ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs

AI summary

Overview

• Research area: Evaluation and safety auditing of medical vision-language models (VLMs) — specifically, measuring whether a model's answer actually depends on the image when a radiology report is also supplied.
• Technical level: Intermediate. The evaluation protocol is conceptually simple (swap the image, hold everything else fixed), but the paper assumes familiarity with vision-language prompting, logit readouts, and transformer-layer interventions.
• Scope: The paper introduces ModaLens, a paired image-swap audit, and applies it to MedGemma-27B on 3,199 MIMIC-CXR cases (44,786 paired trials, 293 patients) to quantify how report availability changes image sensitivity.

What This Paper Is About

Medical VLMs are often given both a chest X-ray image and its associated radiology report, but because the report usually already contains the answer, it is impossible to tell from accuracy alone whether the model is looking at the image at all. ModaLens addresses this by rendering every case twice — once with the case's own image and once with another study's image — while keeping the question and the report identical, so any change in the model's answer can only be attributed to the image. The goal is to measure how much the model's output moves when the image is substituted, and how that sensitivity changes when the report is present versus removed.

Key Contributions

1. A paired image-swap audit protocol (ModaLens) in which a case is rendered twice, identical in every token except the image (the "concordant" input with the case's own image and the "discordant" input with another study's image), so the measured quantity is counterfactual answer sensitivity rather than accuracy.
2. A report-availability manipulation: the same case and same substitution are run with the report present and absent, reporting a paired flip rate and margin change (Figure 1 is the study's visual summary).
3. A multi-readout and multi-model validation, covering lowercase first-token, token-family, and generated-answer readouts, plus replication across MedGemma-4B, Qwen3.5-9B, Qwen3.5-27B, and LLaVA-NeXT on Mistral-7B, alongside per-finding, per-category, and per-layer secondary analyses.
4. A layer-level investigation of when direct access to report-token positions matters, using cumulative attention knockout and controls, reported as a bound (near layer 24 of 62) rather than a locus, plus two openly reported negative mechanistic results (steering and probe transfer).

Main Findings

• Report availability drives a large paired change in image sensitivity. Under an explicit answer instruction, MedGemma-27B's generated answer changes with the image on 4.26% of trials with the report versus 20.94% without it, a paired increase of 16.7 points (patient-clustered 95% CI [15.6, 17.7]).
• The original prompt gives a similar direction and a somewhat smaller estimate. With the lowercase first-token readout and no answer instruction, the rates are 4.70% with the report and 17.07% without, a paired difference of +12.4 points [11.2, 13.6] and a ratio of 3.6 [3.4, 3.9].
• The three readouts agree under the answer instruction. Lowercase first token: 4.64% versus 20.72%, +16.1 points [14.9, 17.2]. Token families: 4.30% versus 21.03%, +16.7 points [15.7, 17.8]. Generated answer: 4.26% versus 20.94%, +16.7 points [15.6, 17.7].
• Continuous answer margins move even when the binary prediction does not. With the report present, the mean absolute paired margin change is 0.690 [0.657, 0.726], and 32.2% of trials move by more than 0.5 while only 4.7% flip; 33.0% of label-unchanged trials also move by more than 0.5. Without the report the mean change is 2.307, larger than with it on 82.9% of trials.
• The effect is not uniform across findings. Five of the 14 report-absent readouts sit at or below 4.1% because the model answers yes on 97.8% to 100% of trials without the report (atelectasis and support devices flip on no trial at all), while the other nine run from 16.8% to 32.0%. Under the generated answer, atelectasis is 6.81% with the report versus 4.63% without, and support devices 11.03% versus 1.44%.
• The direction replicates across model families and sizes. Under the answer instruction the paired report-presence difference is 13.3 points [11.2, 15.6] in Qwen3.5-9B, 19.60 points [17.00, 22.40] in Qwen3.5-27B, and 12.75 points [11.10, 14.44] in LLaVA-NeXT on Mistral-7B. MedGemma-4B flips 2.7 points more than MedGemma-27B.
• The effect survives warnings and framing changes. Warning the model that the report may not describe the image and may be wrong gives a one-question flip rate of 1.50% against 1.22% under the plain framing, a paired difference of +0.28 points [-0.13, 0.74].
• Prompt order changes the magnitude, not the direction. With the text block first, the report effect is +7.8 points [6.9, 8.8] versus +12.4 [11.2, 13.6] with the image first, on identical trials.
• Text content also affects image sensitivity. In the one-question design, length-matched neutral prose and clinical boilerplate give 8.2% and 7.1%, compared with 10.3% without a report and 1.4% with the full report; removing the sentences mentioning the queried finding raises the rate to 11.6%, and a sentence stating the reference answer reduces it to 3.1%.
• Adding the image to the report lowers agreement with report-derived labels. Report alone reaches a pooled balanced accuracy of 0.770; report plus image gives 0.717; image plus question gives 0.665; question alone is at chance. The image raises balanced accuracy by 0.165 [0.158, 0.172] without the report and lowers it by 0.053 with it (pooled interaction −0.218).
• Direct access to report-token positions matters at shallow depths. Masking the text stream's attention to the report raises the image-swap flip rate from 1.50% to 7.75% when the block begins at layer 0, against 12.00% with no report, and is at baseline from layer 32; single-layer blocking never exceeds 2.50%.
• Two mechanistic analyses are negative. Difference-of-means steering changed the share of answers tracking the substituted image by 0.1 points at the largest scale, and a linear probe fit under a frozen protocol and run once on a disjoint patient cohort fell from its test-split discrimination to chance.
• The image-decisive control behaves differently. On VQA-RAD (n = 52), which has no report, the flip rate under a final-token yes/no readout is 61.5% [48.1, 75.0] with accuracy 0.731 on both images.

Methodology in Plain English

The researchers take a chest-X-ray dataset where each image comes with a written report, then run every case through the model twice. In the first run the model sees the case's own image plus its report; in the second run the image is replaced by one from a different study — usually the same patient — while the question and the report stay exactly the same. If the model's answer changes between the two runs, the image must have mattered; if it does not, the model likely answered from the text. The same pair of runs is repeated with the report deleted, and the difference between the two conditions is the reported "report effect."

Substitute images were chosen by sampling from the patient's other studies whose CheXpert positive set differs, falling back to any other study, and for the 19 single-study patients a differently labelled image from another patient. View and acquisition time were not matched, and no pair exceeded the specified structural-similarity (SSIM) threshold of 0.9. The model's answer was read three ways: a lowercase first-token logit comparison, token families, and the generated answer text. Statistical intervals come from patient-clustered percentile bootstraps over the 293 patients — 10,000 draws for the primary all-14 rates and 2,000 elsewhere. A separate experiment masks the model's attention to report-token positions, cumulatively from a given decoder layer upward, to see when blocking direct access to the report changes the measured flip rate.

Why This Matters

Impact on research. Accuracy benchmarks cannot distinguish a model that reads the image from one that reads the report. ModaLens supplies a paired, within-case measurement that isolates sensitivity to an image intervention, and the paper reports a substantial and consistent effect: 4.26% versus 20.94% under the answer instruction, and 4.70% versus 17.07% under the original prompt. It also documents the limits of its own claim — labels are report-derived throughout, so the audit says how much the readout moves with the image, not whether following the text was visually wrong.

Real-world applications:

• Auditing clinical decision-support models before deployment, to check whether a model shown both image and text is genuinely using the image.
• Designing evaluation suites for report-conditioned models, where questions and reports must be held fixed to separate modality contributions.
• Interpretability and monitoring work in medical AI, using the layer-level knockout as a bound on when direct report-token access matters.
• Building comparable controls for datasets without reports, illustrated by the VQA-RAD image-decisive control (n = 52, 61.5% flip rate) and the SLAKE question-type breakdown.

Industry relevance. Any vendor or hospital deploying a multimodal model where a report, prior note, or clinical summary appears alongside an image faces the same confound this paper measures. The reported odds ratio for report presence is 0.12 [0.09, 0.17], and the effect persists across three model families and two sizes within two of them, which suggests the issue is not specific to one snapshot. The authors state that on the label-conditioned one-question set, no model tested reads the substituted image above chance once the report is removed — so what generalises there is report anchoring rather than image competence.

Future Directions

• Independent image annotations. The paper states explicitly that it did not obtain independent image annotations for this evaluation, and that such annotations are needed to determine when the observed behaviour produces visually incorrect answers; the only image-level labels here are the classifier surrogate.
• Testing the manipulated factor outside MIMIC-CXR. Every report-availability estimate comes from one dataset — the MIMIC-CXR test split, 293 patients at a single institution, frontal chest radiographs and English reports — and no other dataset used in the paper pairs an image with a report.
• Separating answer-bearing text from length and context effects. The text-content controls (8.2% for neutral prose, 7.1% for boilerplate, 11.6% when queried-finding sentences are removed, 3.1% when the reference answer is stated) suggest answer-bearing text contributes, but do not isolate its contribution from changes in length and context.
• Better localisation and readout tools. The uncalibrated logit lens is used throughout, the tuned lens scored below it on both readout datasets, activation patching leaves the depth under-determined, and cross-condition residual patching did not separate from a wrong-donor control (14.07% versus a correct-donor maximum of 12.25%).

Target Audience

Researchers and engineers working on medical multimodal models, clinical AI safety and evaluation, and interpretability of vision-language systems; also useful for clinicians and regulators who need to know whether a model shown both an image and a report is actually using the image. Readers should be comfortable with benchmark evaluation and basic transformer concepts, though the core protocol — swap the image, hold the rest fixed, compare flip rates — is easy to follow without that background.

Authors’ abstract

A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also uses the image. ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity: MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients, all 14 questions per case (13 finding-specific and one composite), each image replaced by one from another study, usually of the same patient, with question and report fixed. Under an explicit answer instruction, the model's generated answer changes on 4.26 percent of trials with the report and 20.94 percent without it, a paired increase of 16.7 points (patient-clustered 95 percent CI 15.6 to 17.7), so report availability reduces image-swap sensitivity under this protocol; the original prompt with a lowercase first-token readout gives 4.70 percent against 17.07 percent, and substitutions also move continuous answer scores where the binary prediction does not change. The labels are derived from reports, which limits conclusions about visual correctness; the direction replicates in two further model lineages. Code, the exact prompts and a run record for every number are at https://github.com/criticaldata/MODALENS.