{"id":"70cad543-866d-4124-b82b-bcc574d9bfb5","arxiv_id":"2507.01790","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Vision-language models answer from whichever modality is encoded more saliently in their final-layer representations, and specific attention heads can be manipulated to shift that preference.","lead":"This paper tested four vision-language models on image-caption pairs that disagree, asking the model to report what the image shows or what the caption says. It found models often answer from the wrong modality, and it identifies attention heads that can be amplified to push the model toward the requested modality.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central mechanistic claim rests on a correlation between V-Measure computed on final-layer representations and behavioral accuracy, but the clustering analyses used to select the three reported heads and the head-intervention evaluations are performed on overlapping Pascal VOC data without an…","rationale":"The reader correctly identifies the final-layer V-Measure correlation as partly mechanical and notes the absence of error bars and out-of-sample head selection. I agree with those points; however, the more directly load-bearing concern for the paper's central causal claim is the selection-and-evaluation overlap on the 100 Pascal VOC examples, because the paper's headline contribution is not the correlational V-Measure result but the existence of transferable router/promotion heads that 'can be manipulated or transferred.' If the head selection itself is contaminated by selection on the evaluation sample, the causal attribution fails even if the V-Measure correlation is interpreted charitably. I therefore rank this concern as primary, with the reader's final-layer circularity concern as a secondary but genuine issue. The proposed test of pre-registered split-then-select evaluation specifically targets the load-bearing gap, while a secondary test of recomputing the Section 3.3 correlation on intermediate layers (e.g., layer L-5) would address the circularity concern. I recommend keeping the CONDITIONAL verdict rather than UNCHANGED, because the central causal claim is plausible and independently supported by the cross-dataset transfer pattern, but it is not yet established without an out-of-sample selection protocol.","tokens_in":18193,"tokens_out":1873,"duration_ms":17923,"concrete_test":"Split the Pascal VOC data into two disjoint sets before any head selection. On split A, run the full scanning protocol from Section 4.1 over all heads and pick the top router and promotion heads by the paper's trend criteria. Then, on split B and on the four other datasets, evaluate only these pre-selected heads at α = 10, computing bootstrap 95% confidence intervals over samples and over K-Means seeds for both accuracy and V-Measure. If the held-out Pascal VOC accuracy gains and the cross-dataset transfer gains (Table 1) fall within the bootstrap intervals and remain significant, the causal attribution of modality control to specific heads is supported. If the gains shrink to noise on held-out data, the head-selection procedure was overfit to the 100-pair evaluation set.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's load-bearing move is Section 4: it claims that specific attention heads, found on Pascal VOC, causally control modality selection and transfer across datasets. The selection of L11H14, L19H26, and L13H26 is based on intervention patterns computed on 100 Pascal VOC image-caption pairs (Section 4.1), and the headline improvements in Table 1 are also measured on Pascal VOC. Moreover, the heads are selected partly by 'intervenability' (Appendix F.2), a criterion defined as the ability to improve performance beyond the unintervened baseline. This creates a selection-on-the-test-set risk: on a 100-pair sample with no reported confidence intervals, a head that happens to produce the largest accuracy swing on that sample will look like a 'router head,' even if it does not specifically route modality. The cross-dataset rows in Table 1 are the strongest evidence, but they are evaluated for only three hand-picked heads, with no error bars, and the CUB-Color exceptions (router head hurts caption accuracy by 4.9%) already show the generalization is not universal. The claim that the representations 'predict behavior' (Section 3.3, r = 0.94) is also partly mechanical because the final-layer activations used for K-Means feed directly into the output distribution; this weakens the inference that V-Measure is an independent window into a causal salience mechanism, although the head-intervention data could still support the causal claim independently. A test that genuinely settles the causal claim is to pre-register the head-selection procedure on a training split of Pascal VOC, then evaluate the selected heads on held-out Pascal VOC samples and on CIFAR-10/CIFAR-100/ImageNet100/CUB-Color, reporting bootstrap confidence intervals over both samples and seeds.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how vision-language models (VLMs) process conflicting information across modalities by constructing inconsistent image-caption pairs and prompting models to report information from a designated target modality. The authors evaluate four 7B-scale VLMs on five datasets, finding that models exhibit varying modality biases and that accuracy drops under conflict. They then use linear probes to show that both modality-specific information and inter-modality consistency are encoded in the hidden representations, and apply K-Means clustering with V-Measure to argue that the relative representational salience of the target versus non-target modality predicts behavioral accuracy (Section 3.3, r = 0.94). Finally, they perform attention-head interventions, identifying modality-agnostic 'router heads' and modality-specific 'promotion heads' that, when scaled, shift model behavior toward one modality; they report cross-dataset transfer of three heads selected on Pascal VOC (Sections 4.1-4.3).","tokens_in":18466,"tokens_out":2984,"duration_ms":34825,"significance":"If the causal claims are substantiated, the paper offers a mechanistic account of modality selection in VLMs and a practical handle for controlling it, which would be a meaningful contribution to interpretability and robustness in multimodal models. The strengths are the breadth of behavioral evaluation across four models and five datasets, the inclusion of the reciprocal direction (reporting caption content under conflicting images), the careful construction of inconsistent pairs, and the public release of code. The cross-dataset head-intervention results, despite the caveats discussed below, suggest a potentially generalizable phenomenon. However, the load-bearing mechanistic claims currently rest on statistical and design choices that need to be tightened before the conclusions can be accepted.","major_comments":[{"comment":"The head selection and the headline evaluation are performed on the same Pascal VOC data: 100 conflicting image-caption pairs (200 samples) are used both to identify the router/promotion heads and to report the Pascal VOC rows in Table 1. Because heads are also ranked by 'intervenability' (Appendix F.2), a head that happens to produce a large accuracy swing on this small sample will be selected regardless of whether it specifically controls modality routing. This selection-on-the-test-set risk makes the Pascal VOC columns in Table 1 uninterpretable as evidence. Please hold out a separate split for head selection, evaluate on a disjoint test set, report confidence intervals or bootstrap intervals over multiple data samples and random seeds, and include a control condition with randomly chosen heads (or heads selected on a different dataset).","section":"Section 4.1, Table 1"},{"comment":"The r = 0.94 correlation between V-Measure and behavioral accuracy is computed on the final-layer activations, which feed directly into the output token distribution used to score accuracy. This makes the correlation partly mechanical: the same representation that determines the next-token prediction is clustered and compared with the prediction. The claim that 'the structure of the representations predicts behavior' is therefore not a test of an independent internal cause. Please report the correlation for earlier layers (e.g., layers that do not directly project to the unembedding), or explicitly reframe the result as a descriptive consistency check rather than mechanistic evidence. If the correlation persists in middle layers, the interpretation is considerably stronger.","section":"Section 3.3, Figure 4(b)"},{"comment":"The cross-dataset generalization claim rests on only three heads, with no confidence intervals, significance tests, or multiple-seed variation. The single router head L11H14 actually degrades caption accuracy on CUB-Color by 4.9%, which the paper attributes to 'suboptimal image unimodal performance' without a quantitative test. Please provide error bars for all entries in Table 1, evaluate a larger set of candidate heads (including heads that do not meet the classification criteria), and test whether the observed cross-dataset improvements are significantly larger than those from random heads. As written, the exception on CUB-Color and the absence of variance estimates leave the universality of the router-head claim unsubstantiated.","section":"Section 4.3, Table 1"},{"comment":"The procedure for classifying heads is not fully specified: the monotonicity epsilon is never given, and the final choice of L11H14, L19H26, and L13H26 is described as prioritizing 'intervenability' on the same data used for evaluation. This makes the reported head identities hard to reproduce and contributes to the selection concern in Major Comment 1. Please provide the exact numerical criteria, the full set of heads that meet each criterion, and a held-out validation of the selection rule.","section":"Appendix F.1 and F.2"}],"minor_comments":[{"comment":"Typographical errors: 'thr VLM' should be 'the VLM', and 'ImagetNet100' should be 'ImageNet100'.","section":"Section 2"},{"comment":"The caption of Figure D.1 contains 'acorss', which should be 'across'.","section":"Appendix D.1"},{"comment":"The reference to Tenney et al. is malformed: 'Tenney et al., Belinkov, 2022' is missing the year for the first citation and the formatting should be corrected.","section":"Related Work"},{"comment":"The paper reports the number of sampled pairs and total samples (100 and 200) but does not state how many random seeds were used for the intervention sweeps or whether the sample was stratified by class; please add this information for reproducibility.","section":"Section 4.1"},{"comment":"The V-Measure curves are averaged over three K-Means initializations, but no variance or error bars are shown; adding them would strengthen the claim that the CIFAR-100 versus Pascal VOC difference is robust.","section":"Figure 4(a)"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the interpretability and multimodal communities, but the current version overclaims mechanistic causality from a correlation that is partly mechanical and from head-intervention results that are selected and evaluated on the same small sample. The issues are fixable with held-out splits, confidence intervals, and control heads; I would be willing to review a revision. I do not see evidence of problematic citation practices, but the authors may want to check the formatting of the references list."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a serious referee, but the central mechanistic claim is softer than the abstract suggests. What is genuinely new: they study the reciprocal case (asking the model to report the caption under a conflicting image), find that models differ in which modality they favor, and then trace that preference to specific attention heads that transfer across datasets. The behavioral results across four models and five datasets are clearly presented, and the probes in Section 3 do a good job ruling out simple failure-to-encode explanations. The consistency-probe generalization check in Appendix E.2 is a nice robustness test that most papers skip. Code is released, which is real evidence.\n\nThe soft spots are where the causal story does heavy lifting. The r=0.94 correlation between V-Measure and accuracy is computed on final-layer activations that feed directly into the output distribution, so part of the correlation is mechanical: the representations are not an independent window into processing. That weakens the claim that representational salience \"predicts\" behavior, though the head-intervention evidence remains independent and is the stronger part of the paper.\n\nOn the interventions: the heads are selected on 100 Pascal VOC pairs (200 samples) and the headline improvements in Table 1 are also on Pascal VOC, with no confidence intervals or seeds. That is a real selection-on-the-test-set risk. The cross-dataset rows are the best evidence, and the promotion heads generalize consistently, but the router head does not help caption reporting on CUB-Color (-4.9%) and barely helps on CIFAR-10 (-0.2%). The paper's own Appendix F.2 admits they report only the heads with the most salient trends or intervenability, which is selective reporting. The V-Measure post-intervention changes in Table 2 also inherit the same mechanical correlation issue.\n\nNone of this is fatal, and the stress-test may overstate the problem by ignoring the monotonicity criterion used to classify heads, which is stricter than just picking the largest swing. Still, the paper needs error bars, seed variance, and a held-out head-selection procedure before the causal claims are fully established.\n\nThis is a useful paper for interpretability researchers and people building multimodal systems who want to know where modality bias comes from. It is a solid empirical contribution with a plausible mechanism, not a settled result. I would send it to peer review and ask for the statistical tightening.","headline":"A plausible and useful head-level story about modality bias in VLMs, but the headline correlation and head-selection are statistically softer than the paper lets on; still, this deserves a serious referee.","tokens_in":19108,"tokens_out":2108,"would_cite":true,"duration_ms":25445,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Vision-language models answer from whichever modality their hidden representations encode more saliently, and individual attention heads can be manipulated—and transferred across datasets—to shift that preference.","keywords":["vision-language models","conflicting inputs","modality bias","representational salience","attention head intervention","router heads","V-Measure","cross-modal interference"],"falsifier":"On Qwen2.5-VL with CIFAR-100, compute the V-Measure gap from an intermediate layer that does not feed the output distribution, then scale the router head L11H14 and check whether the mid-layer gap moves with accuracy; if it does not, the final-layer salience measure is reading off the imminent answer instead of explaining the behavior.","tokens_in":17939,"feed_emoji":"🧠","tokens_out":8634,"duration_ms":86804,"temperature":0.7,"pith_summary":"Vision-language models presented with an image and a conflicting caption often answer from the wrong modality, and the paper argues this is not random: a model tends to report whichever modality is encoded more saliently in its hidden representations. Probing shows the models separately encode both modalities and can detect the inconsistency, so the failure is not a simple inability to read the inputs. K-Means clustering with V-Measure shows that the gap between target-modality and non-target-modality salience in late-layer activations predicts accuracy across model–dataset–task triples ($r = 0.94$). The paper then identifies individual attention heads—a modality-agnostic router head and modality-specific promotion heads—that, when scaled, shift that salience and the behavior, and the same heads generalize across datasets. If right, this makes modality preference in multimodal models a localized, manipulable mechanism rather than a fixed property.","feed_headline":"A few attention heads pick the modality a VLM trusts","feed_subtitle":"When image and caption disagree, behavior tracks which one is more salient inside the model—and one head can switch that.","key_machinery":"The paper's mechanism has two linked parts. Representational salience is measured by fitting K-Means clustering with as many clusters as there are classes to the hidden representations of inconsistent inputs, then scoring how well the clusters align with image labels versus caption labels using V-Measure; the target-minus-non-target V-Measure difference is the proxy for which modality dominates. The causal part is head intervention: the output of one attention head at the last prompt-token position is multiplied by a scalar $\\alpha$ ranging from $-10$ to $10$, and the resulting change in answers classifies the head as a modality-agnostic router head (it promotes whichever modality the prompt requests), an image promotion head, or a caption promotion head. Intervening on the same heads also shifts the V-Measure gap, which is the evidence that the heads restructure the representations rather than only changing the output token.","core_discovery":"The central claim is that when a vision-language model is asked to report one modality from an inconsistent image–caption pair, its behavior tracks the relative representational salience of the two modalities, and that salience is controlled by specific attention heads. The supporting evidence has three layers: linear probes show both image and caption information is encoded and the inconsistency is detectable; K-Means clustering with V-Measure shows the last-layer salience gap between target and non-target modality strongly predicts behavioral accuracy; and scaling individual attention-head outputs by a constant $\\alpha$ reveals heads that either promote whichever modality the instruction requests (router heads) or promote one fixed modality (image or caption promotion heads). The router head and promotion heads found on one dataset alter performance in the expected direction on other datasets, and their intervention moves the V-Measure gap in the matching direction. The paper therefore concludes that modality selection in vision-language models is attributable to identifiable, manipulable components.","pith_inferences":["If the same scaling intervention were applied to larger or differently trained multimodal models, one would expect analogous router heads to exist, making this a general recipe for locating modality-selection circuitry rather than a description of one model family.","The salience-versus-accuracy logic should transfer to other modality pairs such as audio–text or video–text, where the V-Measure gap could serve as a diagnostic for which stream a model will trust under conflict.","A practical extension the paper does not pursue is using router-head amplification as an inference-time steering control to raise target-modality accuracy without fine-tuning or retraining.","The fact that consistency is detectable in the representations but not always used suggests future interventions could target the layer where consistency information is integrated, before the answer is formed."],"forward_implications":["Models that fail to reorganize their representations toward the requested modality will systematically under-report that modality, so the V-Measure gap can identify failing model–dataset combinations before behavioral testing.","Scaling a single router head found on Pascal VOC improves target-modality accuracy on most other datasets tested, so modality-selection ability is at least partly portable through specific heads.","Amplifying an image or caption promotion head helps reporting of that modality and hurts reporting of the other, indicating these heads implement a fixed modality prior.","Because the salience gap correlates with accuracy at $r = 0.94$, representational-salience measurements can serve as a proxy for behavior in settings where evaluating many prompts is expensive."],"supporting_citations":[{"why":"Concurrent finding that VLMs show 'blind faith' in text; this paper's reciprocal image-bias results are framed against it.","marker":"Deng et al., 2025"},{"why":"Shows attention heads mediate between parametric knowledge and in-context information, the method this paper adapts to modality conflicts.","marker":"Yu et al., 2023"},{"why":"Defines V-Measure, the metric used to quantify how saliently each modality's classes appear in the clustered representations.","marker":"Rosenberg and Hirschberg, 2007"},{"why":"Supplies CIFAR-10 and CIFAR-100 image–class pairs from which conflicting captions are constructed.","marker":"Krizhevsky et al., 2009"},{"why":"Supplies Pascal VOC, the dataset used to discover and characterize the router and promotion heads.","marker":"Everingham et al., 2010"},{"why":"Supplies CUB-Color with primary-color attribute labels used for attribute-level image–caption conflicts.","marker":"Wah et al., 2011"},{"why":"Provides Qwen2.5-VL, the model on which the attention-head intervention and cross-dataset generalization tests are run.","marker":"Bai et al., 2025"},{"why":"Provides LLaVA-1.5, one of the four models evaluated in the behavioral comparison.","marker":"Liu et al., 2023"}],"fun_headline_variants":["Attention heads decide which modality a VLM believes","One attention head can flip a VLM's trusted modality","VLMs' modality bias traced to specific attention heads","Which modality a VLM trusts? Check its attention heads","Modality conflict in VLMs resolved by router attention heads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that clustering the model's final-layer hidden states measures an internal cause of the model's answer rather than simply echoing the answer the model is about to produce.","fun_headline_variants_meta":{"raw":{"variants":["Attention heads decide which modality a VLM believes","One attention head can flip a VLM's trusted modality","VLMs' modality bias traced to specific attention heads","Which modality a VLM trusts? Check its attention heads","Modality conflict in VLMs resolved by router attention heads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1290,"prompt_tokens":952,"completion_tokens":338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":260}},"tokens_in":568,"tokens_out":338,"duration_ms":4134,"temperature":1.0,"reasoning_tokens":260,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:43:08.243106+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On Qwen2.5-VL with CIFAR-100, compute the V-Measure gap from an intermediate layer that does not feed the output distribution, then scale the router head L11H14 and check whether the mid-layer gap moves with accuracy; if it does not, the final-layer salience measure is reading off the imminent answer instead of explaining the behavior.","supporting_citations":[{"cited_title":"V -measure: A conditional entropy-based external cluster evaluation measure","cited_arxiv_id":null,"evidence_quote":"Defines V-Measure, the metric used to quantify how saliently each modality's classes appear in the clustered representations."}],"review_version":1}