{"id":"b04e734b-2d34-47eb-8092-742dc83444c5","arxiv_id":"2605.23070","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Flow Mismatching detects anomalies via aggregated velocity mismatches along noise-to-image paths in a flow matching model trained only on normal data, yielding pixel heatmaps without reconstruction or test-time optimization.","lead":"The paper introduces Flow Mismatching, an unsupervised anomaly detection approach that measures velocity disagreements between a flow matching model's learned normal dynamics and geometric paths to test images. A smart generalist might read it for a potential alternative to reconstruction-based methods in image defect detection tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Population-level mismatch decomposition may not ensure per-image separability","rationale":"The reader's weakest assumption correctly flags the core velocity-disagreement idea but does not isolate the population-vs-instance gap in the stated decomposition. My concern is narrower and technical; confirming instance-level metrics would resolve it without altering the overall method. This moves the abstract-only UNVERDICTED verdict to CONDITIONAL pending the check.","tokens_in":1740,"tokens_out":296,"duration_ms":24504,"concrete_test":"Recompute image-level anomaly scores on the full MVTec-AD test set using the paper's reported path aggregation and time-step sampling; measure AUC and per-anomaly-type separation. If AUC drops below 0.90 or certain anomaly categories (e.g., logical vs. structural) show no separation, the population-to-instance link fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's key analysis decomposes the *population* mismatch into an irreducible denoising term plus a Fisher-divergence term between test-path and normal-path score functions, claiming the latter drives anomaly separation. Anomaly detection, however, requires reliable scores on individual test images whose paths contain localized anomalous content. The argument provides no explicit bridge showing that the Fisher term remains dominant and separable at the instance level rather than only in expectation, leaving the translation from population analysis to pixel-wise heatmaps and image-level scores as the least secure step.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes Flow Mismatching, an unsupervised anomaly detection method that trains a flow matching model exclusively on normal images and then probes its learned velocity field along affine paths from Gaussian noise to a target test image. Model-predicted velocities (following normal generative dynamics) are compared to geometric velocities toward the target (which incorporate anomalous content); local disagreements are aggregated over time steps and multiple paths to produce pixel-wise heatmaps and image-level scores without test-time optimization, feature memories, or calibration. A population-level analysis decomposes the mismatch into an irreducible denoising term plus a Fisher-divergence term between test-path and normal-path score functions, which is claimed to drive anomaly separation. Experiments on MVTec-AD and VisA report superior performance over reconstruction-based and recent flow-matching baselines.","tokens_in":1842,"tokens_out":607,"duration_ms":23181,"significance":"If the central claims hold, the work introduces a non-reconstruction paradigm for anomaly detection grounded in generative dynamics, with an explicit population decomposition that explains why velocity mismatches can isolate anomalies. Strengths include the absence of test-time optimization or auxiliary memories and the provision of a theoretical account (via the Fisher term) for the effectiveness of path aggregation. This could influence future generative-model-based detection methods by shifting focus from reconstruction error to velocity discrepancies, provided the population-to-instance translation is secured.","major_comments":[{"comment":"§4 (Mismatch Decomposition): The population-level decomposition of mismatch into an irreducible denoising term and a Fisher-divergence term between test-path and normal-path score functions is derived, but no explicit bound, concentration inequality, or per-image argument is supplied showing that the Fisher term remains dominant and separable when anomalous content is localized within a single test image rather than averaged over the population. This gap directly affects the reliability of the claimed pixel-wise heatmaps and image-level scores.","section":"§4"},{"comment":"§3 (Method) and §5 (Experiments): The robust path aggregation is presented as mitigating the mismatch, yet no ablation quantifies how sensitive the reported gains are to the specific choice of affine paths, number of paths, or aggregation operator; without this, it is unclear whether the superiority over baselines stems from the core velocity-discrepancy idea or from post-hoc aggregation tuning.","section":"§3, §5"}],"minor_comments":[{"comment":"Notation for the geometric velocity and model velocity should be introduced with explicit symbols in §2 or §3 to avoid ambiguity when the decomposition is later referenced.","section":"§2"},{"comment":"The abstract and §1 claim 'without additional calibration,' but the precise meaning of calibration (e.g., threshold selection on validation normals) should be clarified to distinguish from standard practice in anomaly detection.","section":"Abstract, §1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify the scope of our theoretical and empirical contributions. We address each major point below.","responses":[{"response":"We agree that the decomposition is strictly population-level and that the manuscript does not supply a concentration inequality or per-image argument establishing dominance of the Fisher term for localized anomalies. The population analysis is intended to identify the score-gap mechanism that motivates the method, while the pixel-wise heatmaps rely on empirical aggregation of mismatches. We will revise §4 to explicitly state this scope limitation and note that the per-instance reliability is supported by the reported experiments rather than by a formal bound.","revision_made":"partial","referee_comment":"[§4] §4 (Mismatch Decomposition): The population-level decomposition of mismatch into an irreducible denoising term and a Fisher-divergence term between test-path and normal-path score functions is derived, but no explicit bound, concentration inequality, or per-image argument is supplied showing that the Fisher term remains dominant and separable when anomalous content is localized within a single test image rather than averaged over the population. This gap directly affects the reliability of the claimed pixel-wise heatmaps and image-level scores."},{"response":"We acknowledge that the current experiments do not include ablations on the number of paths, the affine-path parameterization, or the aggregation operator. In the revised manuscript we will add these ablations (varying path count from 1 to 16, comparing affine versus alternative interpolants, and testing mean versus median aggregation) to quantify sensitivity and isolate the contribution of the velocity-discrepancy signal from the aggregation procedure.","revision_made":"yes","referee_comment":"[§3, §5] §3 (Method) and §5 (Experiments): The robust path aggregation is presented as mitigating the mismatch, yet no ablation quantifies how sensitive the reported gains are to the specific choice of affine paths, number of paths, or aggregation operator; without this, it is unclear whether the superiority over baselines stems from the core velocity-discrepancy idea or from post-hoc aggregation tuning."}],"tokens_in":1488,"tokens_out":416,"duration_ms":10080,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this paper introduces Flow Mismatching, which detects anomalies by measuring disagreement between a flow matching model's learned velocity and the geometric velocity along affine paths to a test image. It backs the idea with a decomposition of the population mismatch into a denoising term and a Fisher-divergence term on score functions, then shows competitive numbers on MVTec-AD and VisA without reconstruction, test-time optimization, or memory banks.","headline":"Flow Mismatching gives a velocity-discrepancy route to anomaly detection in flow models with a clean population decomposition, but the step to reliable per-image scores is the soft spot.","tokens_in":2338,"tokens_out":168,"would_cite":false,"duration_ms":29591,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"Flow-matching velocity mismatch and Fisher decomposition in CV anomaly detection","alignment":"orthogonal","rationale":"The paper's core construction (affine-path velocity discrepancy decomposed into denoising residual + Fisher divergence between score functions, aggregated over paths/times) operates entirely in the domain of conditional flow matching for unsupervised visual anomaly detection. RS framework modules (AbsoluteFloorClosure, Cost/FunctionalEquation, AlexanderDuality, ArithmeticFromLogic, etc.) derive J-cost, φ-ladder, 8-tick periodicity and spacetime from a single distinction with zero adjustable parameters; none of these structures (J(x), φ identities, reciprocal-cost forcing, 8-tick clocks) appear in or are required by the paper's analysis or method. The paper neither invokes nor contradicts any RS theorem.","tokens_in":65201,"confidence":"high","tokens_out":178,"duration_ms":11847,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Flow matching models detect anomalies by measuring velocity disagreements between learned normal dynamics and geometric paths to test images.","keywords":["anomaly detection","flow matching","velocity mismatch","unsupervised learning","generative models","image anomaly localization","MVTec-AD","Fisher divergence"],"falsifier":"On MVTec-AD or VisA images containing known anomalies, compute mismatch scores after replacing anomalous regions with normal content; if scores do not drop substantially to match normal images, the separation claim fails.","tokens_in":2643,"feed_emoji":"","tokens_out":627,"duration_ms":18755,"temperature":0.7,"pith_summary":"The paper proposes Flow Mismatching to find anomalies without relying on reconstruction. It shows that a flow matching model trained only on normal images produces strong local disagreement between its predicted velocity and the straight geometric velocity toward a test image that contains anomalies. Aggregating the mismatch across time steps and multiple paths from noise yields pixel-wise heatmaps and image scores. This approach requires no test-time optimization, stored features, or extra calibration steps. A reader would care because it turns the generative velocity field itself into a direct anomaly signal.","feed_headline":"Velocity mismatches flag anomalies in flow matching models","feed_subtitle":"Comparing learned normal velocities against geometric paths to test images produces heatmaps and scores without optimization or stored data.","key_machinery":"Velocity mismatch between the learned normal velocity field and the geometric velocity along affine paths from Gaussian noise to the target image.","core_discovery":"The central claim is that anomalies induce strong local disagreement between the model-predicted velocity, which follows normal generative dynamics, and the geometric velocity toward the target, which includes any anomalous content. Aggregating the mismatch over different time steps and multiple paths yields pixel-wise heatmaps and image-level scores. The population mismatch decomposes into an irreducible denoising term and a Fisher-divergence term between the test-path and normal-path score functions, identifying the score-gap component that drives anomaly separation.","pith_inferences":["The same mismatch principle could be tested on sequential data such as video by extending paths through time.","Different path samplings beyond affine lines might reduce variance in the aggregated scores.","The Fisher-divergence term suggests that score-function estimation quality directly limits detection sensitivity."],"forward_implications":["Pixel-wise anomaly heatmaps and image-level scores are obtained directly from aggregated velocity mismatches.","The method operates without test-time optimization, feature memories, or additional calibration.","The mismatch decomposes into a denoising term and a Fisher-divergence term that isolates the component driving anomaly separation.","The approach outperforms prior reconstruction-based and flow-matching anomaly detection methods on MVTec-AD and VisA."],"fun_headline_variants":["Velocity mismatches detect anomalies in flow matching","Flow mismatching identifies anomalies with velocity gaps","Velocity disagreements expose anomalies in flow models","Mismatching flows detects anomalies via velocity paths","Anomalies arise from velocity mismatches in flow models"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The velocity field learned exclusively from normal images will produce detectable disagreement specifically attributable to anomalous content when compared to geometric velocities along paths to test images.","fun_headline_variants_meta":{"raw":{"variants":["Velocity mismatches detect anomalies in flow matching","Flow mismatching identifies anomalies with velocity gaps","Velocity disagreements expose anomalies in flow models","Mismatching flows detects anomalies via velocity paths","Anomalies arise from velocity mismatches in flow models"]},"model":"grok-4.3","cost_usd":0.007797,"raw_usage":{"total_tokens":3565,"prompt_tokens":677,"num_sources_used":0,"completion_tokens":57,"cost_in_usd_ticks":77974500,"prompt_tokens_details":{"text_tokens":677,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2831,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":677,"tokens_out":57,"duration_ms":15547,"temperature":1.0,"reasoning_tokens":2831,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-25T05:18:23.339552+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On MVTec-AD or VisA images containing known anomalies, compute mismatch scores after replacing anomalous regions with normal content; if scores do not drop substantially to match normal images, the separation claim fails.","supporting_citations":[],"review_version":1}