vs 0.173), like every attribution map in Table 6;
  5. the heads flag errors no better than random heads.

Test 3 is the core: the map sits on a region that the unsteered answer does not use, while after the write, blanking that region changes the output, which is what ACG tests. A new test (17 models) confirms these are the right heads: a question moves their attention 5.5 times more than random heads' (13 of 17, Holm p=0.009p=0.009). That is 0.67 points of attention, too small for a map to single out the evidence (test 4).

Changes in the revision. Section 2 cites Liu et al., RRR, GAIN and ISNet; the abstract and Section 5 add the answer-only comparison and the shortcut result (new Figure 7); the appendix places the 80× ratio in the context of the bias; the appendix adds the shortcut experiment, transfer and the question-effect test. All changes are highlighted in yellow.

We thank the reviewer again for taking time and giving us useful feedback. We hope these results address your concerns, and we would be glad to answer any further questions.

Thank you for the response and the new experiments. The decoy experiment gives the attention term a clear role, and I appreciate the candid transfer results and the added discussion of Liu et al. Three points remain:

  1. Reading tests. The response states that the map sits on a region the unsteered answer does not use. However, test 3 masks the radiologist's box, and test 4 shows that the maps do not lie on that box, so neither tests whether the attended region is used. Masking the region the heads attend to, against a size-matched random region, would test this directly and is the main result that would change my assessment. The new question-effect test (5.5× random heads) also suggests the heads do follow the question, although weakly.

  2. MedGaze. On VQA-RAD neither adapter improves accuracy, and on unseen SLAKE images MedGaze equals the answer-only adapter, so the accuracy claim should be scoped to in-distribution organ questions, with the shortcut result as the contribution of the attention term. The decrease in visual reliance after MedGaze (0.35 → 0.23, Table 9) is still not explained.

  3. Weakness 4 (ACG) was not addressed: the 64% of steered sentences naming the organ has no unsteered baseline, the region selected from SLAKE organ masks is at chance (0.919 vs. 0.880), and the report gain is measured against the unsteered tour rather than the free prompt.

Reviewer_TeLKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors studies fixation heads in medical vision-language models and proposes methods to use them for anatomical attention control and improved prediction. Across 25 models, the authors examine how these heads are preserved after medical adaptation and show that they can steer generation toward anatomical regions, while their unmodified attention maps do not reliably localize supporting evidence under the evaluated protocols. Based on this analysis, anatomy-conditioned generation directs attention to specified regions during decoding, and MedGaze uses a LoRA adapter on selected query-projection rows to learn anatomical attention without region masks at inference. The paper reports improvements in report finding F1 with anatomy-conditioned generation and organ-question accuracy with MedGaze. Combining the head-level analysis with practical adaptation methods is a useful contribution, although the similar accuracy of the answer-only adapter leaves the additional performance benefit of attention supervision unclear.

Critical strengths
  1. Varied, controlled evaluation with practical benefits. Report generation and organ-question answering are complemented by tests of attention during reasoning, region-masking sensitivity, and random-head steering. ACG improves report finding F1 from 19.7 to 27.6 against a matched unsteered tour, while MedGaze raises organ-question accuracy from 75.1% to 84.8% over frozen models. These experiments connect the head analysis to measurable task improvements.

  2. Lightweight anatomical attention adaptation. MedGaze updates only the query-projection rows of selected fixation heads through a small LoRA-style adapter, with the backbone frozen. It improves anatomical attention alignment without requiring region masks at inference, offering a targeted way to adapt attention while retaining accuracy comparable to answer-only adaptation.

Critical weaknesses
  1. The added benefit of attention supervision is unclear. MedGaze achieves similar accuracy to the answer-only adapter despite better attention alignment, and does not consistently improve measured visual reliance. This limits evidence that attention supervision improves task performance or evidence use beyond ordinary adaptation.

  2. The scope of provenance remains limited. Region-masking experiments support increased regional dependence, but the evaluation does not establish evidential support for individual clinical claims. The distinction should be clearer when interpreting the region sentence pairs.

Questions for authors
  1. Given that MedGaze improves attention alignment but achieves similar accuracy to the answer-only adapter, what future directions could translate this improved alignment into better task performance?

  2. MedGaze improves anatomical attention alignment but does not consistently improve measured visual reliance. How should this discrepancy be interpreted when assessing whether MedGaze improves the use of relevant image evidence?

  3. Given that the masking experiments measure organ naming and aggregate report finding scores, what level of evidential support does an ACG region-sentence pair provide for an individual clinical claim?

3 comments
AuthorsOpenReview ↗

Response (1/1): Alignment pays off in behaviour; reliance and per-pair evidence

We thank the reviewer for the positive assessment and the precise questions. We address all the concerns raised by the reviewer as follows:

Q1. Turning alignment into better task performance. It already pays off in behaviour. We trained answer-only and MedGaze adapters on SLAKE with a corner marker that predicts the answer 90% of the time, placed away from the asked organ (new, 14 models). Answer-only adopts the cue in 11 models; MedGaze shrinks the shortcut gap in all 11 (30.7 to 12.7 points, p=0.001p=0.001) and is more accurate on uncued images after cued training (84.5 vs 78.5). A control with every term except the attention term keeps the shortcut (24.3 vs 24.4), so the gain is the attention term's (new Figure 7a). On chest reports, MedGaze is above answer-only in 9 of 10 models, while answer-only gains nothing over frozen.

Two directions extend this. A counterfactual term, as in the reliance variant, makes the answer depend on the organ (above frozen in 13 of 14 models) while keeping most of the accuracy gain (82.0, against 75.1 frozen). Region supervision from report corpora with boxes (MS-CXR, Chest ImaGenome) would carry the objective to free text.

Q2. Alignment vs measured visual reliance. Moving where the heads look does not by itself make the answer depend on the pixels there; that is the paper's distinction between control and evidence. The attention term protects reliance under adaptation: answer-only drops reliance below frozen (0.35 to 0.18 nats, 12 of 14 models). MedGaze keeps more (0.23; above answer-only in 12 of 14, Holm p=0.02p=0.02), with no significant drop from frozen, and a counterfactual term raises it to 1.21 nats (above frozen in 13 of 14). The attention term thus moves MedGaze toward using the organ, and the counterfactual term makes that use explicit.

Q3. Evidential support of an ACG region-sentence pair. A pair attaches a region to a sentence and lets us test that the sentence depends on it; clinical correctness is a separate check. We now score dependence per pair: DD is the drop in a phrase's log-probability when its own region is blanked, minus the drop for an equal-area random region. On SLAKE phrases naming the steered organ, the median DD is positive in 8 of 8 models (Wilcoxon p=0.008p=0.008; 0.60 nats over 2257 pairs). Three in four pairs (74.9%) depend on their own region more than on an equal-area random one, and DD flags the rest, so a reader knows which pairs to trust. DD needs two extra forward passes and is defined for any sentence; here it is computed on reference phrases under steering. At report level, blanking the steered regions costs the steered tour 9.2 F1 points more than the unsteered tour.

Revision. Section 3.4 states the scope of a pair (dependence on the written region; correctness checked by a radiologist or labeller) and Section 5.4 ties the provenance claim to that dependence; the appendix adds the per-pair score and the shortcut experiment; new Figure 7 summarises the shortcut and report results. The method, the main results and the conclusions remain unchanged. All changes are highlighted in yellow.

We hope these results address your minor concerns, and we would be glad to answer any further questions.

Thank you to the authors for the detailed response. The new shortcut experiments, ablation study, and visual reliance results provide stronger evidence for the benefits of attention supervision and address my main concerns. I also appreciate the additional per-pair analysis and clarification regarding regional dependence and clinical correctness. However, the per-pair analysis measures dependence on anatomical regions rather than whether those regions provide clinically correct evidence for individual claims. Overall, the rebuttal strengthens the paper, and I will maintain my original rating of 3 (Weak Accept).

AuthorsOpenReview ↗

Thank you!

Thank you for reconsidering the paper and for the updated assessment. Your questions on what alignment buys in behaviour, on visual reliance and on the evidence behind each region and sentence pair led to the shortcut experiment, all of which are now in the revised PDF. They make the paper's claims more precise and its evidence more direct.

Regards.

Reviewer_hRALOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper analyses medical VLM attention, and proposes new methodologies to change this attention. It includes a steering methodology, which injects biases into attention, forcing a focus on a region of interest without retraining, and a fine-tuning methodology that teaches such attention focus.

Critical strengths
  1. Substantial model coverage: their experiments cover 25 VLMs from three families, 2B to 73B parameters.
  2. They include multiple datasets and statistical tests.
  3. The main metrics they use to evaluate report generation are correct: they focus on F1 score and other metrics that measure how well reports detect diseases. This is indeed what we want to measure in report generation, instead of focusing on LLM metrics like BLEU, which mainly measure if reports look like human-made reports, instead of measuring how well they detect diseases. I am very glad to see this.
  4. The supplementary material is comprehensive.
  5. The paper is clear.
Critical weaknesses
  1. In Tab 1, fixation steering seems to have minimal benefit over the frozen backbone in accuracy, and MedGaze does not surpass an answer-only adapter. This possibly indicates that MedGaze's better performance comes from fine-tuning alone, not from its new methodology. Therefore, I cannot conclude that the proposed methodologies improve performance. The attention concentration over the target regions improves, as expected, since this is a direct training objective in MedGaze. However, results (the small improvement over the Answer-only adapter) do not seem to support that this attention concentration is improving the model answers. Therefore, I wonder whether the model is truly learning to analyze the target regions better, or if it is only learning to satisfy the new L_att objective without improving performance. Measuring attention is a tool, not the final objective. A model can learn to change its attention without improving its behavior. The final objective is to steer attention to improve model behavior, performance, robustness to bias, or generalization. Unfortunately, Table 1 does not show this final objective was achieved.
  2. Fig 6b shows a strong improvement of F1 for the steering tour, +7.9. However, this improvement is much lower when compared to the ‘no-tour’ frozen model, +3.2% (table 11). Is this improvement significant? This small improvement is coherent with Table 1, which showed a small performance improvement with steering.
  3. Regarding the MedGaze model in table 11, we are missing the answer-only adapter baseline to understand if its improvement comes from the proposed methodology or from fine-tuning alone. Therefore, my conclusions in comment 1 remain the same after reading Fig. 6.
  4. Why is the Answer-only adapter not in part B in Table 1? And why does it change name to “answer loss only’ in part C?
  5. Related work is missing the broad literature on how attention was controlled in CNNs and transformers before medical VLMs. Multiple methods trained deep classifiers to pay attention to target regions in medical image classification, reducing shortcut learning. Examples: right for the right reasons (RRR, https://arxiv.org/pdf/1703.03717), guided attention inference networks (GAIN, https://arxiv.org/pdf/1802.10171), implicit segmentation network (ISNet, https://www.nature.com/articles/s41467-023-44371-z).
  6. The loss functions in Eq. 2 are in the core of this methodology, but they are only described in text, with little detail. Please include the equations and a more detailed description of each loss in the main text. Also, in the description of this equation you mention the “annotated box of the organ”. Are you using organ segmentation masks, organ bounding boxes, or both? Please explain clearly what are the region annotations used in training.
  7. “Report generation uses 199 held-out Chest ImaGenome studies per model, whose findings a rule-based labeller with negation handling scores against the study’s Chest ImaGenome findings as clinical finding F1.” What is the accuracy of the labeler? How was it evaluated? Other works use LLM labelers that could be more accurate.
  8. In Fig 6, please include the absolute F1-scores, not only relative improvements.
Questions for authors

Please see Critical Weaknesses. The most important comments are 1, 2, 3 and 4.

4 comments
AuthorsOpenReview ↗

Response (1/2): The attention term curbs a learned shortcut; answer-only reports

We thank the reviewer for the detailed and constructive review. We address each of the comment here:

We agree with its central point: attention is a means, and the attention term should be judged by whether it improves behaviour, including robustness to bias. We ran new experiments on all 14 models that test exactly this, and added an answer-only adapter to every comparison.

W1. Does the attention term improve behaviour? Yes. At equal accuracy or more, it makes the model far more robust to a shortcut that plain adaptation learns. On held-out SLAKE questions both adapters reach the same accuracy (Table 1, block B); what the attention term changes is what the answer relies on. We test this where attention supervision is known to matter, a spurious cue in the training data, as RRR and ISNet do.

Protocol. All adapters are trained on the same SLAKE items, seed, steps and fixation heads. A small corner marker, placed away from the asked organ, appears on 90% of "yes" training items and 10% of "no" items, so it predicts the answer. Held-out questions carry the marker on "yes" answers (aligned) or on "no" answers (anti); the shortcut gap, aligned minus anti accuracy, measures how far the answer follows the marker. A third adapter, the control, keeps every MedGaze term except the attention term, which isolates that term's contribution.

cue p=0.9p=0.9answer-onlyno attention termMedGaze
gap, 11 models that adopt the cue30.730.012.7
gap, all 14 models24.424.310.1
fixation-head attention on the cue0.0680.0690.016
accuracy on uncued images, after cued training78.578.584.5
  • The shortcut shrinks. MedGaze reduces the gap in all 11 models that adopt the cue (sign test p=0.001p=0.001) and is below the control in 14 of 14 (Wilcoxon p=0.001p=0.001).
  • The answer changes, not only the attention. On anti-cue items, MedGaze is significantly more accurate than answer-only in 9 of the 11 (per-model McNemar, Holm).
  • The attention term is responsible. The control keeps the shortcut (30.0 vs 30.7 in the 11 adopting models), so the attention term accounts for 96% of the reduction there (99% over all 14). MedGaze is also more accurate than the control on uncued images (10 of 14 models, p=0.007p=0.007).
  • A stronger cue. With a perfectly predictive cue (p=1.0p=1.0, 5 models), the gap falls from 52.9 to 24.9 points, smaller in 5 of 5.

In the revised manuscript new Figure 7a shows the shortcut result per model.

W2. Steered tour vs the free report. The comparison that isolates the heads is steered vs heads-off under the identical tour prompt, since the two differ only in the write. Steering adds +7.9 F1 points in 9 of 10 models (Holm p=0.023p=0.023), and the effect holds under all three labellers (W7), strongest under CheXbert (+8.7, 10 of 10). The comparison with the free report answers a different question, whether organising a report by region costs findings, and it does not: the steered tour scores above the free frozen report (27.6 vs 24.4; +3.1, 6 of 10, p=0.32p=0.32). ACG therefore gives every sentence a region while keeping the finding F1 of a free report. In Table 1, steering is a training-free control (+2.4 accuracy over frozen); the paper uses the write for generation (ACG) and learns it for answering (MedGaze).

W3. Answer-only adapter on reports (NEW Exp). We generated reports with the answer-only adapter for the same 10 models and 199 studies, with the same prompt and decoding; the frozen reports reproduce exactly, so every comparison is paired. Answer-only does not raise finding F1 over frozen (25.4 vs 24.4, +1.0 points, 5 of 10 models, p=0.77p=0.77). MedGaze reaches 27.2 and is above answer-only in 9 of 10 models (p=0.010p=0.010), and the LLM labeller agrees (9 of 10, p=0.004p=0.004). MedGaze's report gain therefore comes from its objective, not from adaptation alone. New Figure 7c shows absolute F1 per model, and §5.6 reports the comparison.

AuthorsOpenReview ↗

Response (2/2): Table 1 updates, related work, loss equations and labeller accuracy

Continuation of Response (1/2): The attention term curbs a learned shortcut; answer-only reports.

W4. Table 1. Thank you for the suggestion. Block B now includes the answer-only adapter for all 14 models. The two adapters reach the same accuracy (85.4 vs 84.8; no difference in accuracy, balanced accuracy or MCC after Holm correction), while only MedGaze moves attention onto the organ (0.33 vs 0.09, 14 of 14 models). This is the setting the shortcut experiment builds on: equal accuracy, different evidence. All blocks now use the name "answer-only adapter"; block C's former "answer loss only" row was the same recipe retrained on the second image split.

W5. Related work. Thank you, this is the right context. Section 2 now cites RRR, GAIN and ISNet, and Appendix G.8 discusses them: each supervises where a classifier looks, and each reduces shortcut learning. MedGaze brings this supervision to generative VLMs. It acts on the query rows of a small set of discovered fixation heads, at decoding only, with the backbone and prompt cache untouched. Our shortcut experiment (W1) follows RRR's decoy protocol, and its result agrees with that literature.

W6. Losses and region annotations. Let HH be the fixation heads, ah(t)a_h(t) head hh's attention to image token tt at the answer step, and BB the image tokens covered by the organ box (token cover ≥0.25\ge 0.25):

L=Lans+0.25 Latt+0.1 Lfid,Lans=−log⁡pθ(y∣x,q),L = L_{ans} + 0.25\,L_{att} + 0.1\,L_{fid}, \qquad L_{ans} = -\log p_\theta(y \mid x, q), Latt=−1∣H∣∑h∈Hlog⁡∑t∈Bah(t)∑tah(t),Lfid=KL(pfrozen(⋅∣x,q0) ∥ pθ(⋅∣x,q0)),L_{att} = -\frac{1}{|H|}\sum_{h\in H}\log\frac{\sum_{t\in B} a_h(t)}{\sum_t a_h(t)}, \qquad L_{fid} = \mathrm{KL}\big(p_{frozen}(\cdot\mid x,q_0)\,\|\,p_\theta(\cdot\mid x,q_0)\big),

where q0q_0 is a prompt that names no region and LfidL_{fid} is applied to every fourth training item. In words, LattL_{att} raises each fixation head's share of image attention inside the box, the same quantity that head discovery measures, and LfidL_{fid} keeps region-free answers close to the frozen model. Training uses SLAKE's organ bounding boxes on training images only. Organ masks serve only evaluation and the test-time baselines that take a region (steering, CRG). MedGaze takes no region at test. §3.5 now gives all three terms.

W7. Accuracy of the report labeller. We evaluated it against two independent labellers on every stored report: the official CheXbert checkpoint (the 10 findings it shares with ours) and an open LLM labeller (gemma-3-27b-it, given our 15 Chest ImaGenome findings). The two agree closely with each other (κ=0.82\kappa=0.82, 95% raw agreement). On the decisions where they agree (10 shared findings), our rule labeller recalls 0.82 of their positive findings (F1 0.67) and is the most inclusive of the three (precision 0.57). Because every comparison is paired on the same studies and the same labeller, a bias shared across conditions cancels. Every report effect under each labeller (F1 difference in points, 10 models, unadjusted Wilcoxon pp):

effectrule, 15LLM, 15rule, 10CheXbert, 10
steered vs heads-off tour+7.9 (9/10, .004)+7.3 (8/10, .010)+8.9 (9/10, .004)+8.7 (10/10, .002)
MedGaze vs frozen+2.8 (8/10, .049)+3.5 (7/10, .084)+3.2 (8/10, .049)+2.8 (8/10, .084)
MedGaze vs answer-only+1.8 (9/10, .010)+1.7 (9/10, .004)+2.0 (9/10, .014)+1.7 (7/10, .064)
steered tour vs free report+3.1 (6/10, .32)+5.3 (6/10, .19)+3.5 (6/10, .32)+7.6 (7/10, .11)

Every effect keeps its sign under every labeller, and the causal steering effect is strongest under CheXbert (+8.7, 10 of 10). Appendix G.2 reports the full study.

W8. Absolute F1. New Figure 7 shows absolute finding F1 per model, so the levels are visible alongside the differences: heads-off tour 19.7 vs steered tour 27.6 (panel b), and frozen 24.4, answer-only 25.4 and MedGaze 27.2 (panel c).

Changes in the rivision. Section 2 related work; Section 3.5 equations; the answer-only rows in Table 1 and the reports; Figure 7 (shortcut, absolute F1); the abstract adds the shortcut result; the appendix adds the shortcut experiment and the labeller study. The method, the main results and the conclusions remain unchanged. All changes are highlighted in yellow in the revised PDF.

We hope these results address your concerns, and we would be glad to answer any further questions.

AuthorsOpenReview ↗

Follow-up and summary of the rebuttal

Dear Reviewer,

Thank you again for your detailed review. As we are approaching the deadline for rebuttal and discussion, we would appreciate your feedbacks. Our response addresses every comment you raised, with new experiments on the four concerns you marked as most important:

  1. The attention term improves behaviour. With a spurious cue in training (14 models), MedGaze shrinks the shortcut gap from 30.7 to 12.7 points in all 11 models that adopt the cue, while a control with every term except the attention term keeps it (30.0). At equal accuracy, the attention term makes the model more robust to bias, which is the objective your review sets.
  2. Steering adds findings. Under the identical tour prompt, steering adds +7.9 F1 in 9 of 10 models (Holm p=0.023p=0.023), and the steered tour also scores above the free report (27.6 vs 24.4).
  3. MedGaze's report gain comes from its objective, not from fine-tuning alone. The answer-only adapter does not raise finding F1 over frozen (25.4 vs 24.4), whereas MedGaze reaches 27.2 and is above answer-only in 9 of 10 models (p=0.010p=0.010).
  4. Table 1 is complete in the revised version. Block B now includes the answer-only adapter for all 14 models, under one name in every block.

The revision also covers comments 5 to 8: the related work on attention supervision (RRR, GAIN, ISNet), the loss equations, the labeller study against CheXbert and an LLM labeller, and absolute F1 in Figure 7. We have updated the revised manuscript with all the changes while our method, main results and conclusion remains unaffected.

We believe these results resolve the soundness concerns. The attention term changes the model's behaviour, not only its attention, and MedGaze's robustness and report gains hold against fine-tuning alone. We hope the revised paper merits a higher assessment, and we would be glad to answer any further questions.

Regards.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer;

Thank you for your time and your constructive feedback on our paper. As the rebuttal period will end soon, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns or questions remain.

We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

Thank you again for your time.

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.