{"id":"0aaceeab-e475-48bf-b450-53ce4d73006c","arxiv_id":"2506.18985","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A gradient-attention explainability method produces sequence-level visual and textual saliency maps for free-form answers from large vision-language models, with stronger human-attention alignment and faithfulness than prior baselines.","lead":"GLIMPSE is a new method that creates a single heatmap showing which parts of an image and prompt a large vision-language model actually used when generating an answer. It could make large visual question-answering systems more inspectable, helping people see where models look, hallucinate, or rely on biased cues.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline alignment gain rests on a depth-prior temperature tuned on the same VQA-HAT subset that is used for evaluation; removing that prior collapses NSS to -0.210, so the SOTA claim is not yet established without held-out validation.","rationale":"The Reader's weakest assumption identifies the depth-based prior in Eq. (10) as load-bearing, and the ablation evidence supports this: removing it flips NSS from strongly positive to negative. I agree that the central risk is test-set tuning of this prior. The full text adds supporting red flags: the paper explicitly reports the best-performing hyperparameter configuration, does not report the filtered sample size, tests only one backbone, and in Sec. 5.2 concedes that the comparison to [26] involves a potentially different subset. The method itself is plausible and the ablation study is informative, and the faithfulness experiments provide some internal support, but the strongest public claim, state-of-the-art human-attention alignment, is not yet separable from benchmark-specific hyperparameter selection. The reader's CONDITIONAL verdict is appropriate, and I do not see a reason to move it: the missing evidence is empirical, not a derivation error.","tokens_in":12267,"tokens_out":3825,"duration_ms":44031,"concrete_test":"Split the filtered VQA-HAT subset (open-ended questions with at least three annotators) into three folds. For each fold, select lambda_d in Eq. (10), the head-fusion temperature lambda in Eq. (6), and the flow strength lambda_f in Eq. (22) using only the training portion, then evaluate NSS and rank correlation on the held-out portion. Repeat the same fixed hyperparameters on a second backbone, e.g., LLaVA-1.5 or Qwen2-VL-7B. If held-out NSS with selected lambda_d remains close to 1.0 and the optimal lambda_d is stable across folds and backbones, the tuning concern is resolved; if the gain over TMME (last 12 layers) shrinks substantially or the optimal lambda_d moves, the headline alignment claim is test-set-specific.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that GLIMPSE achieves state-of-the-art human-attention alignment (NSS 1.014, rank 0.250 in Table 2). That claim is carried by the depth-based prior in Eq. (10), s_l = exp(lambda_d(l+1))/sum exp(lambda_d(k+1)), with lambda_d = 0.2. Table 4 shows this is not a mild ingredient: removing depth weighting collapses NSS from 1.014 to -0.210, and changing lambda_d to 0.5 or 1.0 lowers NSS to 0.911 and 0.883. The paper states that it 'report[s] the hyperparameter configuration that achieves the best quantitative performance,' which is selection on the same evaluation subset, and no validation split is described. The sample size after restricting to open-ended questions with at least three annotators is not reported, and all experiments use a single backbone (Qwen2.5-VL 32B). The paper itself acknowledges in Sec. 5.2 that its comparison to a previous negative alignment result 'uses a different backbone and may involve a different subset of VQA-HAT.' Because the depth prior is both hand-chosen and strongly performance-determining, the reported +71.5% NSS and +46.2% rank improvements over TMME could be an artifact of test-set tuning rather than evidence of a general, model-agnostic explanation principle. This is a correctness-risk concern about the empirical claim, not an internal inconsistency in the derivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GLIMPSE, a three-stage saliency attribution method for autoregressive large vision-language models: it fuses gradient-weighted attention across heads, propagates relevance through layers with an adaptive depth-dependent weighting, and combines token-level prompt alignment, visual grounding, and softmax confidence into a holistic response-level saliency map. The method is evaluated on VQA-HAT human-attention alignment and on deletion/insertion faithfulness using Qwen2.5-VL 32B. The authors report state-of-the-art alignment (NSS 1.014 and rank correlation 0.250) and faithfulness close to human attention, and they present qualitative cases for token-level grounding, hallucination diagnosis, and bias analysis.","tokens_in":12603,"tokens_out":6791,"duration_ms":69895,"significance":"If the empirical claims hold, GLIMPSE is a valuable lightweight explainability tool for free-form LVLM responses. The method is well motivated, the three components are clearly described, and the ablations are informative. The large margins over baseline explainers in both human-alignment and faithfulness experiments are encouraging, and the qualitative diagnostics demonstrate a plausible use case. However, the headline alignment result is not yet established: the depth-prior temperature is selected on the same evaluation subset, the evaluation subset size is not reported, and the experimental protocol for optional preprocessing cues is ambiguous. These issues are fixable but prevent accepting the state-of-the-art claim on the current evidence.","major_comments":[{"comment":"The central alignment claim is not yet established because the depth-prior temperature lambda_d is selected on the same evaluation subset used for the headline numbers. The text states that the paper reports the hyperparameter configuration that achieves the best quantitative performance, and Table 4 shows that removing the depth weighting collapses NSS from 1.014 to -0.210 and rank correlation from 0.250 to -0.167, while lambda_d = 0.5 and 1.0 give NSS 0.911 and 0.883. Since no validation split or cross-validation is described, the +71.5% NSS improvement over TMME (last 12 layers) could be an artifact of test-set tuning rather than evidence of a general explanation principle. Please add a validation protocol (for example, a separate split or cross-validation for lambda and lambda_d) and report performance on held-out data with fixed hyperparameters, or at least a systematic sensitivity analysis across multiple random splits.","section":"Sec. 4.3, Eq. (10), Table 4"},{"comment":"The number of VQA-HAT samples used in the evaluation is not reported. The paper only states that the QA set was restricted to open-ended questions with at least three annotator maps. The reported standard errors (for example, NSS 1.014 +/- 0.032) cannot be interpreted without the sample size, and the comparison between GLIMPSE and TMME (last 12 layers) is not accompanied by a statistical test. Please report the subset size, the retention rate relative to the full VQA-HAT set, and a paired significance test for the main alignment and faithfulness comparisons.","section":"Sec. 4.1"},{"comment":"The paragraph describing optional evaluation cues is ambiguous about whether those cues were used in Tables 2 and 3. Removing punctuation from the generated response changes the token sequence and therefore the gradient-based attribution; adding a system instruction cue changes the prompt; and applying a Gaussian blur alters the saliency map post hoc. If any of these were applied when producing the reported numbers, the baseline explainers were not evaluated on identical inputs, and the comparison is no longer on the standard VQA-HAT setting. Please state explicitly which configuration was used for each table, and if these cues were used, rerun all baselines under the same preprocessing and report the unmodified configuration as well.","section":"Sec. 4.3, paragraph after Table 4"},{"comment":"All experiments use a single backbone, Qwen2.5-VL 32B. The paper claims a model-agnostic framework and state-of-the-art alignment for LVLM interpretability, but no evidence is provided that the method transfers to another architecture or scale. A single additional backbone (for example, a LLaVA-style model or a smaller Qwen-VL variant) with the same hyperparameters would substantially strengthen the generality claim; without it, the 'model-agnostic' claim is undersupported.","section":"Sec. 4.1 and contribution 1 (model-agnostic claim)"}],"minor_comments":[{"comment":"There is a numeric inconsistency: the text states that using the last 30% of layers without depth weighting reduces NSS to 0.490, but Table 4 reports 0.670 for that configuration.","section":"Sec. 4.3, Table 4 and surrounding text"},{"comment":"The NSS computation uses the set of human-attention locations above the 95th percentile threshold rather than actual fixation locations; this is a non-standard NSS variant and should be justified, with a sensitivity analysis for the threshold theta.","section":"Eq. (24)"},{"comment":"The faithfulness results are reported without standard errors or significance tests; given the small gap at 5% insertion AUC between GLIMPSE (0.134) and human attention (0.149), the claim that GLIMPSE 'closely matches' human attention needs statistical support.","section":"Table 3"},{"comment":"The comparison to the -0.20 rank correlation of reference [26] is presented as supporting evidence, but the paper itself acknowledges the different backbone and possibly different VQA-HAT subset; the abstract and introduction should qualify the state-of-the-art claim to the specific single-backbone and subset setting.","section":"Sec. 5.2"},{"comment":"There are several copyediting issues, including the subject-verb disagreement 'Recent large vision-language models (LVLMs) has advanced' in the abstract and 'which is only applied to Ours method' in the Figure 1 caption.","section":"Abstract and Figure 1 caption"},{"comment":"The notation m in {V, P} overloads the index-set symbols V and P as modality labels; this is understandable but should be clarified to avoid confusion between the target modality and the token index sets.","section":"Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good candidate for publication if the authors can resolve the test-set selection concern, report the evaluation subset size, and clarify the preprocessing protocol. The depth-prior sensitivity is striking enough that I would not accept the state-of-the-art claim on the current evidence. I would also encourage the editor to ask the authors to release the exact subset construction and code, since the VQA-HAT subsetting choices directly affect the headline numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GLIMPSE is a genuinely useful combination—gradient-weighted attention, depth-aware propagation, confidence-weighted token aggregation—targeted at a real gap: explaining whole free-form LVLM responses rather than single tokens or phrases. If the empirical claims hold, it gives practitioners a cheap audit tool. I think the paper is worth engaging seriously, but the headline alignment numbers are not yet established.\n\nWhat is good: the method is lightweight and model-agnostic, one forward pass plus one backward pass per token. The three stages are clear and each is ablated. The faithfulness experiments (deletion/insertion AUC) are well designed and show GLIMPSE near human attention, which is stronger evidence than alignment alone. The paper also acknowledges that human-alignment misses valid alternative cues and that the comparison to the prior negative result uses a different backbone and possibly a different subset. That honesty is real.\n\nThe soft spots are concentrated in the evaluation protocol. The paper states it reports \"the hyperparameter configuration that achieves the best quantitative performance,\" and the load-bearing configuration includes lambda_d = 0.2 in the depth prior. Table 4 shows that removing depth weighting collapses NSS from 1.014 to -0.210, and changing lambda_d to 0.5 or 1.0 lowers it to 0.911 or 0.883. With no validation split described and no second backbone, the +71.5% NSS and +46.2% rank improvements could simply be test-set tuning. I do not think this is fabrication or an internal inconsistency; it is selection leakage, and it is fixable. We also never learn how many VQA-HAT samples survive the open-ended/three-annotator filter, and the optional system-prompt cue, punctuation removal, and Gaussian blur are mentioned but it is unclear whether they were active in the reported runs. Those details should be mandatory in the paper.\n\nThis paper is for XAI researchers and LVLM safety people who want a fast way to trace a model's full answer back to image regions. It deserves peer review: a serious referee should ask for code, a held-out validation split, the filtered sample size, and at least one additional backbone. I would not desk-reject it, but neither would I cite it until the code and held-out validation appear.","headline":"GLIMPSE is a useful combination of known ingredients aimed at a real gap, but the headline alignment numbers are not yet safe because the depth prior is tuned on the evaluation subset.","tokens_in":13074,"tokens_out":2531,"would_cite":false,"duration_ms":26919,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GLIMPSE attributes an entire LVLM response to the visual and textual evidence that drove it, reporting human-gaze alignment scores well above earlier explainers.","keywords":["explainable AI","large vision-language models","visual question answering","saliency maps","attention gradients","layer-wise relevance propagation","human attention alignment","hallucination diagnosis"],"falsifier":"Fix all GLIMPSE hyperparameters except the depth temperature, then estimate $\\lambda_d$ on a training split of VQA-HAT and evaluate on a held-out split (or on a second human-attention benchmark); if the best $\\lambda_d$ is far from 0.2, or if NSS with $\\lambda_d=0.2$ falls below the TMME last-12-layer baseline, the reported state-of-the-art alignment is an artifact of tuning rather than a model-agnostic depth prior.","tokens_in":12053,"feed_emoji":"👁️","tokens_out":15909,"duration_ms":135368,"temperature":0.7,"pith_summary":"This paper introduces GLIMPSE, a saliency method that explains an entire free-form answer produced by a large vision-language model, not just a single predicted token or phrase. The paper's central claim is that fusing positive attention gradients with attention weights, propagating the fused relevance through all layers with a depth-based weighting, and re-weighting each generated token by confidence and cross-modal alignment yields one holistic heatmap that is both close to human gaze and faithful to the model's own behavior. That matters because interpretability in generative vision-language models has lagged behind their capabilities, and a lightweight, model-agnostic heatmap could turn hallucination and bias from black-box anomalies into traceable failures. On an open-ended subset of VQA-HAT, the paper reports Normalized Scanpath Saliency of 1.014, rank correlation 0.250, and deletion AUCs of 0.855, 0.718, and 0.617 at 5/15/30 percent perturbation, close to the corresponding human-attention values.","feed_headline":"GLIMPSE maps a whole VQA answer onto the image evidence behind it","feed_subtitle":"Gradient-attention fusion lifts human-gaze alignment to NSS 1.014 and exposes hallucinated tokens at the patch level.","key_machinery":"The load-bearing object is the gradient-fused layer relevance matrix $E_\\ell = \\sum_{h} w_h^\\ell \\mathrm{ReLU}(g_h^\\ell \\odot A_h^\\ell)$, where $A_h^\\ell$ is the attention matrix of head $h$ in layer $\\ell$ and $g_h^\\ell$ is its gradient with respect to the generated-token logit. Relevance is propagated by the additive update $R \\leftarrow R + (I + \\alpha_\\ell E_\\ell)R$, with layer weights $\\alpha_\\ell \\propto g_\\ell s_\\ell$ combining each layer's gradient norm with the exponential depth prior $s_\\ell = \\exp(\\lambda_d(\\ell+1)) / \\sum_k \\exp(\\lambda_d(k+1))$. This additive accumulation avoids the noise buildup of full-matrix Attention Rollout, while the depth prior (temperature $\\lambda_d=0.2$) suppresses early-layer noise; per-token weights $\\beta_t$ then aggregate the maps into one response-level heatmap.","core_discovery":"GLIMPSE's central claim is that a holistic, response-level saliency map for an autoregressive LVLM can be computed by treating each generated token as an explanation target: compute the layer-wise gradient-fused attention matrix $E_\\ell$ from positive gradients and attention, accumulate relevance across layers with the update $R \\leftarrow R + (I + \\alpha_\\ell E_\\ell)R$, and aggregate per-token maps with weights $\\beta_t$ that multiply softmax confidence by alignment with the complementary modality. The paper reports that this design reaches NSS 1.014 ± 0.032 and rank correlation 0.250 ± 0.008 on an open-ended VQA-HAT subset, a 71.5% NSS gain and 46.2% rank-correlation gain over the strongest baseline (TMME restricted to its last 12 layers), while deletion AUC matches human attention closely (0.855/0.718/0.617 at 5/15/30% perturbation vs. 0.852/0.707/0.589 for human maps). It further claims that token-level relevancy scores localize individual objects in generated text, expose a foreground bias, and trace hallucinated words to image regions that do not support them.","pith_inferences":["Editorial inference: the same depth-prior mechanism should transfer to video question answering if the prior reflects a general property of deep autoregressive stacks; applying GLIMPSE to frame-based video tokens and checking whether $\\lambda_d=0.2$ still outperforms a re-tuned value would test that.","Editorial inference: a neighbouring problem the paper does not pursue is using the joint token relevance $\\gamma_t$ as a training signal; the same score that diagnoses hallucination after the fact could regularize decoding or fine-tuning to suppress low-grounded tokens.","Editorial inference: the contrast with the negative human-alignment rank correlation reported by a perturbation-based explainer suggests that earlier claims of intrinsic misalignment between LVLMs and human gaze may have measured the explainer rather than the model; a direct comparison on identical backbone and identical data subset would settle that."],"forward_implications":["If the central claim holds, one heatmap can audit a complete multi-sentence answer, so practitioners can check which image region and which prompt words actually drove the response instead of inspecting token-by-token maps.","Because the method needs only one forward pass plus one backward pass per generated token, it can be attached to any autoregressive vision-language model without retraining or architecture changes.","The deletion and insertion results imply that the highlighted patches are causally load-bearing: deleting them lowers the model's confidence sharply and inserting them raises it, so the map reflects evidence the model truly used.","Token-level relevance scores provide a hallucination diagnostic: hallucinated words carry low confidence and point at image regions without supporting features, allowing failures to be localized to specific visual evidence."],"supporting_citations":[{"why":"Supplies the gradient-weighted attention product and additive relevance propagation that GLIMPSE builds on.","marker":"[8]"},{"why":"Defines Attention Rollout, the full-matrix propagation baseline GLIMPSE replaces to avoid noise buildup.","marker":"[1]"},{"why":"Provides the VQA-HAT human-attention dataset used for the alignment benchmark.","marker":"[9]"},{"why":"Is the Qwen2.5-VL backbone used in all reported experiments.","marker":"[5]"},{"why":"Provides the Grad-CAM gradient-based baseline against which GLIMPSE is compared.","marker":"[20]"},{"why":"Documents the gradient-shattering problem that motivates the depth-based layer weighting.","marker":"[6]"},{"why":"Reports the negative human-alignment result from a perturbation-based LVLM explainer that GLIMPSE contrasts.","marker":"[26]"},{"why":"Supplies the layer-wise relevance propagation background that the method's conservative propagation design responds to.","marker":"[4]"}],"fun_headline_variants":["GLIMPSE links a full VQA answer to the exact image evidence","One map explains an entire LVLM response, not just single tokens","GLIMPSE traces every word of an answer back to supporting pixels","Holistic cross-modal explainability: GLIMPSE sees the whole reasoning","GLIMPSE reveals where an LVLM looks for its whole answer, at once"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline result rests on a hand-set depth schedule (temperature 0.2) that makes the deepest layers dominate the explanation; if that schedule was tuned to the human-gaze evaluation set rather than reflecting a general property of deep LVLMs, the alignment gains would not transfer to other models or benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["GLIMPSE links a full VQA answer to the exact image evidence","One map explains an entire LVLM response, not just single tokens","GLIMPSE traces every word of an answer back to supporting pixels","Holistic cross-modal explainability: GLIMPSE sees the whole reasoning","GLIMPSE reveals where an LVLM looks for its whole answer, at once"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1602,"prompt_tokens":965,"completion_tokens":637,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":541}},"tokens_in":581,"tokens_out":637,"duration_ms":6610,"temperature":1.0,"reasoning_tokens":541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:40:46.964070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix all GLIMPSE hyperparameters except the depth temperature, then estimate $\\lambda_d$ on a training split of VQA-HAT and evaluate on a held-out split (or on a second human-attention benchmark); if the best $\\lambda_d$ is far from 0.2, or if NSS with $\\lambda_d=0.2$ falls below the TMME last-12-layer baseline, the reported state-of-the-art alignment is an artifact of tuning rather than a model-agnostic depth prior.","supporting_citations":[{"cited_title":"Generic attention- model explainability for interpreting bi-modal and encoder- decoder transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-weighted attention product and additive relevance propagation that GLIMPSE builds on."},{"cited_title":"Quantifying attention flow in transformers","cited_arxiv_id":null,"evidence_quote":"Defines Attention Rollout, the full-matrix propagation baseline GLIMPSE replaces to avoid noise buildup."},{"cited_title":"Human attention in visual question answer- ing: do humans and deep networks look at the same regions? In Proc","cited_arxiv_id":null,"evidence_quote":"Provides the VQA-HAT human-attention dataset used for the alignment benchmark."},{"cited_title":"Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba- tra","cited_arxiv_id":null,"evidence_quote":"Provides the Grad-CAM gradient-based baseline against which GLIMPSE is compared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the gradient-shattering problem that motivates the depth-based layer weighting."},{"cited_title":"Attention, Please! PixelSHAP Reveals What Vision-Language Models Actually Focus On","cited_arxiv_id":"2503.06670","evidence_quote":"Reports the negative human-alignment result from a perturbation-based LVLM explainer that GLIMPSE contrasts."},{"cited_title":"On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation","cited_arxiv_id":null,"evidence_quote":"Supplies the layer-wise relevance propagation background that the method's conservative propagation design responds to."}],"review_version":2}