{"id":"016dd007-5e15-42dc-aa16-2269528ac86a","arxiv_id":"2505.17812","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"VaLSe uses attention-based visual contribution maps to steer an LVLM's latent features toward visually grounded content, reducing object hallucinations on several benchmarks while exposing flaws in CHAIR-style evaluation.","lead":"VaLSe is a method that shows which parts of an image a vision-language model uses when it says specific words, and then nudges the model's internal state toward those visual regions to reduce made-up objects. The paper also argues that current hallucination benchmarks can flag correct answers as hallucinations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Contribution-map fidelity is under-validated: Eq. 5 maps are tested on only 8 non-monotone samples, artifact removal is unspecified, and Appendix B concedes spatial correspondence fails for MiniGPT-4/Qwen2-VL, yet those models still show gains.","rationale":"The reader's weakest assumption is exactly the hinge: the fidelity of the Eq. 5 contribution maps is what connects the interpretability claim to the steering results. I agree with that assessment, and my pass sharpens it with two concrete observations already in the paper. First, Appendix F's validation is far too small and is self-reported as non-monotone, so the maps' localization quality is not established. Second, Appendix B states that for MiniGPT-4 and Qwen2-VL the required spatial correspondence is destroyed, yet Table 2 still reports gains on those models; if the maps are known to be spatially corrupted there, then either the maps are not causally load-bearing for mitigation or the interpretation claim is model-specific and needs separate support. I credit the random-masking ablation in Table 7 as real evidence that relevance-guided masking does something beyond random corruption, and the method is training-free and reproducible in principle, so this is not a basis for rejection. However, the central claim's strongest form—that VaLSe produces faithful visual contribution maps and that those maps drive the hallucination reduction—remains conditional pending a ground-truth localization audit and a test that replaces VaLSe maps with ground-truth masks. This does not move the reader's CONDITIONAL verdict; it specifies what would move it to ACCEPT or REJECT.","tokens_in":20924,"tokens_out":5570,"duration_ms":66837,"concrete_test":"Localization audit with ground truth: sample 100 COCO images, select visual-sensitive tokens via Eq. (4), and score each VaLSe map against the referred object's ground-truth mask using the Pointing Game (or IoU at a fixed threshold), alongside raw attention rollout and vision-encoder Grad-CAM baselines. Then rerun the full CHAIR pipeline once with positive samples built from VaLSe maps and once with positive samples built from ground-truth object masks at the same masking ratio. If VaLSe maps are not significantly above baselines or chance, or if GT-mask steering reproduces the 48.7 to 36.2 CHAIR drop within noise, the interpretation-to-steering link is not demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that VaLSe traces visual influence and steers by it rests on Eq. 5: a gradient-weighted attention rollout whose high values are assumed to localize the image evidence actually used for a token. The paper does not establish this. Appendix F validates the maps on only 8 samples and itself states that the deletion/insertion curves 'do not exhibit a consistent trend,' so the map-quality evidence is anecdotal. Artifact elimination is described only qualitatively ('identify positions P exhibiting artifacts') and deferred to the supplementary materials. More importantly, Appendix B concedes that for Q-former/pixel-shuffle or multi-scale encoders (MiniGPT-4, Qwen2-VL), the spatial correspondence required by Eq. 5 is destroyed, yet Table 2 still reports VaLSe improvements on those models. That is consistent with the mitigation gains coming from the masking/SVD procedure rather than from faithful localization. Since the same maps define the positive samples for the steering direction (Sec. 3.2), weak map fidelity decouples the interpretation claim from the mitigation result: the CHAIR drop could arise from any relevance-weighted image contrast, not from 'vision-aware' grounding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VaLSe, a training-free framework for object-hallucination mitigation and interpretability in LVLMs. The method selects visual-based tokens via a log-likelihood ratio (Eq. 3), computes visual contribution maps through gradient-weighted attention rollout (Eq. 5), removes activation artifacts by contrasting with special tokens, and uses these maps to mask low-contribution image regions and construct positive/negative feature pairs. Singular value decomposition of the feature differences provides a steering vector, which is added to LLM hidden states during inference. The authors report reduced hallucination on CHAIR, AMBER, POPE, MMHal, and MMVP, with preserved general ability on MME, GQA, and LLaVA-Bench, and they provide qualitative case studies arguing that existing CHAIR labels can be false positives. The main claims are that VaLSe produces faithful visual contribution maps and that steering by these maps reduces object hallucination.","tokens_in":21147,"tokens_out":4893,"duration_ms":55497,"significance":"If the central claim is upheld, VaLSe would be a useful contribution: it is training-free, provides a unified interpretation-and-mitigation pipeline, and reports a substantial CHAIR improvement on LLaVA-1.5 (sentence-level hallucination from 48.7 to 36.2 in Table 2). The random-masking ablation in Table 7 is a genuine attempt to isolate the contribution of relevance-guided masking, and the inclusion of code is a practical strength. The preservation of general performance on GQA, LLaVA-Bench, and MME is also encouraging. However, the paper's two central claims—map fidelity and the causal role of those maps in mitigation—are coupled, and the current evidence for map fidelity is thin. The significance of the work therefore depends on the authors' ability to decouple the interpretation and mitigation claims and to validate the contribution maps more rigorously than the current 8-sample evaluation.","major_comments":[{"comment":"The visual contribution maps are the foundation of both the interpretability and the steering mechanism, yet their fidelity is validated on only 8 samples, and the paper itself states that the deletion/insertion curves 'do not exhibit a consistent trend.' This is load-bearing because the same maps are used to construct the positive samples for the steering direction in §3.2; if the maps do not accurately localize the image evidence actually used by the model, then the steering direction may be driven by arbitrary masking artifacts rather than by vision-aware grounding. Please provide a more extensive quantitative validation of Eq. (5), for example a pointing-game comparison against segmentation masks or a larger deletion/insertion study with confidence intervals, and disclose the artifact-elimination details that are currently deferred to the supplementary materials.","section":"§3.2, Eq. (5), and Appendix F"},{"comment":"Appendix B concedes that for MiniGPT-4 and Qwen2-VL, the Q-former, pixel-shuffle, or multi-scale encoders destroy the spatial correspondence required by Eq. (5). Yet Tables 1 and 2 still report VaLSe gains on these models. This creates a tension: if the contribution maps are unreliable for these architectures, the reported improvements cannot be attributed to the claimed vision-aware localization mechanism. To make the central claim defensible, either restrict the vision-aware interpretation claim to spatially aligned architectures such as LLaVA-1.5, or provide evidence on these models that the gains are mediated by the maps—for example, by comparing steering directions derived from contribution maps with direction derived from random or constant masks on the same models.","section":"§3.2 and Appendix B"},{"comment":"The description of positive-sample construction is internally inconsistent. Section 3.2 says that for each sample the method masks p% of image patches with low contribution values, and the ablations in Tables 4 and 5 vary p. Appendix D.2, however, says 'Instead of masking a fixed percentage p, we adopt an adaptive strategy by masking all tokens whose relevance scores are below the mean relevance value.' This ambiguity makes the method unreproducible and blurs what the p-ablation actually demonstrates. Please specify the exact masking rule used in the main experiments and in each ablation table.","section":"§3.2 and Appendix D.2"},{"comment":"The intervention strength β is tuned per model and per benchmark (e.g., β=0.5 for LLaVA-1.5 on CHAIR and AMBER but β=0.4 on other experiments; β=0.2 for Qwen2-VL on MMVP and MME but β=0.5 on other experiments), and α is set per model. Because the headline claim is that VaLSe 'outperforms all compared methods' on CHAIR, this per-benchmark tuning poses a risk of overfitting to the specific evaluation. Please report results under a single fixed hyperparameter configuration, or provide a sensitivity analysis that covers the full range of β and α on all benchmarks, so that the reader can assess the robustness of the reported gains.","section":"Appendix D.2"},{"comment":"The theoretical justification relies on the assumptions f(A(noise))=0 and that the remainder R is negligible, but neither is justified. For real LVLMs, replacing the image with noise does not generally zero out attention matrices, and the higher-order terms in the Taylor expansion are not controlled. As written, Eq. (8) is a heuristic analogy rather than a proof. If the purpose is only to provide intuition, please state this explicitly; otherwise, the argument needs to be made rigorous or removed from the central narrative.","section":"§3.3, Eq. (8)"}],"minor_comments":[{"comment":"The artifact-elimination procedure is described only qualitatively ('identify positions P exhibiting artifacts') and is deferred to the supplementary materials. Please provide the full procedure in the main text or appendix, since it is essential for reproducing the contribution maps.","section":"§3.2, 'Artifacts Elimination'"},{"comment":"The 'noise image' eI is not defined. Please specify its distribution, size, and how it is passed through the LVLM in the single forward pass described in the text.","section":"§3.2, Eq. (3)"},{"comment":"Figure 3 reports MME results in a plot without numeric values. Since the text claims 'improved performance in color and positional understanding' and 'notable gains in OCR and code-related tasks,' please provide the exact numerical scores for all MME subtasks.","section":"Figure 3 and Table 3"},{"comment":"The model list includes 'Mipha-3B' in addition to the models evaluated in the main text, but no experimental results for Mipha-3B are reported anywhere. Please clarify its role or remove it.","section":"Appendix D.1"},{"comment":"The four hallucination categories ('truly hallucinated,' 'factual,' 'unclear,' and 'false' hallucinated words) are each supported by a single qualitative example. Please state whether these categories were observed at quantified frequencies on a larger sample, or explicitly present them as illustrative rather than measured taxonomies.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The central weakness is the coupling between the interpretability claim and the mitigation claim: the steering direction is fit from positive samples constructed using the paper's own contribution maps, so a failure of map fidelity would undermine both claims simultaneously. The random-masking ablation in Table 7 is a good step, but it does not validate the maps themselves. I would encourage the editor to ask for a substantially strengthened map-validation study, a clarified masking rule, and a decoupled evaluation of the steering mechanism before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible training-free steering method with a real CHAIR improvement on LLaVA-1.5, and the authors are unusually honest about where it breaks; the big caveat is that the interpretability maps are under-validated, and the mitigation result does not actually require them to be faithful. The novel piece is the closed loop: LLR-based visual token selection, gradient-weighted attention rollout with artifact suppression, and SVD latent steering using relevance-masked images as positive samples. The random-masking ablation (Table 7) is the strongest evidence that the masking matters, and it supports relevance-guided masking over random. They also ship code, report five hallucination benchmarks and three general-ability ones, and include a thoughtful qualitative critique of CHAIR.\n\nWhere it gets soft. First, map fidelity. Eq. 5 maps are validated on 8 samples, and the paper's own Appendix F says the deletion/insertion curves 'do not exhibit a consistent trend.' The artifact-removal procedure is deferred to the supplement. Appendix B concedes that Q-former/multi-scale encoders destroy the spatial correspondence Eq. 5 assumes, yet Table 2 still shows gains on MiniGPT-4 and Qwen2-VL. That means the mitigation works even where the maps are arguably not faithful, which decouples the interpretation claim from the mitigation result. The gains could come from almost any relevance-weighted image contrast; the maps being 'vision-aware' in a localized sense is not established. Second, hyperparameters: alpha, masking ratio, and especially beta are tuned per model and per benchmark (beta varies 0.2-0.5 across settings). That invites overfitting concerns, and Table 2 lacks error bars. Third, there's a real inconsistency in the masking description: Section 3.2 says p% of patches are masked, Appendix D.2 says the implementation uses an adaptive threshold at the mean relevance, and Appendix E ablates p. These need to be reconciled before the results are fully interpretable.\n\nThe central mitigation claim is defensible: the CHAIR drop is large and the random-masking baseline is the right control. The interpretation claim is plausible but not yet proven. The paper deserves a serious referee; it's well above desk-reject level. Recommend sending it out, with requests for better map validation (more samples, monotone curves or an explanation of why non-monotone), a precise artifact-removal description, and a sensitivity analysis of beta. For a reader working on hallucination mitigation it is a useful method-reference; for the interpretability community it's a reminder that tethering mitigation to explanation needs validation of the explanation itself.","headline":"Credible training-free steering with a solid CHAIR gain on LLaVA-1.5, but the interpretability maps are under-validated and the mitigation result does not actually require them to be faithful; still worth a serious referee.","tokens_in":21682,"tokens_out":3154,"would_cite":true,"duration_ms":25411,"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":"VaLSe traces which image regions drive each output word, then steers the model's internal state toward those regions to reduce hallucinated objects.","keywords":["object hallucination","vision-language models","interpretability","attention maps","latent steering","CHAIR benchmark","singular value decomposition","visual grounding"],"falsifier":"Run the same steering pipeline with random masking at the same masking ratios on a large set of images; if random masks reduce CHAIR scores as much as VaLSe's relevance-guided masks, or if deleting the top-ranked regions of a contribution map changes the target word's probability no more than deleting random regions, then the maps are not carrying the causal signal the paper claims.","tokens_in":20709,"feed_emoji":"🔍","tokens_out":9559,"duration_ms":73029,"temperature":0.7,"pith_summary":"VaLSe tries to show that object hallucination in vision-language models can be mitigated after the fact, without retraining, by first making the model's visual reasoning visible. For each word that depends strongly on the image, the method builds a visual contribution map: a heatmap over image regions that raised that word's probability, obtained by propagating relevance through the model's attention layers and removing known activation artifacts. The maps are then used to mask irrelevant parts of the image, the model's internal features with and without those parts are contrasted, and the dominant difference direction is added to the hidden states during generation. The paper reports that this lowers the sentence-level CHAIR hallucination rate of LLaVA-1.5 from 48.7 to 36.2 and beats all compared methods on the 64-token CHAIR setting, while general benchmarks stay roughly level. If the maps are faithful, the same procedure is simultaneously a token-level explanation tool and a training-free fix, and it provides evidence that some words flagged as hallucinated by current metrics are actually grounded in the image.","feed_headline":"Vision-aware latent steering cuts object hallucinations by a quarter","feed_subtitle":"VaLSe maps the image regions behind each word, then shifts internal representations toward them.","key_machinery":"The central object is the visual token contribution map: a per-token heatmap over image tokens built by propagating an identity relevance matrix through all LLM layers with gradient-weighted head averaging, $\\bar{A}^l = \\left(\\sum_h \\frac{\\partial \\mathcal{L}_{y_s}}{\\partial A^l_h} \\odot A^l_h\\right)^+$, and reading off the image-token entries of the last row. Artifact activations are removed by contrasting the target token's map with that of a non-semantic special token, which suppresses the fixed spatial positions that fire regardless of input. The map does two jobs: it defines which image patches are semantically relevant (positive samples are the original image with all sub-mean-relevance patches masked), and it is the basis for the steering signal, since the per-layer difference between positive- and negative-sample MLP features is decomposed by SVD and the top right-singular vector becomes the steering direction added at inference.","core_discovery":"The paper's central claim is that the gradient-weighted attention propagation of Eq. (5), after subtracting artifact activations, yields visual contribution maps that faithfully trace how each image region contributes to a given output token. Those maps identify the visual evidence behind each word, and aggregating them over the selected words defines a semantic mask: positive samples keep only the high-contribution patches, while the original image and response form the negative sample. Contrasting the model's MLP features for positive and negative samples at every layer, then taking the top right-singular vector of the difference via SVD, produces a per-layer steering direction that is added to hidden states at inference. In the paper's experiments this realigns attention toward semantically relevant content and reduces hallucinated outputs on CHAIR, AMBER, POPE, and MMHal for LLaVA-1.5 and Qwen2-VL with little or no loss on MME, GQA, and LLaVA-Bench. The visualization analysis further argues that some words CHAIR labels as hallucinated are actually correct, visually grounded descriptions, which the authors take as evidence that existing OH metrics need revision.","pith_inferences":["I would expect the same interpretation-then-steer recipe to generalize beyond hallucination, for example to typographic attacks or any failure mode in which the model attends to the wrong image region; the paper already shows a typographic-deception case where the maps reveal misdirected attention.","If the maps are taken seriously, hallucination evaluation could be redefined at the pixel level: a word is a true hallucination only when the regions that most raised its probability do not contain the claimed object, rather than when the word is missing from a caption-based ground-truth list.","A decisive and inexpensive test of the map-fidelity premise would be a large-sample comparison of relevance-guided masking against random masking at matched mask ratios.","For Q-former and multi-scale-encoder models, the paper's limitation discussion implies that a token-alignment or feature-resampling step would need to be inserted before the same loop could reach LLaVA-level gains."],"forward_implications":["With VaLSe, LLaVA-1.5's sentence-level CHAIR hallucination rate drops from 48.7 to 36.2, and its 64-token CHAIR score is lower than all compared mitigation methods in the paper's experiments.","Hallucination scores improve on AMBER, POPE, and MMHal for both LLaVA-1.5 and Qwen2-VL, with general-task performance on MME, GQA, and LLaVA-Bench roughly preserved.","The contribution maps make token-level visual grounding visible, allowing users to see which image region drove a given word and to catch cases where CHAIR flags a visually correct word as a hallucination.","Because the whole pipeline is training-free and built from a small set of paired samples, it can be applied to open-source LVLMs by shifting hidden states at inference time.","The method's usefulness is tied to spatial alignment: the paper states it works when visual features enter the LLM through a linear projection that preserves layout, and degrades for architectures such as Q-former or multi-scale encoders that compress or rearrange visual tokens."],"supporting_citations":[{"why":"Supplies the layer-by-layer relevance propagation rule from which Eq. (5) builds the contribution maps.","marker":"[45]"},{"why":"Provides the gradient-weighted head averaging used to aggregate attention heads when computing token contribution maps.","marker":"[46]"},{"why":"Documents the massive-activation artifacts that VaLSe's artifact-elimination step subtracts from the maps.","marker":"[26]"},{"why":"The latent-space steering baseline whose feature-difference and SVD recipe VaLSe adapts for vision-aware steering.","marker":"[21]"},{"why":"Visual contrastive decoding, the decoding-level debiasing method whose first-order expansion is compared with VaLSe's feature-level steering in the analysis.","marker":"[17]"},{"why":"Defines the CHAIR benchmark used for the main hallucination numbers and whose ground-truth incompleteness the visualization analysis critiques.","marker":"[27]"},{"why":"Provides the image set and the claim that object information is spatially localized in LLaVA-style models, grounding the paired-sample construction.","marker":"[63]"},{"why":"Source of the log-likelihood-ratio criterion for selecting visual-based tokens and of the observation that multi-resolution encoders complicate token-to-region alignment.","marker":"[22]"}],"fun_headline_variants":["VaLSe maps image-to-token links to reduce object hallucinations","New latent steering method reveals visual evidence behind each word","VaLSe traces each word to its image region to stop hallucinations","Interpretable latent steering reduces object hallucination in LVLMs","Vision-aware latent steering: map then fix object hallucinations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the contribution map for a word really marks the image regions the model used to generate that word; if the map points to the wrong regions, both the explanations and the steering built from them lose their basis.","fun_headline_variants_meta":{"raw":{"variants":["VaLSe maps image-to-token links to reduce object hallucinations","New latent steering method reveals visual evidence behind each word","VaLSe traces each word to its image region to stop hallucinations","Interpretable latent steering reduces object hallucination in LVLMs","Vision-aware latent steering: map then fix object hallucinations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3056,"prompt_tokens":984,"completion_tokens":2072,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":1998}},"tokens_in":600,"tokens_out":2072,"duration_ms":14438,"temperature":1.0,"reasoning_tokens":1998,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:40:47.219242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same steering pipeline with random masking at the same masking ratios on a large set of images; if random masks reduce CHAIR scores as much as VaLSe's relevance-guided masks, or if deleting the top-ranked regions of a contribution map changes the target word's probability no more than deleting random regions, then the maps are not carrying the causal signal the paper claims.","supporting_citations":[{"cited_title":"Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the layer-by-layer relevance propagation rule from which Eq. (5) builds the contribution maps."},{"cited_title":"Transformer interpretability beyond attention visualization","cited_arxiv_id":null,"evidence_quote":"Provides the gradient-weighted head averaging used to aggregate attention heads when computing token contribution maps."},{"cited_title":"Reducing hallucinations in large vision-language models via latent space steering","cited_arxiv_id":null,"evidence_quote":"The latent-space steering baseline whose feature-difference and SVD recipe VaLSe adapts for vision-aware steering."},{"cited_title":"Mitigating object hallucinations in large vision-language models through visual contrastive decoding","cited_arxiv_id":null,"evidence_quote":"Visual contrastive decoding, the decoding-level debiasing method whose first-order expansion is compared with VaLSe's feature-level steering in the analysis."},{"cited_title":"Towards interpreting visual information processing in vision-language models","cited_arxiv_id":null,"evidence_quote":"Provides the image set and the claim that object information is spatially localized in LLaVA-style models, grounding the paired-sample construction."}],"review_version":1}