{"id":"ac39232b-bef0-4df3-8132-ebb4aed5247a","arxiv_id":"2411.15268","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ICT steers attention-head activations toward visual information using blur-derived trusted and untrusted pairs, improving object hallucination benchmarks in LLaVA-v1.5 and Qwen-VL without slowing generation.","lead":"This paper introduces ICT, a method that reduces object hallucination in vision-language models by adding small, precomputed adjustments to the activations of selected attention heads during generation. It reports accuracy gains on POPE and MME benchmarks with no added inference latency, offering a lightweight way to make open-weight multimodal models more visually grounded.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'untrusted' state is never shown to be a hallucination state: blur-derived shifts may be generic visual amplification, so the trusted-intervention mechanism and its generalization claims are unvalidated.","rationale":"The paper's practical result may be real: the shifts are precomputed, inference is cheap, and gains on GQA, A-OKVQA, and MME provide some evidence of generality. The t-SNE analysis and cross-model transfer results in Table 3 also suggest that the shift vectors carry shared visual information. However, the central claim as stated is mechanistic: the method is a 'trusted intervention' that balances attention at image and object levels rather than a generic visual boost. That claim rests entirely on the assumption that blurring the whole image or an object region produces an 'untrusted' state in the specific sense of object hallucination. The paper's own limitation section notes that only Gaussian blur was used, and no experiment connects the blur direction to actual hallucinated activations. The classifiers only detect that activations differ, which is necessary but far from sufficient. If the shift is just a visual-fidelity amplifier, the method might still reduce hallucination on these benchmarks, but the explanation, the 'preserving language priors' advantage, and the predicted failure modes would all need revision. The reader's verdict already conditions acceptance on re-evaluation with clean splits; my concern adds a second, more conceptual condition: validate the direction. I therefore agree partially with the reader: the same weakest assumption was identified, but I weight the missing mechanistic validation over the test-set contamination because the cross-dataset results provide some protection against the latter. A direct alignment test between blur-derived shifts and correct-versus-hallucinated activation differences would settle the issue. No adjustment to the reader's conditional verdict is needed.","tokens_in":20700,"tokens_out":9705,"duration_ms":105274,"concrete_test":"Compute a direct hallucination direction on held-out POPE samples: record the model's internal activations at the final prompt token (before any answer token is generated) for 'No' questions where the base model hallucinates 'Yes' versus where it correctly says 'No'. Average (correct-no minus hallucinated-yes) per head to obtain a direct direction D_n, then compare cosine similarity between D_n and the blur-derived S_n (Eq. 4) for the selected heads. If the alignment is no higher than a control direction from a non-hallucination perturbation such as hue rotation or color jitter that preserves object identity, the blur-derived shifts are not specifically anti-hallucination directions, and the trusted-intervention mechanism is unsupported even if benchmark scores improve.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism requires that Gaussian blur creates an 'untrusted' model state whose activation difference from the clean state points along the object-hallucination direction (Eqs. 3, 4, 6). The paper never verifies this: the binary classifiers in Secs. 4.1 and 4.2 only measure whether a head's activations change under blur, which any visual perturbation will satisfy, and the limitations section admits that only Gaussian blur was used. Because the calibration prompts append the correct object name (q + O_i), the clean-versus-blurred contrast isolates image fidelity, not hallucination. Consequently, S_n and S_obj,n may simply be generic 'sharpen the visual signal' directions. Such a direction could reduce POPE/MME errors by boosting visual attention while leaving the core claim—that the intervention specifically counteracts over-reliance on language priors and preserves beneficial priors—unestablished. It also leaves the method's failure modes, such as amplifying irrelevant visual details, uncharacterized. Separately, the headline COCO Random numbers are tuned on the same test subset and calibrated on POPE 'Yes' pairs from the same benchmark, although cross-dataset gains on GQA, A-OKVQA, and MME partially mitigate that concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICT (Image-Object Cross-Level Trusted Intervention), a training-free, forward-pass intervention method for reducing object hallucination in open-weight LVLMs. For a set of calibrated image-question pairs with 'Yes' answers, the authors blur the whole image (image-level) or only a detected object region (object-level) via a diffusion-style Gaussian noise process, and compute per-head activation shift vectors as the mean difference between clean and blurred activations. SVMs are trained per attention head to identify heads that distinguish clean from blurred states, and the top-K such heads are shifted by scaled versions of the shift vectors during generation. Experiments on LLaVA-v1.5 and Qwen-VL report consistent gains on POPE (nine subsets across MSCOCO, A-OKVQA, GQA), on the MME hallucination subsets, and a zero-latency advantage over contrastive decoding baselines. Cross-dataset and cross-model transfer experiments (e.g., shift vectors from LLaVA applied to Qwen on GQA) are also reported. The authors claim that the method mitigates over-reliance on language priors while preserving useful priors, and that it generalizes across datasets and models.","tokens_in":20991,"tokens_out":5021,"duration_ms":79010,"significance":"If the mechanism and evaluation were fully validated, ICT would be a useful contribution: it is lightweight, adds no inference latency, and is orthogonal to decoding-time contrastive methods, while the cross-dataset and cross-model results suggest the computed shift directions capture something transferable rather than dataset-specific. The paper's core idea of steering attention-head activations during the forward pass, using a cheap calibration set, is novel and practically appealing. However, the paper is currently weakened by a circular evaluation on the POPE COCO Random subset, a lack of reported variance, and, most importantly, an unvalidated mechanistic premise that Gaussian-blur-induced activation differences point along the object-hallucination direction. The out-of-sample gains on A-OKVQA, GQA, and MME provide genuine independent support, but the central claim that ICT specifically counteracts language-prior over-reliance while preserving beneficial priors is not yet established by the evidence presented.","major_comments":[{"comment":"The POPE COCO Random results are partly circular: the activation shift vectors are computed from 1,500 'Yes' pairs from the COCO Random subset, the per-head SVM classifiers are trained on the same subset, and the hyperparameters α (= β) and K are selected by grid search on the POPE COCO Random subset (Appendix A). Reporting the test-set F1 gain on that same subset as a headline result (§5.2, first bullet, and Table 1) is therefore a test-set-tuned number, not a fair evaluation. Please either tune on a separate validation split and then report the COCO Random test row, or clearly relegate the COCO Random row to a development-set result and treat the other eight POPE subsets, plus MME, as the primary out-of-sample evidence. The paper should also report variance or significance (e.g., multiple seeds or bootstrap intervals) for at least the main rows, since no uncertainty is currently given.","section":"Sec. 5.1, Appendix A, Table 1"},{"comment":"The load-bearing premise that a Gaussian-blurred image constitutes an 'untrusted' state whose activation difference from the clean state points along the object-hallucination direction is not validated. The calibration pairs are (q + O_i, V_i) versus (q + O_i, V'_i), where q is 'What is it in the image?' and O_i is the correct object name; thus the trusted/untrusted contrast isolates the model's sensitivity to image fidelity, not its tendency to hallucinate a specific object. The SVM classifiers in Secs. 4.1-4.2 only measure whether heads change activation under blur, which any visual degradation would satisfy, and the Limitations section itself notes that only Gaussian blur was used. As a result, S_n and S_obj,n could be generic 'amplify visual signal' directions rather than hallucination-specific directions, leaving the paper's core mechanistic claim ('mitigates excessive reliance on language priors while enhancing sensitivity to essential visual cues') unestablished. Please add a direct test that the shift direction is tied to hallucination, for example: (i) compare against other perturbations (rotation, cropping, color jitter, or a different noise type) and show the intervention transfers only when the perturbation genuinely induces hallucination errors; (ii) measure whether applying the shift to a hallucinating sample moves its activations toward a correct sample, using the paper's own trusted/untrusted labels; (iii) or at minimum, characterize what fraction of the 1,500 POPE 'Yes' pairs would be answered correctly or incorrectly after blur, to show that the untrusted state corresponds to actual error.","section":"Secs. 4.1, 4.2, Eq. (4), Eq. (6), Sec. 7 Limitations"},{"comment":"The baseline numbers for VCD and OPERA are reported without stating whether they are re-run under the same code, model version, and decoding configuration, or taken from the original papers. Since the paper's central contribution is a comparison against these decoding-time methods, please clarify the provenance of every baseline number and, if they are not re-run, re-run them with the same evaluation harness to ensure a fair comparison. Additionally, all reported numbers in Tables 1 and 3 are point estimates; without any measure of variability across runs or evaluation subsets, it is difficult to assess whether the claimed improvements over VCD/OPERA are statistically meaningful, particularly on A-OKVQA and GQA where some gaps are as small as 0.1-0.4 points.","section":"Sec. 5.1 Baselines; Tables 1 and 3"},{"comment":"The cross-model generalization claim (Sec. 6.2, Table 3) is interesting but the t-SNE visualization in Figure 4 is used to conclude that the shift vectors are 'universal shift vectors' (Sec. 6.2, final sentence). t-SNE is a nonlinear, stochastic embedding, and the visual proximity of points in t-SNE does not directly demonstrate that the shift vectors can be transferred across models; the actual transfer results in Table 3 are the relevant evidence, and they show a performance drop relative to the same-model ICT (e.g., Qwen GQA Random: ICT 86.38 vs ICT-LLaVA 83.27 F1). Please temper the claim of universality and instead report a quantitative similarity measure (e.g., cosine similarity between shift vectors after alignment) or a head-selection overlap statistic.","section":"Sec. 6.2 and Figure 4"}],"minor_comments":[{"comment":"There is inconsistent naming ('OPERA' vs 'Opera') and numerous duplicated or near-duplicate references in the bibliography, e.g., refs [3] and [4] are the same AGLA paper, and refs [54] and [55] are the same survey. Please deduplicate the reference list.","section":"Throughout"},{"comment":"Equation (3) writes a full forward diffusion process, but only the final blurred image V'_i is used; the intermediate steps and noise schedule β_t are never specified or used. Either specify the exact blur/noise parameters used (number of diffusion steps, β_t values, or equivalently the Gaussian kernel and standard deviation) or simplify the equation to a direct Gaussian blur. Without these details, the calibration procedure is not reproducible.","section":"Eq. (3) and Eq. (4)"},{"comment":"The radar chart in Figure 3 is difficult to read, and the text under the model names ('CSQA 119.28', 'Existence 190.00', etc.) is not clearly labeled as referring to which variant (vanilla, VCD, or ICT). The caption should state explicitly which curve and which set of numbers correspond to each method, and the claimed total improvement of '67.37 points' (Introduction, contribution bullet) is not directly derivable from the numbers in the figure; please provide a clear numerical summary table.","section":"Figure 3"},{"comment":"The first bullet in Sec. 5.2 says ICT improves average F1 by 7.09% for LLaVA and 5.44% for Qwen, and that these are 'respectively higher than the previous contrastive decoding sota baseline (Opera) 2.19% and 1.14%.' The comparison is confusing because the 7.09/5.44 numbers are absolute F1 improvements over Regular, while the 2.19/1.14 numbers appear to be the difference between Opera's improvement and ICT's improvement. Please state the comparison on the same scale and clarify whether these are absolute percentage points or relative improvements.","section":"Sec. 5.2, POPE results"},{"comment":"The case studies in Figure 5 and Figure 8 are anecdotal. They are useful for illustration, but the paper should make clear that they are selected examples, not a systematic error analysis. The 'undercut fruit' example also shows a failure mode where ICT misses the 'uncut' attribute, which is a useful negative case; please discuss more explicitly whether and how the method can be expected to handle attribute-level reasoning.","section":"Sec. 6.5 and case studies"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is timely and the zero-latency forward-pass intervention is a useful alternative to decoding-time methods. The main concern is that the mechanistic premise (blur shift = hallucination direction) is not directly validated, and the headline COCO Random result is test-set-tuned. However, the cross-dataset and cross-model results provide enough out-of-sample support that a revision with a proper validation split, variance reporting, and a direct mechanism test could bring the paper to an acceptable standard. I would also encourage the editor to check the completeness of the reference list, since there are several duplicate entries that suggest a sloppy final pass."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuine new recipe, and I think it works, but the paper overclaims what the mechanism is. The method—precomputing activation shift vectors from blurred versus clean images at two granularities, then adding them to SVM-selected attention heads during the forward pass—is a sensible combination of known pieces (ITI-style steering, VCD-style blur contrast, head selection by probe accuracy). The no-latency property is real and practically useful, and the cross-dataset and cross-model results are the strongest evidence in the paper. Gains on A-OKVQA, GQA, and MME are out-of-sample and consistent across two models, which is hard to explain away as pure benchmark overfitting.\n\nThe soft spots are real but not fatal. The evaluation on the POPE COCO Random subset is partly circular: the 1,500 calibration pairs come from that same subset, and alpha/beta/K are tuned on it via grid search. That result should be treated as a validation-set number, not a test-set number. The paper should redo the split, exclude calibration samples from evaluation, and report variance across runs. I also share the stress-test concern about the mechanism: the paper never shows that Gaussian blur induces a hallucination state specifically. The classifier only detects heads whose activations change under blur, which any image perturbation would satisfy. The shift vector could simply be a generic 'boost visual signal' direction. That would still explain much of the improvement, but it undercuts the claim that the intervention specifically counteracts over-reliance on language priors while preserving beneficial ones. The limitations section candidly admits only Gaussian blur was tested, which is good, but it doesn't fix the missing validation. Closest baselines like ITI, AGLA, and NoiseBoost are cited but not compared.\n\nWho is this for? Anyone working on LVLM hallucination or inference-time steering will find it worth reading. It deserves a serious referee, but the review should push for a cleaner evaluation and a direct test of the mechanism, e.g., comparing blur-derived shifts to random visual perturbations or to shifts from other transformations. If the cross-dataset and cross-model results hold up with a proper protocol, this is a solid contribution.","headline":"A plausible training-free intervention for LVLM hallucination with real cross-dataset gains, but the headline POPE numbers are partly circular and the 'trusted/untrusted' mechanism is never directly validated.","tokens_in":21501,"tokens_out":1543,"would_cite":true,"duration_ms":16488,"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":"Adding precomputed activation shifts to selected attention heads during the forward pass reduces object hallucination in large vision-language models without adding inference latency.","keywords":["object hallucination","large vision-language models","activation intervention","training-free","forward pass","attention heads","POPE benchmark","MME benchmark"],"falsifier":"Compute activation shift vectors from several corruptions of the same images (Gaussian blur, cropping, rotation) and compare whether they point in a common direction or whether applying the shift systematically reduces hallucination on a held-out set. If the shifts disagree across corruptions, or if applying the shift to inputs where the model is already correct degrades accuracy, the claim that the shift is a generic trustworthiness direction is falsified.","tokens_in":20517,"feed_emoji":"🖼️","tokens_out":8390,"duration_ms":69906,"temperature":0.7,"pith_summary":"This paper claims that object hallucination in open-weight large vision-language models can be reduced during the forward pass, rather than at decoding time, by adding precomputed activation shifts to attention heads that encode image-level and object-level visual information. The shifts come from comparing model activations on original images against activations on Gaussian-blurred 'untrusted' versions, using 1,500 'Yes' pairs from the POPE dataset. Applied together, the two intervention levels steer attention toward visual evidence while keeping useful language priors intact. If the claim holds, this yields a training-free, plug-in method that improves POPE and MME scores with no added inference latency, sidestepping the trade-off that contrastive decoding methods face.","feed_headline":"Forward-pass shift cuts object hallucinations with zero added latency","feed_subtitle":"Precomputed activation shifts applied to selected attention heads reduce hallucination errors while preserving useful language priors.","key_machinery":"The central objects are the activation shift vectors $S^{(l)}_n$ and $S^{(l)}_{n,\\text{obj}}$, defined as the mean difference between trusted and untrusted attention activations over the calibration samples at head $n$ and layer $l$. A binary SVM classifier trained on the trusted/untrusted pairs, with 2-fold cross-validation, identifies which heads encode image-level versus object-level visual information; the top-$K$ heads by classification accuracy receive the additive shift during the forward pass. Because the shift vectors are precomputed, the intervention adds no computational cost at inference time. The mechanism's role is to push attention toward visual evidence at two granularities without removing language priors.","core_discovery":"The paper's central claim is that object hallucination can be mitigated by a two-level activation intervention applied during the forward pass. For image-level information, whole-image Gaussian blur creates untrusted activations, and the mean trusted-minus-untrusted activation difference across 1,500 POPE 'Yes' pairs defines an image-level shift vector $S^{(l)}_n$ for each attention head and layer. For object-level information, Gaussian blur applied only to the detected object region yields an object-level shift vector $S^{(l)}_{n,\\text{obj}}$. A binary SVM classifier per head picks the top-$K$ heads that best separate trusted from untrusted activations, and the shifts are added to those heads' outputs with intensities $\\alpha$ and $\\beta$ during generation. In tandem the interventions give the model a balanced attention mechanism that reduces over-reliance on language priors while increasing sensitivity to visual cues. The authors report average POPE F1 gains of 7.09 percentage points for LLaVA-v1.5 and 5.44 for Qwen-VL, MME gains of 80.51 and 34.67 points respectively, and essentially unchanged inference time.","pith_inferences":["Inference: if the blur-derived direction is a genuine trustworthiness direction, the same recipe should transfer to attribute and relation hallucinations, not just object presence, which is a testable extension.","Inference: the per-head SVM selection step implies that hallucination control is localizable to a small set of heads; architectures with similar head specialization should benefit similarly, while models without such specialization may not.","Inference: the method's dependence on open weights suggests a practical extension: learning the shift directions from a surrogate model of the same family and applying them to a larger sibling model.","Inference: the reported 1,500-sample calibration budget may be much larger than necessary; ablating the calibration sample count could reveal a data-efficiency frontier for activation-based intervention."],"forward_implications":["Applying ICT improves average POPE F1 by 7.09 percentage points on LLaVA-v1.5 and 5.44 on Qwen-VL across nine subsets, exceeding the gains of the OPERA baseline.","ICT raises MME hallucination-subset scores by 80.51 points on LLaVA-v1.5 and 34.67 on Qwen-VL, and also improves commonsense QA scores.","Inference time stays essentially unchanged (for example, 415.9 ms versus 405.3 ms for 20 generated tokens on LLaVA-v1.5), whereas contrastive decoding methods are 2.1 to 3.6 times slower.","Shift vectors trained on the MSCOCO random subset transfer to A-OKVQA and GQA and even to a different model, with LLaVA-v1.5-derived vectors improving Qwen-VL on GQA by about 4.6 points in F1 and accuracy.","Because the intervention happens in the forward pass, it is orthogonal to decoding-stage methods and can be combined with them."],"supporting_citations":[{"why":"Provides LLaVA-v1.5, the open-weight model on which ICT is calibrated and primarily evaluated.","marker":"[52]"},{"why":"Provides Qwen-VL, the second model used to demonstrate generality and cross-model transfer.","marker":"[5]"},{"why":"Defines VCD, the contrastive-decoding baseline whose language-prior elimination ICT contrasts with and outperforms.","marker":"[43]"},{"why":"Defines OPERA, the decoding-stage baseline compared on POPE.","marker":"[31]"},{"why":"Supplies the POPE benchmark and the 1,500 'Yes' pairs used to compute the intervention vectors.","marker":"[48]"},{"why":"Supplies the MME benchmark used to evaluate hallucination and reasoning changes.","marker":"[20]"},{"why":"Grounding DINO detects the object region that receives the local Gaussian blur for object-level intervention.","marker":"[57]"},{"why":"The forward diffusion process defines the Gaussian blur used to create untrusted images.","marker":"[26]"},{"why":"The SVM is the per-head binary classifier used to select which attention heads receive intervention.","marker":"[15]"}],"fun_headline_variants":["Image-object trust shifts curb hallucination without retraining","Forward-pass activation shifts fix object hallucination, no training","Zero-latency, training-free fix for object hallucination","Cross-level attention steering reduces hallucination in LVLMs","Blur-pair activation shifts cut object hallucination in forward pass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Gaussian blur applied to the whole image or to a detected object region creates an 'untrusted' model state whose activation difference from the original points along the object-hallucination direction; if blur mainly removes generic image fidelity rather than inducing the specific failure mode, the precomputed shift vector may not transfer to new images, prompts, or datasets.","fun_headline_variants_meta":{"raw":{"variants":["Image-object trust shifts curb hallucination without retraining","Forward-pass activation shifts fix object hallucination, no training","Zero-latency, training-free fix for object hallucination","Cross-level attention steering reduces hallucination in LVLMs","Blur-pair activation shifts cut object hallucination in forward pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001151,"raw_usage":{"total_tokens":4785,"prompt_tokens":973,"completion_tokens":3812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":3729}},"tokens_in":589,"tokens_out":3812,"duration_ms":26747,"temperature":1.0,"reasoning_tokens":3729,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:45:00.344932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute activation shift vectors from several corruptions of the same images (Gaussian blur, cropping, rotation) and compare whether they point in a common direction or whether applying the shift systematically reduces hallucination on a held-out set. If the shifts disagree across corruptions, or if applying the shift to inputs where the model is already correct degrades accuracy, the claim that the shift is a generic trustworthiness direction is falsified.","supporting_citations":[{"cited_title":"Improved baselines with visual instruction tuning, 2023","cited_arxiv_id":null,"evidence_quote":"Provides LLaVA-v1.5, the open-weight model on which ICT is calibrated and primarily evaluated."},{"cited_title":"Evaluating object hallucination in large vision-language models","cited_arxiv_id":null,"evidence_quote":"Supplies the POPE benchmark and the 1,500 'Yes' pairs used to compute the intervention vectors."}],"review_version":1}