acceptodds
Under review as a conference paper at ICLR 2027

If Medical Concepts Are Still There, Why Do Med-LVLMs Get Them Wrong?

Abstract

Why do medical large vision–language models (Med-LVLMs) get medical concepts wrong when those concepts remain recoverable from their internal representations? We investigate this gap between concept availability and answer correctness by tracing medical concepts through the visual encoder, vision–language connector, and language decoder. Linear probes reveal no sustained loss of concept decodability, while representational analyses indicate that encoder features are preserved across the connector. The decoder, however, exhibits differences in how available information is used: incorrect answers show weaker attention to target regions, and correct and incorrect answers become increasingly separable in middle-layer MLP activations. Activation patching further identifies middle-to-late MLP components whose clean outputs restore support for concept presence after target-region masking. These findings suggest that errors can arise from a failure to use available medical information, rather than from its loss alone. Guided by this evidence, we introduce MedSteer, which uses concept-specific directions derived from original and target-masked images to edit decoder activations, with a presence probe gating each edit. Across three Med-LVLMs, MedSteer improves average VQA accuracy by – percentage points over the unedited models. Our results identify a gap between representing medical concepts and using them to answer correctly, and demonstrate that concept-guided decoder editing can help bridge this gap.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical vision–language models, interpretability, steering, activation patching, medical concepts

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DaNvOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper asks why medical LVLMs answer yes/no concept-presence questions wrongly when linear probes can still read the concept from hidden states. Tracing the encoder, connector and decoder of three models, the authors argue the decoder fails to use this information. They propose MedSteer, a probe-gated activation edit that raises accuracy by 5 to 13 points over unedited models and beats steering and decoding baselines.

Critical strengths
  1. The analysis traces concepts from the vision encoder to the decoder and corroborates correlational findings with activation patching, across three models and multiple modalities. The appendix is candid about the limitations of each test.

  2. MedSteer is lightweight: 64 positive examples per concept, two forward passes, no fine-tuning. It improves all model and modality combinations in Table 2, which matters in practice if the gains withstand the controls below.

Critical weaknesses
  1. The paper interprets high probe accuracy as evidence that the concept remains decodable in the decoder. However, since the probe is trained with explicit labels, high accuracy only demonstrates that a linear classifier can separate the two classes, not that the model itself retains or utilizes the concept.

  2. Table 3 is difficult to interpret. The paper does not specify which SLAKE questions are used (closed-set, open-set, or the authors’ own presence template), nor how the 155 CT and 60 MRI samples were selected. These sample sizes are also limited.

  3. All main results are derived from yes/no presence questions constructed from segmentation labels. The paper does not verify whether MedSteer can improve model performance on real downstream tasks such as the full SLAKE test set or full IU-Xray report generation.

Questions for authors
  1. The decoder probe accuracy reflects how well the probe can predict, from the decoder representation, whether a given layer still retains the concept. Why does high probe accuracy indicate that the concept is decodable?

  2. SLAKE contains both closed-set and open-set questions. Which subset is used in Table 3, how were the 155 CT and 60 MRI samples selected, and are they exclusively yes/no presence questions?

  3. MedSteer improves concept recognition on presence questions. Does this generalize to real downstream tasks? Results on the full SLAKE test set and full IU-Xray report generation task would substantially strengthen the case.

4 comments

We thank the reviewer for the constructive feedback and for highlighting the need to clarify our probe interpretation, evaluation protocol, and downstream applicability. We address each concern below.

>>> Response to W1 and Q1: We agree that high probe accuracy alone does not establish that the model uses a concept when generating an answer. In our paper, “decodable” means that information about the concept label is linearly recoverable from a frozen representation. As stated in Section 4, “Linear probes test whether a target property can be recovered by a simple classifier from frozen representations” (lines 132–133; Alain & Bengio, 2017; Belinkov, 2022).

The explicit labels supervise the linear readout, and they do not modify the frozen LVLM. High probe accuracy on held-out, patient-disjoint test data therefore indicates that the representation contains features predictive of concept presence that are accessible to a linear classifier. This is the operational meaning of “decodable” in our analysis. It does not, by itself, establish that these features constitute a faithful semantic representation of the concept or exclude reliance on correlated features.

We also explicitly distinguish recoverability from use: “probe performance measures whether concept information is recoverable from a representation by a linear readout. It does not establish that the model itself uses this information when generating an answer” (§4, lines 157–159). This distinction is the central premise of the paper and motivates the complementary analyses in Sections 6–7. Attention and activation analyses characterize associations with model behavior, while activation patching and steering test how interventions affect the model’s answers.

To avoid ambiguity, we will define “linearly decodable” explicitly at its first occurrence and replace ambiguous statements that a layer “retains the concept” with the more precise wording “contains linearly recoverable information about the concept label.”

Revision: The definition is now stated explicitly in Section 4 (page 3, lines 159–161), and the wording is revised in Section 1 (pages 1–2, lines 52–59).

>>> Response to W2 and Q2: Thank you for identifying this omission. Table 3 (Table 4 in the revised version) evaluates whether the kidney and liver directions learned from CT data transfer to new data. We therefore select SLAKE’s original English closed-ended questions asking whether a kidney or liver is present.

We include all eligible questions from the SLAKE English test set whose associated image is CT or MRI, with no additional filtering. This yields 155 CT and 60 MRI question–image pairs. All selected questions are yes/no presence questions from SLAKE itself, and they are not generated using our segmentation-derived templates.

We agree that this is a limited, targeted transfer evaluation. The sample sizes reflect the number of eligible questions in the test set. We have specified the question source, inclusion criteria, and unit of counting in the description of Table 3 (Table 4 in the revised version). To complement this evaluation, we also report results on the full SLAKE English test set in Table R7 below.

Revision: These details are now given in Section 7.5 (page 10, lines 509–512) and Appendix E.1 (page 32).

References

  • Alain, G., & Bengio, Y. (2017). Understanding intermediate layers using linear classifier probes. ICLR Workshop.
  • Belinkov, Y. (2022). Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1), 207–219.

>>> Response to W3 and Q3: Following the reviewer’s suggestion, we evaluated MedSteer on the full SLAKE English test set and the full IU-Xray report-generation test split. Because MedSteer operates on concepts with learned steering directions, we use the following protocols.

SLAKE. We apply the gated edit when a question mentions a concept in our modality-level concept pool and leave other questions unedited. Of the 1,061 test questions, 470 receive an edit, and 591 remain unedited. The reported metrics cover the full test set, including both groups. For closed-ended questions, we report accuracy. For open-ended questions, we report recall, defined as the fraction of ground-truth answer tokens that appear in the generated answer, averaged over questions.

Table R7. Full SLAKE English test set (n = 1,061; 470 questions edited). Scores are percentages, reported as unedited → MedSteer. We report accuracy for closed-ended questions and recall for open-ended questions.

ModelClosed-ended accuracyOpen-ended recall
MedSIGHT70.9 → 74.655.3 → 57.2
Lingshu84.6 → 86.375.2 → 76.1

MedSteer improves closed-ended accuracy by 3.7 and 1.7 percentage points and open-ended recall by 1.9 and 0.9 percentage points for MedSIGHT and Lingshu, respectively. These are full-test-set improvements under a protocol that targets the 470 concept-matched questions.

IU-Xray. We apply gated MedSteer edits for seven candidate chest X-ray finding concepts, including lung lesion and pleural effusion, to every test image. Each concept’s probe determines its edit sign for that image. We use the same generation setting as in Table 4 (Table 5 in the revised version) and evaluate the complete generated reports.

To assess report quality beyond target-concept accuracy, we report F1 over all 14 CheXpert findings for finding-level agreement, which is computed as micro-averaged F1 over positive labels. We also report ROUGE-L for lexical overlap, and RadGraph F1 for clinical entity and relation agreement.

Table R8. Full IU-Xray test split (n = 590 studies, one reference report per study), using the generation setting from Table 4 (Table 5 in the revised version). Scores are on a 0–100 scale, reported as unedited → MedSteer.

ModelF1 over 14 findingsROUGE-LRadGraph F1
MedSIGHT14.7 → 15.813.7 → 15.216.6 → 20.2
Lingshu27.5 → 28.628.8 → 30.032.0 → 32.5

MedSteer improves all three report-generation metrics for both models. Together with the full SLAKE results, these findings provide evidence that its benefits extend beyond templated presence questions to the evaluated downstream VQA and report-generation tasks. The gains vary across models and metrics, and the intervention remains limited to a predefined concept pool. We have added both evaluations and clarified their scope in the revised paper.

Revision: Both evaluations are summarized in Section 7.5 (page 10, lines 515 and 529–531), with full results in Appendices E.2 and E.4.

We thank the reviewer again for the suggestions, which have helped us clarify the interpretation of our analyses and strengthen the downstream evaluation. We have incorporated these changes into the revised paper.

I thank the authors for their efforts. The additional clarifications and explanations have adequately addressed my concerns. I therefore recommend accepting the manuscript and will maintain my original positive score.

We are glad that our additional clarifications have adequately addressed the reviewer's concerns. We sincerely appreciate the reviewer's time, thoughtful consideration, and continued support of our work.

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

This paper studies why medical vision-language models answer incorrectly even when relevant concepts can be read from their internal states. Across three models, it uses linear probes, representation comparisons, and activation patching to study where information is retained and how it affects answers. MedSteer then uses a trained concept-presence probe to guide decoder edits, improving average balanced VQA accuracy by 5.2–12.6 percentage points.

Critical strengths
  • The paper distinguishes readable information from information actually used in an answer, and combines several analyses to study this gap.

  • The evaluation covers three models and several image types. Patient grouping and the meaning of positive and negative concept labels are clearly discussed. Hard-negative tests, random-direction controls, and patient-bootstrap intervals strengthen the analysis.

Critical weaknesses
  • The gain beyond the trained presence probe is unclear. The probe already predicts whether a concept is present and chooses the edit sign for each image. Although probe curves and random-direction controls are included, Table 2 lacks a direct probe-only comparison on the full test sets. Please add this and compare gated CAA or another direction using the same probe.

  • Activation patching shows recovery after artificial masking, which does not directly explain errors on the original image. A clean MLP output may restore broad image information as well as the target concept. Please analyze originally incorrect cases separately, compare equally sized masks over unrelated regions, and test whether removing the selected component damages correct answers.

Questions for authors

Please see the weaknesses.

5 comments

>>> Response to W1: We have conducted both requested comparisons on the full Table 2 test sets. Below, we clarify the scope of the original comparison and examine the contributions of the gate and steering direction.

(1) Probe-only prediction and the scope of the original comparison. Table 2 focuses on steering and decoding methods that modify the LVLM’s inference behavior. Our objective is to improve how the model uses available information when generating an answer. A linear probe provides an external prediction of whether a predefined concept is present, without changing the LVLM’s behavior.

For example, given the question “Is the kidney absent in this image?,” a kidney-presence probe may predict that the kidney is present. This prediction must then be mapped to the answer “No”; the probe itself does not perform this question-dependent mapping. Similarly, a binary probe does not directly generate open-ended descriptions or reports, as evaluated for MedSteer in Table 4.

This distinction motivated the original scope of Table 2. Nevertheless, we agree that probe-only prediction is an important reference for the binary task. We therefore evaluated both probe-only prediction and probe–logit fusion on the exact Table 2 test examples, as reported in Table R1 of our response to Reviewer G2ge’s Q1. Both outperform MedSteer on this binary task. We have explicitly acknowledged this result and added both comparisons to Table 2 as separate reference rows. However, we would like to clarify that, a presence probe is limited in practice and it cannot directly answer differently phrased questions without hand-crafted rules that map its prediction to the correct answer. In contrast, MedSteer’s contribution is to improve the generative model through internal interventions and assess whether those improvements transfer to diverse VQAs and open-ended generation.

(2) Same gate, different directions. The baselines in Table 2, except ODESteer, follow their original designs without a presence gate. To separate the contribution of our probe-guided gate from that of the steering direction, we equip the baseline directions with the same gate. For these gated comparisons, we hold the edit site, edit norm, and intervention strength fixed and vary only the direction, using the same training budget for the learned directions. We also include the original ungated VTI and CAA results for reference.

Table R4. Modality-averaged balanced accuracy (%) on the full Table 2 test sets. Gated methods use the same presence probe.

DirectionMedSIGHTLingshuHuatuoGPT-Vision
Unedited60.669.260.6
Gated Random60.769.461.2
VTI64.769.262.0
Gated VTI68.973.470.1
CAA65.973.067.7
Gated CAA70.375.569.9
MedSteer73.274.470.6

Two observations follow. First, the gate benefits other learned directions: gated VTI and gated CAA outperform their ungated counterparts across all three models, while the gated random direction provides little improvement. Second, the direction still matters when the gate is fixed. MedSteer performs best on MedSIGHT and HuatuoGPT-Vision, whereas gated CAA performs best on Lingshu, exceeding MedSteer by 1.1 percentage points. These results support contributions from both the gate and the direction, without establishing that MedSteer’s direction is uniformly superior on binary questions.

The construction of the directions suggests a possible explanation for differences in their transfer behavior. MedSteer derives its direction from activation changes induced by target-region masking, whereas CAA contrasts answer-position states associated with concept-present vs. concept-absent examples. The former may therefore capture changes more closely associated with the target image region, while the latter may be more closely associated with the binary answer format. To examine whether their effects extend beyond that format, we evaluated gated CAA on the open-ended generation tasks from Table 4.

Table R5. Open-ended generation in the Table 4 setting. Scores are percentages, reported as unedited → gated edit.

Task / MetricModelMedSteerGated CAA
Renal descriptions: target-term accuracyMedSIGHT41.0 → 47.041.0 → 42.5
Renal descriptions: target-term accuracyLingshu54.0 → 55.554.0 → 53.5
IU-Xray: target accuracyMedSIGHT54.3 → 56.754.3 → 53.6
IU-Xray: target accuracyLingshu51.0 → 51.951.0 → 50.2

Gated CAA is competitive on binary questions, but its improvements do not consistently transfer to the evaluated open-ended tasks: it improves one setting by 1.5 percentage points and decreases performance in the other three. MedSteer improves all four reported scores.

These results support more consistent transfer of MedSteer’s direction to the evaluated open-ended settings. They are consistent with the differences in direction construction described above, but do not establish that the directions exclusively encode visual evidence or answer decisions. We have added these comparisons and clarified this interpretation in the revised paper.

Revision: The probe references are in Table 2 and Section 7.2 (lines 369-376); the same-gate comparison is summarized in Section 7.3 (page 8, lines 414–422) and Figure 5(b), with full results in Appendix D.7 (page 31).

>>> Response to W2: We agree that recovery from artificial masking alone does not explain errors on the original image. We therefore conducted additional interventions on both originally incorrect and originally correct positive examples, with matched nontarget-mask controls.

(1) Originally incorrect cases and (2) matched nontarget masks. As detailed in Table R2 of our response to Reviewer G2ge, we intervene on positive examples that the original, unmasked model answers incorrectly. At the selected component, we amplify the difference between each image’s clean and target-masked activations. The control applies the same procedure using an area- and shape-matched mask outside the target region.

The target-based intervention corrects 20.2%, 21.7%, and 33.8% of the original errors for MedSIGHT, Lingshu, and HuatuoGPT-Vision, respectively, compared with 10.6%, 13.8%, and 10.6% for the matched control. The consistently higher correction rates support a stronger contribution from target-region perturbations than from the tested nontarget perturbations. However, they do not exclude contributions from contextual or other image information also affected by masking.

(3) Replacing the selected component’s output on originally correct examples. To test the complementary direction of influence, we replace the selected component’s output during inference on the unmasked image with its target-masked counterpart. The control uses the output from the matched nontarget-masked run. This tests the effect of changing the information supplied by the selected component; it does not remove or zero the entire component.

Table R6. Rates (%) at which originally correct positive answers become incorrect after replacing the selected component’s output with its masked counterpart. Entries report target-mask / matched-control rates; n denotes the number of originally correct examples.

ModelAll concepts: target / control (n)Findings: target / control (n)
MedSIGHT8.9 / 1.9 (426)29.0 / 1.5 (69)
Lingshu2.6 / 0.7 (872)8.9 / 1.7 (180)
HuatuoGPT-Vision3.3 / 1.7 (754)6.3 / 2.5 (79)

Across all three models, target-masked replacement disrupts more correct answers than the matched control, with larger effects for findings than for the full concept set. For MedSIGHT findings, it changes 29.0% of originally correct answers to incorrect answers, compared with 1.5% for the control. The smaller effects on organ concepts may reflect the availability of surrounding anatomical cues, although the intervention does not directly test this explanation.

Together, these analyses demonstrate that interventions at the selected component can both correct naturally occurring errors and disrupt originally correct answers. In both cases, target-region interventions have larger effects than the matched nontarget controls. This supports a causal contribution of the selected component’s activations to the evaluated predictions, while stopping short of establishing strict necessity, sufficiency, or exclusively target-specific information. We have added these analyses, including the finding/organ breakdown, and clarified the scope of our mechanistic conclusions.

Revision: Both analyses are now in Appendix C.4 (pages 24–25), referenced in Section 6.3 (page 6, lines 314–316).

We thank the reviewer again for the suggestions, which have helped us strengthen the controlled comparisons and clarify the interpretation of our results. We have incorporated these analyses into the revised paper.

Reviewer_G2geOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates why medical vision-language models can answer concept-presence questions incorrectly even when the relevant concept remains recoverable from their internal representations. Across three models, the authors use linear probes and representation comparisons to trace concept information from the visual encoder through the connector and decoder. They then examine attention and MLP activations associated with correct and incorrect answers, and use activation patching to identify components that affect concept-presence predictions. Based on these analyses, they propose MedSteer, which edits a decoder activation along a concept-specific direction, with a presence probe determining the edit sign. On segmentation-derived concept-presence tasks, MedSteer improves balanced accuracy over the unedited models by 5.2–12.6 percentage points.

Critical strengths
  1. Careful distinction between information availability and information use. The paper does not treat a successful linear probe as proof that the model uses the decoded concept. It adds decoder analyses and activation patching to examine where concept information can affect an answer. This makes the analysis substantially stronger than relying on probing or attention alone.
  2. Consistent gains with meaningful intervention controls. MedSteer improves results across all three evaluated models, and the paper compares it with several adapted steering and decoding baselines. Removing the input-dependent gate or replacing the learned direction with a random direction reduces performance. These results support the effectiveness of the complete method on the concept-presence task evaluated here.
Critical weaknesses
  1. The contribution beyond the presence probe is not sufficiently isolated. At inference, a supervised presence probe decides whether MedSteer adds or subtracts its concept direction. Figure 5(b) suggests that directly using this probe can be more accurate than the steered model at some intervention depths, particularly for MedSIGHT. The learned direction is not without value—the ablations and a comparator sharing the gate provide evidence for its effect—but the main comparison omits a straightforward probe-only predictor or probe–model logit fusion. For a binary task, that omission makes it difficult to judge the practical benefit of editing the decoder rather than using the probe’s prediction directly.
  2. The mechanistic claim does not yet fully explain naturally occurring errors. Activation patching shows that restoring an unmasked component output can increase concept-presence support in a target-masked positive example. It does not directly show that patching corrects cases the original model answered incorrectly without masking. Moreover, some negative examples are slices where an organ is outside the image, including slices from patients who also provide positive examples. This is a valid slice-level task, but it is narrower than distinguishing clinically plausible abnormalities.
Questions for authors
  1. On the exact test examples used in Table 2, how does MedSteer compare with directly using the presence probe, or combining its score with the model’s answer logits? A paired comparison would clarify what decoder steering adds beyond the probe.
  2. For examples the original, unmasked model answers incorrectly, does patching the selected component correct the answer? How does the effect compare with masking an area-matched nontarget region?
  3. Can the authors separately report performance on difficult negatives where the relevant anatomy is visible but the target abnormality is absent? This would clarify how much of the gain extends beyond straightforward slice-level presence detection.
5 comments
AuthorsOpenReview ↗

Response to reviewer G2ge

We thank the reviewer for the thoughtful assessment and for recognizing the distinction between information availability and use, as well as the consistency of MedSteer’s gains. These suggestions help us clarify what decoder steering contributes beyond the presence probe and the scope of our mechanistic conclusions.

>>> Response to W1 and Q1: We first explain the scope of the original comparison and then present the requested results in Table R1.

(1) Rationale for the original comparison. Table 2 focuses on steering and decoding methods that modify the LVLM’s inference behavior. Our objective is to improve how the model uses available information when generating an answer. A linear probe instead provides an external prediction of whether a predefined concept is present, without changing the LVLM’s behavior.

For example, given the question “Is the kidney absent in this image?,” a kidney-presence probe may predict that the kidney is present. This prediction must then be mapped to the answer “No”, but the probe itself does not perform this question-dependent mapping. Similarly, a binary probe does not directly generate open-ended descriptions or reports, as evaluated for MedSteer in Table 4.

This difference motivated the original scope of Table 2. Nevertheless, we agree that probe-only prediction and probe–logit fusion are important comparisons for the binary task evaluated here.

(2) Requested comparison. Following standard probing practice (Alain & Bengio, 2017; Belinkov, 2022), we use the probe to assess whether concept labels are linearly recoverable from frozen representations. Its performance provides a readout reference. The difference between probe performance and the unedited model’s answers quantifies the operational availability–use gap studied in this paper.

We evaluated two probe-based predictors on the exact test examples used in Table 2: (i) Probe-only, which directly uses the presence probe at the selected intervention layer, ℓ*; and (ii) Probe–logit fusion, which adds the probe score to the model’s yes/no log-odds, with the fusion weight selected on the validation set.

Table R1. Average balanced accuracy (%) on the Table 2 test sets.

MethodPrediction mechanismMedSIGHTLingshuHuatuoGPT-Vision
UneditedOriginal LVLM output60.669.260.6
MedSteerLVLM output after decoder activation editing73.274.470.6
Probe-onlyExternal classifier on frozen representations90.591.290.6
Probe–logit fusionProbe-assisted rescoring of yes/no logits90.491.290.9

Both probe-based predictors outperform MedSteer on this binary task, consistent with the probe results in Fig. 5(b). However, as illustrated by the example above, the probe only outputs a fixed presence label and cannot read the question. For example, in VQA datasets, if the question is rephrased so that the correct answer flips (e.g., “Is the kidney absent in this image?”), the probe’s prediction no longer maps to the correct answer without an additional question-dependent rule. Answer-contrast steering directions such as CAA may face the same limitation. Thus, we only use linear probing results as readout references. Relative to the probe-only reference, MedSteer closes approximately 24–42% of the balanced-accuracy gap between the unedited model and the probe, but does not eliminate it.

MedSteer’s different contribution is an intervention within the generative model whose effects can also be evaluated on open-ended generation. Table 4 demonstrates improvements in this setting. The binary probe and yes/no fusion rules do not directly specify an open-ended generation procedure; extending them to descriptions or reports would require an additional mechanism. We have clarified this distinction and added both probe-based comparisons to Table 2 as separate reference rows.

Revision: We have added both predictors as gray reference rows in Table 2 (page 8) and explained their role in Section 7.2 (page 7, lines 369–376), with details in Appendix D.6 (page 30).

References

  • Alain, G., & Bengio, Y. (2017). Understanding intermediate layers using linear classifier probes. ICLR Workshop.
  • Belinkov, Y. (2022). Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1), 207–219.

>>> Response to W2-1 and Q2: We agree that restoring an unmasked activation in a target-masked example does not directly establish that the same component can correct an error on the original image. To address this distinction, we conducted an additional intervention on positive test examples that the original, unmasked model answered incorrectly.

The evaluation uses the CT and chest X-ray test sets from Table 2 and covers five findings (kidney tumor from KiTS23, nodule from LIDC-IDRI, COVID-19 infection, lung opacity, and pneumothorax) and six organs from AMOS22 and the chest X-ray collections.

For these examples, simply restoring the clean activation would reproduce the original error. We therefore amplify the activation difference induced by masking the target region at the selected component:

h′=hclean+α(hclean−hmasked),α=4.h' = h_{\mathrm{clean}} + \alpha\bigl(h_{\mathrm{clean}} - h_{\mathrm{masked}}\bigr), \qquad \alpha = 4.

Both activations come from the same image, with and without the target mask. As a control, we apply the same intervention using an area- and shape-matched mask outside the target region. This experiment amplifies the target-mask-induced activation difference during inference on the original image, rather than restoring a clean activation in a masked input.

Table R2. Correction rates (%) among originally incorrect positive examples.

ModelNumber of original errorsTarget maskMatched nontarget mask
MedSIGHT64420.210.6
Lingshu24021.713.8
HuatuoGPT-Vision35833.810.6

The target-based intervention corrects 20.2–33.8% of these errors, exceeding the matched-mask control by 7.9–23.2 percentage points. These results show that interventions at the selected component can correct naturally occurring errors and that their effectiveness depends on the perturbed image region.

The matched control strengthens the link to target-region information. However, it does not establish that the activation difference contains exclusively target-concept information or fully explain why the original errors occurred. We have added this analysis and clarified the scope of the mechanistic conclusions accordingly.

Revision: This analysis is now in Appendix C.4 and referenced in Section 6.3 (page 6, lines 314–316).

>>> Response to W2-2 and Q3: We agree that organ-out-of-view negatives test a narrower capability than distinguishing an abnormality from normal tissue within visible anatomy. Because our labels are derived from segmentation masks, a negative indicates only that the target is not visible in the image or slice; some original negatives are therefore slices in which the organ lies outside the field of view, or radiographs from a different source than the positives. Such negatives can be answered without recognizing the abnormality itself. We therefore constructed the following hard-negative subsets, in which the relevant anatomy is visible but the target abnormality is absent:

  • KiTS23 kidney tumor: kidney-containing slices without a tumor, drawn from the same patients as the positive slices. These comparisons test abnormality detection within shared patient anatomy. The training and test folds remain patient-disjoint.
  • Chest X-ray lung opacity: lung-opacity and normal radiographs from the same Radiography collection.
  • Chest X-ray COVID-19: COVID-19 radiographs compared with normal and non-COVID lung-disease radiographs from the same COVID-QU-Ex collection.

Within each chest X-ray comparison, positives and negatives share the image source and preprocessing, which removes differences in image source as a possible shortcut. We use the directions, gates, and strengths from Table 2 without re-tuning. To make the comparison controlled, we report each concept under both the original and the hard negatives, using the same test positives and the same number of negatives.

Table R3. Balanced accuracy (%) under the original and the hard negatives, reported as unedited → MedSteer. Numbers in parentheses give the positive / negative test examples.

ConceptMedSIGHT: OriginalMedSIGHT: HardLingshu: OriginalLingshu: HardHuatuoGPT-Vision: OriginalHuatuoGPT-Vision: Hard
KiTS23 kidney tumor (123 / 120)58.1 → 66.658.5 → 67.559.7 → 59.7*60.9 → 60.9*53.2 → 60.854.5 → 60.1
CXR lung opacity (72 / 83)66.6 → 83.781.2 → 88.769.5 → 72.792.5 → 93.856.8 → 64.566.6 → 81.8
CXR COVID-19 (79 / 91)50.0 → 82.651.3 → 84.080.7 → 87.485.9 → 89.650.0 → 50.2*50.0 → 50.1*
Average improvement+19.4+16.4+3.3+1.7+5.2+7.0

* denotes that validation selected no steering (α = 0) or the weakest steering strength.

Both performance and gains are broadly similar under the two constructions: the average gains of MedSteer differ by at most 3.0 points for each model (e.g., +19.4 vs. +16.4 points for MedSIGHT), and for kidney tumor and COVID-19, unedited accuracy differs by at most 5.2 points across all three models.

The two constructions differ in what they test as negatives. Original negatives may lack the organ or come from a different source, whereas hard negatives keep the anatomy and image source fixed and remove only the abnormality. The persistence of the gains therefore indicates that MedSteer does not rely on organ-out-of-view or source cues. Lung opacity shows the largest difference: unedited accuracy is consistently higher under hard negatives, while the change in gains varies across models (smaller for MedSIGHT and Lingshu, larger for HuatuoGPT-Vision). This is likely because its hard negatives are normal radiographs, whereas its original negatives may include other lung abnormalities, such as pneumonia, that resemble opacity. Across both constructions, MedSteer improves every setting except those for which validation selected no edit or the smallest strength, and lowers none. We have added this evaluation to the main paper as a new subsection.

Revision: This evaluation is now a new Section 7.4 in the main paper (page 9, with Table 3 on page 10).

We thank the reviewer again for these insightful suggestions. The added probe-based comparisons, interventions on naturally occurring errors, and hard-negative evaluation have helped us clarify what decoder steering contributes beyond the presence probe and sharpen the scope of our mechanistic conclusions. We have incorporated these results into the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.