{"id":"dbc0497a-52a0-41a9-ad0f-d30262f88efe","arxiv_id":"2607.16105","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Aggregating a VLM's attention over visual tokens and mapping it back to image patches produces saliency maps that, in a 13-sample deletion test on ChartGemma, appear to localize the visual evidence behind chart answers.","lead":"This paper presents a saliency-map method for vision-language models that shows which parts of a chart a model looked at while generating each word of its answer. The authors validate the maps by deleting the pixels they highlight and measuring how quickly the model's accuracy collapses.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal-faithfulness claim rests on a 13-sample deletion experiment with no error bars; AUC gap vs AG-CAM is not statistically established and baseline is not task-comparable.","rationale":"The reader's formal weakest assumption (Eq. 9 renormalization) is not the most dangerous: rescaling a token's attention vector does not change the relative ranking of pixels for that token, and the §5.2 aggregation already L1-normalizes each token, so absolute visual-attention magnitude does not enter the final deletion ranking. The actual load-bearing gap is evidential: the deletion experiment is tiny, single-model, and lacks uncertainty quantification, and the AG-CAM baseline is not matched. This does not invalidate the method; it means the strong causal claim is conditional on a more adequate evaluation. The reader's verdict of CONDITIONAL remains appropriate, so no change.","tokens_in":8849,"tokens_out":6130,"duration_ms":56420,"concrete_test":"Recompute the deletion evaluation on all VLAT samples correctly answered by ChartGemma (or a bootstrap resample of the reported 13) and report per-sample AUC distributions with 95% bootstrap CIs for the proposed method, AG-CAM, and random deletion, using identical infill and aggregation. If the 0.020-vs-0.070 gap is within the CI overlap, or if the ordering changes under a matched output-token protocol for AG-CAM, the headline claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the attention maps are causally faithful and beat AG-CAM—depends on the §5.2 deletion test on 13 correctly answered VLAT samples, one model (ChartGemma), and no error bars or significance test. Two point AUCs (0.020 vs 0.070) cannot support the comparison; with n=13 the standard error is large and the ordering could easily reverse. The paper's own exclusion of True/False items and incorrect answers further narrows the evidence. The comparison is also confounded: AG-CAM explains input/question tokens, whereas the proposed method explains generated output tokens, so the two maps are not answering the same explanatory question; the lower AUC does not directly establish 'more faithful.' Finally, the aggregate map gives equal weight to every generated token after L1 normalization (Sec. 5.2), so tokens with little real image attention (e.g., function words) can dilute the map; this is an unvalidated design choice that could affect the deletion result. The reader's Eq. 9 concern is secondary because per-token pixel ranking is scale-invariant; the decisive issue is the evidential base.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a gradient-free, attention-based saliency method for vision-language models (VLMs) that aggregates LLM attention over visual tokens across all heads and layers, then maps this attention onto the vision encoder's patch grid to produce per-token and aggregate saliency maps for chart images. The method is evaluated with a deletion test on the VLAT dataset using ChartGemma, reporting an AUC of 0.020 versus 0.070 for AG-CAM, and the authors claim the maps are causally faithful and outperform AG-CAM.","tokens_in":9139,"tokens_out":3577,"duration_ms":33676,"significance":"If validated, the method would be a useful lightweight diagnostic tool for interpreting VLMs on visualization tasks, with the strengths of being gradient-free, requiring no architectural changes or extra inference passes, and providing token-level maps. The paper also ships code, which aids reproducibility. However, the central empirical claim—causal faithfulness and superiority over AG-CAM—rests on a very small deletion experiment with no statistical support, so the significance is currently limited.","major_comments":[{"comment":"The deletion-based evaluation uses only 13 correctly answered, non-True/False VLAT samples with a single model (ChartGemma). No error bars, confidence intervals, or significance tests are reported. The claimed advantage (AUC 0.020 vs 0.070) is based on two point estimates; with n=13 the standard error is large and the ordering could easily reverse. This is the sole evidence for the paper's central claim of causal faithfulness and superiority over AG-CAM, so it is not sufficient. Please add variance estimates (e.g., bootstrap), statistical tests, and ideally a larger evaluation set.","section":"§5.2, Figure 3"},{"comment":"The comparison with AG-CAM is confounded. AG-CAM attributes importance to input/question tokens, while the proposed method attributes to generated output tokens. These answer different explanatory questions: one explains why the model chose a particular answer, the other explains which input regions are salient to the question. The lower AUC therefore does not directly establish that the proposed maps are 'more faithful'; it may reflect that the two methods address different targets. A task-comparable baseline (e.g., a gradient-based method applied to output tokens, or AG-CAM evaluated on the same deletion protocol with matched attribution target) is needed.","section":"§5.2, 'It is worth noting...'"},{"comment":"The method renormalizes each token's attention vector over patches (Eq. 9) and, when aggregating, L1-normalizes each token's distribution so that tokens with diffuse attention contribute equally to tokens with sharply peaked attention. These are unvalidated design choices that directly affect the resulting saliency map: function words or tokens with little image attention can dilute the aggregate map, and the absolute attention magnitude allocated to the image versus the prompt is discarded. The paper provides no ablation or analysis showing that these choices preserve or improve causal faithfulness. Given that the deletion test is the only validation, this is a load-bearing gap.","section":"§5.2, Eq. (9) and 'L1-normalize each token's attention distribution'"}],"minor_comments":[{"comment":"Typo: 'a deletion metrics' should be 'a deletion metric'.","section":"Abstract"},{"comment":"Typo: 'caculated' should be 'calculated'.","section":"Conclusion"},{"comment":"Typo: 'trace calculate' should likely be 'trace and calculate'.","section":"Section 2, Related Work"},{"comment":"Typo: 'Figure 2 shows an example our saliency map' should be 'an example of our saliency map'; also 'VLA T' spacing.","section":"Figure 2 caption"},{"comment":"Typo: 'The dataset is code available' should be 'The dataset and code are available'.","section":"Supplementary Material"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the method is lightweight and fast, but the empirical validation is far below the standard needed to support the causal-faithfulness claim. The 13-sample deletion test without significance testing is the main blocker; this is fixable in revision but requires substantially more evaluation. I would not accept in current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on arXiv:2607.16105.\n\nThe genuinely new piece is not the attention aggregation — that goes back to Abnar and Zuidema — but the application: mapping LLM attention over visual tokens back onto the ViT patch grid to get per-token saliency maps during free-form chart QA. That is a useful, cheap, gradient-free diagnostic, and the paper is honest about most of its moving parts. They reproduce Dong and Crisan's accuracy numbers as a sanity check and release code, which is more than many submissions do.\n\nThe main soft spot is the evidence base for the central causal-faithfulness claim. The deletion test runs on 13 correctly answered VLAT samples, one model (ChartGemma), no error bars, no significance test. Two point AUCs (0.020 vs 0.070) cannot support the conclusion that the method 'more faithfully reflects the focus of the model' — with n=13, the ordering could easily flip. The comparison to AG-CAM is also confounded: AG-CAM explains input/question tokens, while the proposed method explains output tokens, so the two maps are not answering the same explanatory question. A lower AUC in this setup does not directly establish higher faithfulness.\n\nThe Eq. 9 normalization the reader flagged is, I think, secondary — per-token pixel ranking is scale-invariant. More substantive is the aggregate map: after L1-normalizing each token's attention, they average over all generated tokens, giving equal weight to function words and content words. That design choice could dilute the map and is not separately validated. The paper acknowledges the small sample size, but then uses 'confirms' in the results and calls the maps 'causally faithful' in the conclusion. That overstates what 13 samples can show.\n\nThis is still a useful paper for people working on explainability for VLMs and visualization literacy. I would not cite it as evidence of faithfulness, but the method is easy to reimplement and the framing is clear. It deserves a serious referee — the idea is relevant and the baseline setup is conventional — though the referee should ask for a larger evaluation, error bars, and a task-comparable baseline.","headline":"Useful lightweight saliency method for VLM chart QA, but the causal-faithfulness claim rests on a 13-sample deletion test and a confounded baseline comparison.","tokens_in":9567,"tokens_out":3087,"would_cite":false,"duration_ms":25808,"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":"A gradient-free method turns a VLM's attention into saliency maps that pinpoint the chart regions driving each generated answer token.","keywords":["saliency maps","vision-language models","visualization literacy","attention aggregation","chart question answering","explainability","deletion test","interpretability"],"falsifier":"A direct comparison of the deletion-test AUC for the proposed method versus a variant that uses raw (non-renormalized) attention magnitudes, on a larger sample of chart questions from a public benchmark, would settle whether the renormalization step preserves faithful rankings; if the raw-magnitude variant performs differently or better, the current method's faithfulness claim is undermined.","tokens_in":8770,"feed_emoji":"📊","tokens_out":2774,"duration_ms":29297,"temperature":0.7,"pith_summary":"This paper introduces a saliency mapping technique for vision-language models (VLMs) that converts the model's internal attention weights into spatial heatmaps over chart images. For each word the model generates in response to a chart question, the method aggregates attention from that token across all attention heads and layers, then maps it back to the image patch grid. The authors claim these maps are causally faithful, meaning they correctly identify the image regions the model actually relies on. They test this by progressively deleting the most salient pixels and measuring how quickly the model's accuracy drops, comparing against a gradient-based baseline. If true, the method offers a lightweight, real-time way to inspect whether VLMs are reading charts correctly, not just whether they output the right answer.","feed_headline":"Attention maps show where VLMs look when reading charts","feed_subtitle":"A gradient-free method traces each answer token back to image patches, beating gradient saliency in deletion tests.","key_machinery":"The central object is the token-to-patch attention projection. For each generated answer token, the method extracts the attention row from that token to all visual prefix tokens, averages these vectors across all attention heads and layers, L1-normalizes them to a distribution over image patches, reshapes the flat patch vector into a 2D grid, and bilinearly upsamples it to image resolution. This projection is what links the language model's internal attention dynamics to spatial regions of the input chart, producing per-token and aggregate saliency maps without any gradient computation.","core_discovery":"The paper's central claim is that the language model's attention from a generated token to the visual prefix, averaged over all heads and layers and reshaped into the vision encoder's patch grid, produces a saliency map that accurately reflects which image regions causally drive the model's answer. The authors validate this with a deletion test on a chart-specialized VLM, showing that removing the top-ranked pixels collapses accuracy far faster than removing random pixels, and faster than a gradient-based attention attribution baseline (area under the curve 0.020 vs. 0.070). This demonstrates, within the paper's setup, that attention-based saliency can serve as a faithful and efficient expla","pith_inferences":["The method's reliance on raw attention averages may inherit known limitations of attention as explanation, such as ignoring the actual information flow through the residual stream; a rollout-based aggregation might alter the maps but was not tested here.","The deletion evaluation is performed on only 13 correctly-answered samples, so the reported AUC difference may not generalize; a larger-scale study across more samples and models would be a natural next step.","The renormalization step (making each token's patch attention sum to 1) discards the absolute magnitude of attention allocated to the image versus the text prompt, which could matter when the model largely ignores the image — an implicit assumption that could be tested directly.","The method's framing around chart images suggests it could extend to other structured visual inputs like diagrams or tables, where understanding which visual element drives an answer is equally important."],"forward_implications":["If the method is faithful, it provides a real-time diagnostic tool for chart question answering, letting users see whether a VLM is looking at the correct bars, axes, or labels when generating an answer.","The steep accuracy collapse observed during deletion indicates that chart reasoning in this model depends on a very small set of critical pixels, which has implications for robustness and adversarial vulnerability.","Because the method works over generated output tokens rather than input tokens, it can explain free-form, open-ended answers, not just classification choices.","The approach is gradient-free and requires no architectural changes, making it applicable to any VLM that uses a visual prefix token sequence with self-attention.","It offers a complement to standard accuracy-based benchmarks: even a model that answers correctly could be flagged if its saliency maps point to unrelated image regions."],"fun_headline_variants":["Attention maps trace VLM chart reading to image patches","Gradient-free saliency shows where VLMs look in charts","Deletion test: attention saliency beats gradient baselines","Chart VLM attention: causal maps without gradients"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Each token's attention vector is renormalized over the image patches, erasing how much of the model's total attention was directed at the image versus the text prompt, and the paper provides no test of whether this discarded magnitude carries causal information.","fun_headline_variants_meta":{"raw":{"variants":["Attention maps trace VLM chart reading to image patches","Gradient-free saliency shows where VLMs look in charts","Deletion test: attention saliency beats gradient baselines","Chart VLM attention: causal maps without gradients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1369,"prompt_tokens":675,"completion_tokens":694,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":419,"completion_tokens_details":{"reasoning_tokens":638}},"tokens_in":419,"tokens_out":694,"duration_ms":5837,"temperature":1.0,"reasoning_tokens":638,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:18:09.670271+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct comparison of the deletion-test AUC for the proposed method versus a variant that uses raw (non-renormalized) attention magnitudes, on a larger sample of chart questions from a public benchmark, would settle whether the renormalization step preserves faithful rankings; if the raw-magnitude variant performs differently or better, the current method's faithfulness claim is undermined.","supporting_citations":[],"review_version":1}