{"id":"b76e6331-6acf-4b2f-9319-7a4404f90038","arxiv_id":"2607.29412","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Hallucinated tokens in vision-language models are detectable from each attention head's deviation from its own faithful source-allocation role, reaching ~93% AUROC with a linear probe.","lead":"The paper shows that when a vision-language model hallucinates, specific attention heads shift away from their usual behavior in reproducible patterns, which the authors call Role-Break. These patterns are linearly readable, so the authors build a lightweight detector that spots hallucinated tokens without retraining the VLM.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Faithful-role stability across images is untested: the reported split-half r likely pools tokens from the same images, so it does not rule out content-dependent role shifts.","rationale":"Reader's weakest assumption identifies the same point; I sharpen it. Finding 2's split-half r is the only direct evidence for cross-image role stability, and its description ('randomly split the faithful pool') is compatible with token-level, not image-disjoint, splits. The detector's image-disjoint protocol makes this distinction consequential: if µ^F varies by image, faithful test tokens carry deviations from the training-image role. Because faithful and hallucinated tokens occur within the same test images, some image offsets are shared across classes, but if hallucination rate correlates with image properties (e.g., number of objects, caption length), the linear probe can learn image-level shortcuts. The paper's other controls are genuinely strong—within-polarity POPE, amplitude matching, channel-choice ablations, permutation nulls—and I am not claiming the result is wrong. I am claiming the specific load-bearing assumption is not yet tested by the reported statistic. The proposed image-disjoint split-half check would settle it. Verdict stays CONDITIONAL because the concern is a missing control rather than a demonstrated failure.","tokens_in":22557,"tokens_out":6903,"duration_ms":87612,"concrete_test":"Recompute the split-half Pearson correlation of µ^F (Supp. Table 5) using image-disjoint halves: randomly partition the 500 COCO images into two sets, estimate µ^F per head from faithful tokens of one set and correlate with the estimate from the other set, for LLaVA-1.5-7B and Qwen3-VL-8B on both COCO-500 and POPE-9000. If r stays ≥0.999 (SNR >100×), the role-stability assumption survives; if r drops materially (e.g., <0.99), the faithful role is partly image-specific and Role-Break on held-out images is contaminated by image content, weakening the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construct is the faithful role (µ^F, σ^F in Eq. 2 and ILR analogue in Eq. 7), estimated on training-split faithful tokens and applied to test tokens. For Role-Break to be a hallucination signature, this role must be a property of each head rather than of the particular image being described. The paper's key support is split-half stability of µ^F (Finding 2; Supp. Table 5, r≥0.9999, SNR 98–853×). But the halves are described only as 'randomly split the faithful pool into two halves'; they are not stated to be image-disjoint. If the same images contribute tokens to both halves, the correlation can be near-perfect because both halves average over the same image-specific allocation patterns. The actual detector uses image-disjoint 80/20 splits (Section 4.2), so µ^F is estimated on training images and applied to held-out images; any image-content-dependent component of the role is precisely what the reported statistic would fail to detect. The four source channels include image attention, which can vary with image complexity, and self attention, which varies with generation position. Position-matching in Finding 2 addresses position for deviation reproducibility, not for the role baseline. Until µ^F is shown stable across disjoint image sets, a test-token deviation conflates hallucination with content/position offsets, and the linear probe may exploit image-level statistics rather than a per-head hallucination signature.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a head-level characterization of VLM hallucination. It defines a per-head 'faithful role' as the mean source-allocation of each attention head over faithful tokens (Eq. 2), and 'Role-Break' as the studentized deviation from that role (Eq. 3; ILR version in Eq. 7). Findings 1–3 argue that single-channel attention indicators are not cross-setting stable, that hallucinated tokens exhibit reproducible, head-concentrated deviations from faithful roles, and that these deviations are linearly readable. The authors then build a linear detector over per-head ILR residuals, evaluate it on six VLMs and four benchmarks under image-disjoint 80/20 splits, and report an average AUROC of 93.23, together with ablations, controls, and a small answer-flipping intervention.","tokens_in":22949,"tokens_out":11787,"duration_ms":140050,"significance":"If the faithful-role construct is valid, this is a valuable contribution: it offers a unified head-level view of heterogeneous hallucination patterns, a lightweight and practical detector, and a rare degree of internal validation for an interpretability-driven claim. The paper includes strong controls: split-half stability analyses, caption-clustered sign-flip permutation nulls, amplitude-matched evaluation, answer-polarity controls, sample-efficiency curves, and Wilcoxon tests over paired observations. The ingredient-ladder and channel-choice ablations are particularly informative. The main weakness is that the faithfulness of the 'faithful role' baseline is not validated on image-disjoint splits, which is load-bearing for the central interpretation; a second, lesser issue is that the 'cross-setting' claim is supported only by in-distribution per-benchmark evaluations. With additional validation of the baseline's transfer, the paper would be a solid contribution.","major_comments":[{"comment":"The faithful-role baseline (μ^F, σ^F; Eq. 2 and ILR analogue Eq. 7) is the reference against which Role-Break is defined. Its claimed stability rests on split-half Pearson correlations (Supp. Table 5, r≥0.9999), but the halves are described only as a random split of the faithful token pool. If tokens from the same images appear in both halves, the correlation measures within-image agreement and cannot detect an image-content-dependent component of the role. The detector itself uses image-disjoint 80/20 splits (§4.2), so μ^F is estimated on training images and applied to held-out images; any content-dependent shift in μ^F is exactly the confound the reported statistic would miss. Position trends are controlled for deviation reproducibility, not for the position-dependence of the baseline. Please add (i) image-disjoint split-half correlations (with caption-clustered intervals), (ii) a with","section":"Finding 2 / Supp. Table 5; Eq. (2)"},{"comment":"The paper motivates Role-Break by contrasting it with signals that fail to remain stable 'across models and tasks,' but the detector is trained and evaluated in-distribution on each benchmark: every (VLM, benchmark) cell uses its own linear probe and its own faithful-role estimate on the same benchmark's training split. The 93.23 average AUROC is therefore an in-distribution number and does not by itself demonstrate cross-setting stability of the Role-Break representation. I recommend adding at least one transfer experiment (e.g., a POPE-trained probe evaluated on AMBER, or a CHAIR-labels-trained probe evaluated on M-HalDetect), or a within-model cross-benchmark role-transfer check. If transfer is not intended, the 'unified/cross-setting' language in the abstract and introduction should be softened.","section":"Abstract / §4.2, Table 1"}],"minor_comments":[{"comment":"The ingredient ladder shows that preserving head identity is the sole load-bearing predictive ingredient; subtracting μF, dividing by σF, and the ILR transform change AUC by <0.2 points. This is an honest and useful control, but it tempers the main-text phrasing that the detector is built on Role-Break in the sense of Eq. (3). Consider stating in the main text that the faithful-role baseline is primarily an interpretability anchor and that the predictive signal is the per-head source-allocation pattern itself.","section":"Supp., Ingredient Ladder (Table 12)"},{"comment":"The reference to 'Figure 4 in the main paper' for the anatomy plot appears to point to Figure 5 (the ΔAUC disruption plot); please correct cross-references.","section":"Supp. Numerical Results, Finding 3"},{"comment":"The sentence 'our setup follows prior work in teacher-forcing part of the answer tokens' is ambiguous for the generative benchmarks. Specify whether M-HalDetect and COCO-Caption tokens are generated free-running or teacher-forced, and how attention is extracted for them.","section":"Section 4.2 / Experimental Setup"},{"comment":"Consider a notation table for π_{k,l,h}, μ^F, σ^F, r^src, r^{ilr}; the paper uses many symbols and the reader must track them across sections.","section":"Eq. (1)–(3) / Notation"},{"comment":"In the table header, 'LLaVA-1.5 †' and the footnote '†13B-scale variant' are clear, but the main text uses 'LLaVA-1.5-13B'; please unify naming.","section":"Table 1 / caption"}],"recommendation":"major_revision","confidential_remarks":"The split-half stability issue is the main obstacle; if the authors provide image-disjoint validation of the faithful role or otherwise show that content-dependent shifts do not drive the deviations, I would be willing to support acceptance. The transfer experiment is desirable but secondary; the central scientific claim about Role-Break as a per-head signature needs the image-disjoint control first."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is the ingredient ladder: the only step that matters for detection is preserving head identity. That is a crisp, reproducible empirical finding, and it is supported by a genuinely careful set of controls—amplitude matching, polarity controls, sample efficiency, permutation nulls, and a channel-choice sweep showing that four-source allocation beats both scalar attention statistics and full head-output vectors. The detector is lightweight, the benchmark coverage is broad, and the reporting is unusually honest, including the authors' own caveat that Role-Break may be a downstream response or self-correction rather than a cause. If you work on VLM hallucination detection, this is worth a read and probably a citation for the probe design alone.\n\nThe soft spot is precisely what the stress-test note flags. The split-half correlation of the faithful role (Supp. Table 5, r ≥ 0.9999) is computed on random halves of the faithful pool, not on disjoint image sets. If the same images contribute tokens to both halves, that correlation can be near-perfect even when the 'role' is partly an image-content artifact. The detector's own 80/20 image-disjoint splits mean the reported AUROC is still honest for in-distribution benchmark performance, but they do not establish the stronger mechanistic claim that each head has an image-invariant faithful role. A deviation on a test token could include a content-shift component, and the linear probe might be exploiting image-level statistics rather than a per-head signature. The authors should be asked to recompute role stability on disjoint image sets, or at least to show that within-image and across-image role variation are separated. This is a load-bearing assumption for the interpretability story, even if the detector survives as an empirical tool.\n\nTwo smaller gaps: no code or data are released, and there is no cross-benchmark or out-of-distribution transfer evaluation. The sample-efficiency analysis mitigates the label burden, but the fitted probe is still per-benchmark and per-model. These are revision issues, not reasons to reject.\n\nMy take: the paper deserves a serious referee. The empirical core is strong, the controls are better than most work in this subfield, and the head-identity finding is worth surfacing. The reviewers should push on image-disjoint role stability and ask for code, but conditional acceptance is the right call if those checks land.","headline":"A solid detector paper with a novel feature representation and unusually careful controls, but the headline 'faithful role' stability claim rests on a split-half statistic that doesn't test image-invariance.","tokens_in":23417,"tokens_out":1592,"would_cite":true,"duration_ms":22362,"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":"VLM hallucinations can be detected as structured, per-head attention deviations called Role-Break, and a simple linear probe over them reaches an average AUROC of 93.23 across six vision-language models and four benchmarks.","keywords":["hallucination detection","vision-language models","attention heads","faithful role","role-break","source-allocation","linear probe","interpretability"],"falsifier":"Take a set of faithful tokens from images with unusual content (e.g., sparse or text-heavy images) or from a different prompt distribution, and check whether per-head source-allocation deviations from the training-split faithful role are as large as those of hallucinated tokens; if AUROC drops to near chance or role estimates shift beyond split-half noise, Role-Break is content noise, not a hallucination signature.","tokens_in":22460,"feed_emoji":"🧠","tokens_out":4996,"duration_ms":51333,"temperature":0.7,"pith_summary":"This paper argues that hallucination in vision-language models is not tied to one failure pattern but shows up as a structured, head-specific departure from each attention head's stable baseline behavior, which the authors call Role-Break. On faithful tokens, every head has a reproducible 'faithful role' — its typical way of distributing attention across system, image, user-text, and self-generated context. On hallucinated tokens, specific heads deviate from that role in reproducible, directionally structured ways, and the pattern is concentrated in a subset of heads. Because the deviation pattern is linearly readable once head identity is preserved, the authors build a lightweight linear detector that needs no VLM fine-tuning and no extra forward passes, reaching an average AUROC of 93.23 across six VLMs and four benchmarks. If this is right, hallucination detection can be a cheap post-hoc readout of internal attention structure rather than a bespoke detector per failure mode.","feed_headline":"Attention-head shifts reveal VLM hallucinations at 93% AUROC","feed_subtitle":"A linear probe over per-head attention deviations flags hallucinated tokens across six VLMs and four benchmarks, no fine-tuning needed.","key_machinery":"The load-bearing object is the per-head source-allocation vector — the fraction of attention a head places on each of four context groups: system prompt, image, user text, and self-generated output. Each head's faithful role is the empirical mean and standard deviation of these vectors over faithful tokens. Role-Break is the standardized deviation from that role after an isometric log-ratio (ILR) transform that maps the four-part composition to three Euclidean coordinates (capturing context-vs-self, image-vs-text, and system-vs-user balances). Flattening these residuals across all heads yields a vector below 5,000 dimensions that a single regularized logistic regression reads. The identity-p","core_discovery":"The central discovery is that a hallucinated token in a VLM is marked by reproducible, per-head deviations from that head's own faithful source-allocation baseline. Each attention head has a stable role: a mean and standard deviation of how it distributes attention across system, image, user-text, and self-generated tokens, estimated from faithful tokens. Hallucination breaks these roles in a way that is consistent across independent data splits, concentrated on a subset of heads, and linearly separable from faithful tokens. The deviations are not just larger; their pattern across heads and source channels is what carries the signal, and the pattern is destroyed when head identity or channel","pith_inferences":["Beyond the paper: Role-Break could serve as a generic monitoring signal for distribution shift, since the faithful-role baseline is a per-model fingerprint; sudden changes in the deviation pattern might flag out-of-distribution inputs even before tokens are labeled hallucinated.","Beyond the paper: The head-identity-preserving representation may apply to other failure modes in multimodal models (e.g., sycophancy or instruction misalignment) if those also produce structured per-head departures; the paper's finding that scalar summaries lose information suggests such signals should be read with head coordinates intact.","Beyond the paper: A testable extension is to use Role-Break as a reinforcement signal for decoding strategies in generative settings — e.g., suppressing tokens whose Role-Break magnitude is high — which the paper leaves for future work."],"forward_implications":["Hallucination can be detected from a single forward pass of the VLM, with no fine-tuning and negligible compute overhead, making real-time token-level monitors feasible.","The Role-Break signal transfers across VLMs of different architectures and across both discriminative and generative tasks, so it is not bound to one model family or one failure pattern.","Because the signal is linearly readable, it can be acted on directly: flipping a VLM's answer when the probe flags it as hallucinated improves answer accuracy across all tested VLMs.","The faithful-role baseline gives an interpretable reference — a hallucinated token is one whose internal information-allocation pattern has departed from the model's own healthy behavior.","A few hundred labeled faithful tokens per class suffice to estimate the role and train the probe, so the detector is practical to deploy in new settings."],"fun_headline_variants":["Role-Break heads spotlight VLM hallucinations","Per-head role deviations spot VLM hallucinations at 93% AUROC","Linear probe on attention roles hits 93% AUROC for VLM hallucinations","Head-level Role-Break detects VLM hallucinations","No fine-tuning needed: head roles catch VLM hallucinations"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The per-head faithful role estimated from training-split faithful tokens is stable across images and generation positions, so a deviation seen on a test token is attributable to hallucination rather than to image-content variation or position effects.","fun_headline_variants_meta":{"raw":{"variants":["Role-Break heads spotlight VLM hallucinations","Per-head role deviations spot VLM hallucinations at 93% AUROC","Linear probe on attention roles hits 93% AUROC for VLM hallucinations","Head-level Role-Break detects VLM hallucinations","No fine-tuning needed: head roles catch VLM hallucinations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2717,"prompt_tokens":730,"completion_tokens":1987,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":1902}},"tokens_in":474,"tokens_out":1987,"duration_ms":18397,"temperature":1.0,"reasoning_tokens":1902,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T07:27:15.364919+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of faithful tokens from images with unusual content (e.g., sparse or text-heavy images) or from a different prompt distribution, and check whether per-head source-allocation deviations from the training-split faithful role are as large as those of hallucinated tokens; if AUROC drops to near chance or role estimates shift beyond split-half noise, Role-Break is content noise, not a hallucination signature.","supporting_citations":[],"review_version":1}