{"id":"51b59b11-77f2-45e0-9587-5db348c9befe","arxiv_id":"2506.11571","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new benchmark with edited-image question pairs shows that multimodal reasoning models lose accuracy when visual cues change, suggesting their reasoning is often not faithfully tied to the image.","lead":"This paper builds VFaith-Bench, 755 question-image pairs where key visual details are edited so the same question has a different correct answer. Across 19 models, accuracy drops after the edit, and the authors interpret this as evidence that models often lean on memorized patterns instead of the image.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Repeat ratio conflates perceptual failure with memory; low perception scores predict exactly the repeat behavior, so the data-leakage conclusion is not supported.","rationale":"The reader's weakest assumption concerns edit fidelity and ambiguity of the edited image-question pairs. That is a real threat to the raw-versus-edit accuracy gap. However, the single most load-bearing spot for the paper's headline memory claim is the repeat-ratio metric: it is defined so that perceptual failure produces the exact phenomenon it labels as memory. The paper reports low perception accuracy across all models, which makes this confound active, not hypothetical. This does not require assuming the benchmark data are invalid; the concern is internal to the interpretation of the reported numbers. The proposed test is feasible with data the authors say they will release: the perception subset already probes the same edited cues, so conditioning the repeat ratio on perception correctness directly separates 'saw the change but repeated the old answer' from 'never saw the change.' If the conditional repeat ratio collapses, the central memory conclusion should be retracted or substantially weakened, even though the benchmark may still be useful as a sensitivity and perception probe. Because the reader already issued a conditional verdict, this stress-test does not change the verdict; it sharpens the condition under which the paper's strongest claim should be accepted.","tokens_in":16499,"tokens_out":3007,"duration_ms":32093,"concrete_test":"Link each edited non-perception item to its corresponding item in the 91-case perception task, where available. For each model, compute the repeat ratio restricted to items for which the model correctly identifies the edited cue in the perception question. If the repeat ratio is near zero on those items, the memory interpretation is falsified for them and the phenomenon is perceptual; if the repeat ratio remains high even when the model demonstrably knows the edited cue, the leakage claim gains support. For the remaining items, supplement with a lightweight check: parse the model's CoT on edited questions for explicit acknowledgement of the changed cue (e.g., 'bare trees', 'open blinds') and compare repeat ratios between acknowledged and unacknowledged edits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest novel claim, that models rely on 'previous memories' rather than seen images, rests on the repeat-ratio metric defined in Section 3.2.2: the fraction of edited-image errors in which the model outputs the original ground-truth answer. Section 4.2.1 interprets a high repeat ratio as evidence of 'data leakage or over-reliance on training patterns.' But in this benchmark the question text is identical and the image is nearly identical except for the targeted cue. A model that simply fails to perceive the edit will, by construction, produce the original answer. The paper's own perception results (Tables 2 and 3; accuracy 16.48-50.50%) show that all tested models frequently fail to detect modified cues. Low perception therefore predicts the repeat behavior directly, without any appeal to training-data memory. The metric cannot distinguish 'model saw the changed cue but answered from a memorized pattern' from 'model did not see the change and answered from the unchanged image.' No control condition, per-item perception linkage, or CoT-based check is provided to separate these two mechanisms. This is load-bearing because the headline conclusion about memory and data leakage depends entirely on the stronger reading.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VFaith-Bench, a benchmark for evaluating whether large multimodal models reason from seen visual evidence or from memorized patterns. The authors build an automatic pipeline using visual-cue extraction and GPT-image-1-based editing to create 644 isomorphic image-question pairs (plus a 91-item perception task) in which a targeted visual cue is modified so that the correct answer changes. They evaluate 13 models in raw and edited conditions, reporting accuracy drops, a 'repeat ratio' (how often an error on the edited image still outputs the original answer), and perception accuracy. The central claim is that a significant performance drop on edited images, together with high repeat ratios, indicates hallucination and potential data leakage or over-reliance on training patterns rather than genuine visual reasoning.","tokens_in":16714,"tokens_out":6351,"duration_ms":57383,"significance":"VFaith-Bench is a useful and timely diagnostic resource: the paired design controls for question text and image content, making the raw-versus-edit accuracy comparison a relatively clean measure of sensitivity to specific visual cues. The perception task is a valuable addition, and the repeat-ratio metric, while currently under-validated, points at an interesting failure mode. The paper also provides a reproducible-looking pipeline and promises public code and data. If the central memory/data-leakage interpretation is supported by additional analyses, the benchmark would be an important tool for studying MLLM hallucination and reasoning faithfulness. However, the current evidence for the strongest claim is confounded by perceptual failure, and the lack of statistical rigor weakens the quantitative conclusions.","major_comments":[{"comment":"The repeat-ratio metric, defined as the fraction of edited-image errors in which the model outputs the original ground-truth answer, is interpreted in Section 4.2.1 as evidence of 'data leakage or over-reliance on training patterns.' However, because the question text is identical and the edited image preserves most content, a model that simply fails to perceive the altered cue will, by construction, produce the original answer. The paper's own perception results (Perception column, 16.48-50.50%) show that all tested models frequently fail to detect modified cues, so the high repeat ratios (68.85-87.50%) are exactly what low perception predicts. The metric cannot separate 'saw the change but answered from memory' from 'did not see the change and answered from the unchanged image.' To support the memory/data-leakage conclusion, the authors should provide a per-item analysis linking perception failures to repeat responses (e.g., repeat ratio conditioned on whether the model correctly describes the edited cue in its CoT), or add a control condition that edits a task-irrelevant cue and show that repeat ratios are not elevated there.","section":"Section 3.2.2 and Tables 2-3"},{"comment":"The text repeatedly uses 'significant' (e.g., 'a significant average performance drop', 'statistically higher repeat ratio') without any confidence intervals, bootstrap estimates, or hypothesis tests. Sample sizes are small in several categories (CRE=57, TIF=71, REA=132), so per-category deltas such as the +10.53 point increase for Seed1.5-VL on CRE or the -1.19 for Gemini on PFK are within plausible noise. The headline claims about degradation and about closed-source versus open-source repeat ratios therefore are not quantitatively substantiated. The authors should report at least bootstrap 95% confidence intervals and, if feasible, paired significance tests on the raw-versus-edit comparisons.","section":"Section 4, Tables 2 and 3"},{"comment":"The validity of the benchmark hinges on the assertion that edits change only the targeted visual cue and yield a unique correct answer (Principles 1-3) and that manual verification ensures this (Section 3.2.1). However, the paper provides no details on the number of annotators, inter-annotator agreement, or the criteria applied; the appendix only mentions that a manual review was conducted. Moreover, the example in Figure 7 (train dirt) edits a cue about rain streaks that may still leave option C ('All options are possible') plausible, illustrating the difficulty of guaranteeing uniqueness. Without a documented verification protocol and ideally an independent human-accuracy study, the raw-versus-edit accuracy gap conflates visual faithfulness with edit quality and ambiguity.","section":"Sections 3.1.2 and 3.2.1"},{"comment":"The paper justifies the interpretation of model accuracy drops by stating that humans can solve the edited variants 'with near-perfect accuracy,' but no human-subject data are presented anywhere in the manuscript. Since this assertion is the normative baseline that makes a model drop a 'pathology,' it should be replaced with a measured human accuracy on a sample (or at least on a subset) of VFaith-Bench, with the same question format.","section":"Section 1"}],"minor_comments":[{"comment":"There is a grammatical error: 'it is challenging to analysis with quantitative results' should be 'it is challenging to analyze with quantitative results.'","section":"Abstract"},{"comment":"The subsection heading contains a typo: 'Self-Evloving' should be 'Self-Evolving.'","section":"Section 4.2.2"},{"comment":"The Repeat column lists numbers such as 86.36 and 72.92; the caption or text should explicitly state that these are percentages and should define the denominator used for the repeat ratio in this table.","section":"Table 2"},{"comment":"The statement that closed-source models have a 'statistically higher repeat ratio' is not accompanied by any test statistic or p-value; if the claim is qualitative, remove the word 'statistically' to avoid implying a test was run.","section":"Section 4.2.1"},{"comment":"The example option D contains a typo, 'winder outing,' which should be 'winter outing.'","section":"Figure 1"},{"comment":"The paper uses Claude-3.5-Sonnet to grade answers when direct extraction fails; since Claude-3.7-Sonnet is one of the evaluated models, the authors should discuss whether using a model from the same family as a grader could introduce systematic bias in the scoring of that model's responses.","section":"Evaluation pipeline"}],"recommendation":"major_revision","confidential_remarks":"The benchmark itself is a valuable contribution and the raw-versus-edit accuracy comparisons are worth publishing, but the most newsworthy claim about data leakage and memory is not currently supported because the repeat-ratio metric is confounded with perceptual failure. This is fixable within the scope of the paper by adding a per-item perception-conditioned analysis or an irrelevant-edit control, plus basic statistical intervals. I would encourage the authors to revise the interpretation accordingly rather than merely adding disclaimers, since the current framing overstates what the data show. The fit with the journal's scope is fine; the paper is a benchmark/evaluation contribution in multimodal reasoning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know: this is a usable new benchmark for testing whether MLLMs actually use visual cues, but the headline memory/leakage story is not proven. The accuracy drops on cue-edited images are consistent across 19 models and the perception scores are genuinely low. That part holds. The repeat-ratio argument does not.\n\nWhat's new: the automated pipeline — extract visual cues, generate edit suggestions, edit with GPT-image-1, human check — is a real step beyond manual or single-type edits in POPE/Bingo/MERLIM. The five-category structure and the perception task are useful additions. The raw measurements are plausible: every model drops on edited images, closed-source models show higher repeat ratios, perception accuracy is 16–50%. Those numbers will be useful to people building or evaluating multimodal reasoners.\n\nThe soft spots are concentrated in interpretation. The repeat ratio is defined as the fraction of edited-image errors where the model still outputs the original ground-truth answer. In this design the question text is identical and the image is nearly identical except for the targeted cue. A model that fails to perceive the edit will, by construction, output the original answer. Their own perception results show all models frequently miss the edit. So a high repeat ratio is exactly what perceptual failure predicts; it does not discriminate between 'saw the change but answered from memory' and 'did not see the change.' The data-leakage conclusion in Section 4.2.1 rests on the stronger reading, and the metric cannot support it. That is load-bearing for the paper's title claim.\n\nAlso worth flagging: no error bars, confidence intervals, or significance tests; some category samples are 57–71 items; there is no control condition that edits the image in a way that should not change the answer; and the near-perfect human baseline is asserted but not documented. The pipeline relies on closed models, which limits exact reproducibility, and the dataset itself is not yet in the preprint. These are fixable issues, not disqualifying ones.\n\nWho benefits: anyone working on multimodal hallucination, visual grounding, or benchmark design. The benchmark deserves referee time, but a revision needs to either add a per-item perception linkage or CoT check, or soften the memory claim. I'd recommend major revision rather than rejection; the artifact is worth having and the measurements are credible, but the central interpretation is currently over-sold.","headline":"VFaith-Bench is a genuinely useful new benchmark for probing visual faithfulness in MLLMs, but the headline memory/data-leakage conclusion rests on a metric that conflates perceptual failure with memory and needs to be softened or re-supported.","tokens_in":17229,"tokens_out":1426,"would_cite":true,"duration_ms":13687,"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":"VFaith-Bench shows that large multimodal models often succeed on visual questions by repeating learned patterns rather than reasoning from the image in front of them.","keywords":["visual faithfulness","multimodal large language models","chain-of-thought reasoning","hallucination evaluation","visual cue editing","benchmark construction","repeat ratio","perception task"],"falsifier":"A direct falsifier is a controlled human study on a random sample of VFaith-Bench pairs: if human annotators shown only the edited image and question do not reliably pick the new ground-truth answer, or if an independent edit-quality check finds systematic artifacts or difficulty shifts in the edited images, the memory interpretation loses its footing. A complementary control experiment would edit image regions that are irrelevant to the correct answer; if accuracy drops as much as in the task-relevant edits, the drop reflects general sensitivity to editing rather than reliance on memorized answer patterns.","tokens_in":16276,"feed_emoji":"🖼️","tokens_out":6208,"duration_ms":61471,"temperature":0.7,"pith_summary":"The paper asks whether multimodal large language models solve visual-reasoning problems by genuinely reading the image, and it answers with VFaith-Bench, a benchmark of 755 isomorphic image-question pairs. Starting from 644 questions in existing multimodal reasoning sets, the authors extract the visual cues that a model's chain of thought explicitly says it relies on, edit one such cue so that the correct answer flips to a different option, and keep the question text unchanged. Across every tested model, accuracy drops after this edit, with several models losing more than ten points overall and more than twenty in a single category; models also frequently repeat the now-incorrect original answer, a behavior the paper quantifies as the repeat ratio. The authors conclude that strong performance on the original images may come partly from brittle, memory-like patterns rather than robust perception-and-reasoning, and that current models show substantial visual-cue hallucination on a dedicated perception task. If right, the benchmark provides a way to separate visual faithfulness from raw accuracy when comparing reasoning models.","feed_headline":"Editing one image cue makes multimodal models stumble","feed_subtitle":"VFaith-Bench's swapped visual cues drop accuracy by up to 18 points overall, and models repeat the old, now-wrong answer.","key_machinery":"The load-bearing object is the cue-driven automatic and controllable editing pipeline. A reasoner is prompted to mark every visual cue it uses with explicit <vcues_*> tags; an edit-suggestion model proposes a minimal change to one selected cue that makes a different option correct; GPT-image-1 performs the edit; and human reviewers verify that the edit is coherent, the original answer is no longer valid, and the new answer is unique. This yields isomorphic pairs: the same question, a visually similar image, and a different ground truth. The repeat-ratio metric then isolates the memory-like failure: among all cases where the edited answer is wrong, it measures the fraction in which the model repeats its originally correct answer, a signature of pattern adherence rather than re-reading the image. The perception subset, with questions about which cues were actually changed, separates failures of seeing from failures of reasoning.","core_discovery":"In the paper's own terms, VFaith-Bench establishes that current multimodal reasoning models are not reliably faithful to the visual input: all evaluated models show a significant average accuracy drop on questions whose images were edited only in task-critical visual cues, while maintaining coherent-looking chain-of-thought output. The paper reports that closed-source flagship models lead in raw accuracy but also show statistically higher repeat ratios, indicating greater reliance on memorized response patterns, and that the InternVL series shows the strongest visual-cue perception yet underperforms in overall edited accuracy, suggesting a perception-reasoning misalignment. The authors deliberately interpret the raw-versus-edited gap as evidence that success on the original problems may not necessarily stem from true visual observation coupled with robust reasoning, but potentially from brittle patterns.","pith_inferences":["Not in the paper: if the memory interpretation holds, the repeat ratio becomes a natural training target, e.g. a reward that penalizes carrying over the pre-edit answer after the image has changed.","Not in the paper: the isomorphic-pair design transfers to other modalities, such as short video clips or audio diagrams, where one event-level cue determines the answer.","As an extension: a control set with irrelevant edits would tell whether some of the reported accuracy drop is generic image-edit sensitivity rather than cue-specific memory reliance.","As an extension: the closed-source-versus-open-source repeat-ratio difference hints at memorization scaling with training mixture, but confirming that would require controlling pretraining data, which the paper does not do."],"forward_implications":["If the central claim is correct, raw accuracy on existing multimodal reasoning benchmarks overstates true visual reasoning, since part of the score can come from repeating memorized answer patterns.","VFaith-Bench lets researchers rank models by visual faithfulness, and the paper's results identify concrete failure modes: most models lose points after cue edits, while repeat ratio and perception accuracy vary independently.","The reported pattern that closed-source models have higher repeat ratios suggests that larger, text-heavy training makes models more likely to stick to learned answers rather than update from the new image.","The repeat-ratio metric could be used as a diagnostic signal when training or fine-tuning reasoning models, with lower repetition after cue edits indicating better visual grounding.","The benchmark construction itself provides a reusable recipe for synthesizing challenging, answer-flipping test cases from any multimodal reasoning dataset."],"supporting_citations":[{"why":"supplies the multi-domain, multi-step multimodal chain-of-thought questions that are the source set for most benchmark entries","marker":"[5]"},{"why":"supplies the real-world-scenario questions from which additional isomorphic pairs are built","marker":"[4]"},{"why":"prior diagnostic for entangled language hallucination and visual illusion that motivates measuring visual faithfulness during reasoning","marker":"[10]"},{"why":"provides the object-category, attribute, and relation hallucination taxonomy that VFaith-Bench's perception task targets","marker":"[1]"}],"fun_headline_variants":["Swapping one visual cue drops model accuracy by 18 points","Image-edited questions break multimodal models' reasoning","Models ignore changed visual cues, rely on memory instead","Visual cue swap reveals models' reliance on memorized answers","Models repeat old answers when image cues change"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that GPT-image-1 edits change only the targeted visual cue, leaving the question, the rest of the image, and the difficulty of the task otherwise intact, and that human review guarantees each edited pair has exactly one correct answer; if edits introduce artifacts, ambiguity, or shifts in difficulty, the raw-versus-edit gap conflates visual faithfulness with image quality.","fun_headline_variants_meta":{"raw":{"variants":["Swapping one visual cue drops model accuracy by 18 points","Image-edited questions break multimodal models' reasoning","Models ignore changed visual cues, rely on memory instead","Visual cue swap reveals models' reliance on memorized answers","Models repeat old answers when image cues change"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3110,"prompt_tokens":999,"completion_tokens":2111,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":2035}},"tokens_in":615,"tokens_out":2111,"duration_ms":15468,"temperature":1.0,"reasoning_tokens":2035,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:35.316527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct falsifier is a controlled human study on a random sample of VFaith-Bench pairs: if human annotators shown only the edited image and question do not reliably pick the new ground-truth answer, or if an independent edit-quality check finds systematic artifacts or difficulty shifts in the edited images, the memory interpretation loses its footing. A complementary control experiment would edit image regions that are irrelevant to the correct answer; if accuracy drops as much as in the task-relevant edits, the drop reflects general sensitivity to editing rather than reliance on memorized answer patterns.","supporting_citations":[],"review_version":1}