{"id":"23ddf509-0fb6-472d-8891-7766e6f83a4f","arxiv_id":"2608.00986","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"UCFB mitigates cross-modal fusion bias in RGB-depth anomaly detection by slowing the dominant modality during early training and improving cross-modal interaction, yielding small average improvements on MVTec 3D-AD and Eyecandies.","lead":"The authors propose UCFB, a plug-and-play module for multimodal anomaly detection that uses Fisher information to detect when one input modality (RGB or depth) dominates training and then rebalances it. The module reports small average accuracy gains on two industrial benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central causal claim is under-supported: the FIM-trace gap in Eq. 1 is treated as task-relevant information suppression, but it is only a parameter-gradient norm, and the blur ablations do not establish causality.","rationale":"The reader's weakest assumption correctly identifies the load-bearing premise: Eq. 1's FIM trace is used both as a faithful measure of task-relevant information and as a causal driver of fusion performance, while the blurring ablations are used to simulate suppression. My stress test sharpens this into a specific technical objection: the per-modality gradient norm in a fused network is a property of the whole optimization landscape, not an isolated measure of that modality's information content. The Gaussian-noise test would directly falsify the faithfulness of the measure, and the gradient-scaling arm would test whether the blur experiments actually reproduce the mechanism UAA targets. If either fails, the theoretical narrative is unsupported, although UCFB might still work as a generic gradient regularizer. The reader's conditional verdict already demands additional validation (seeds, splits, hyperparameters); my concern reinforces that the causal interpretation specifically needs testing before the 'Fisher information perspective' claim can be accepted. I therefore recommend no change to the verdict: CONDITIONAL remains appropriate, with the added requirement of a causal intervention study.","tokens_in":19276,"tokens_out":7501,"duration_ms":87165,"concrete_test":"Train EasyNet and CFM on MVTec 3D-AD with the depth channel replaced by Gaussian noise (same shape and matched first/second moments as real depth), while keeping RGB clean. Compute Tr(F_depth) over the first 20 epochs. If Tr(F_depth) does not drop substantially below its value with real depth, Eq. 1 conflates fitting difficulty with information acquisition, and the suppression diagnosis in Figure 1 is not established. A complementary arm: for the first N epochs, scale the depth encoder gradients by a constant factor (e.g., 0.1) instead of blurring inputs; if this does not reproduce the degradation in Table 1, the blur ablations do not simulate fusion bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 defines Tr(F_kappa) as the expected squared norm of the gradient of the total loss with respect to that modality's encoder parameters (Eq. 1). This is not a per-modality Fisher information about the task: the total loss is computed after fusion, so the RGB encoder's gradient includes contributions routed through the fusion module and depends on the depth branch; comparing these traces across modalities inside a multimodal network does not measure independent information acquisition. A modality can show a small trace because its features are already linearly predictable from the other modality, or a large trace because it is harder to fit, not because it is more informative. Observation 1 also compares unimodal models (separate encoders/heads) with per-modality traces inside a multimodal model, so the 'gap' may reflect a change in architecture rather than suppression.\n\nThe ablations in Section 2.2 replace suppression by blurring the input image. Blurring removes input information; it does not reproduce the gradient-level competition that UAA later acts on. Thus the experiments show that early damage to input quality is harmful, not that the FIM-trace imbalance is the causal mechanism behind suboptimal fusion. Since the design principle of UAA (slow the dominant modality during a 'critical epoch') follows directly from these two premises, both need to hold for the paper's central 'understanding' contribution. The empirical gains of UCFB may still be real, but they would then be explained by generic gradient regularization rather than by the Fisher-information account.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies cross-modal fusion bias in RGB-depth anomaly detection. It first measures modality-specific information acquisition by the squared gradient norm (trace of the Fisher Information Matrix) and claims that multimodal training induces one modality suppressing the other. It then presents ablative experiments using blurred inputs to argue that early-stage interference is irreversibly harmful, and identifies a 'critical learning epoch.' Based on this analysis, the authors propose UCFB, a plug-and-play module combining unimodal adaptive adjustment (UAA), which slows the dominant modality via Fisher-information-guided regularization, and canonical similarity analysis (CSA), which improves cross-modal interaction. The method is integrated into four MAD baselines and evaluated on MVTec 3D-AD and Eyecandies in single-class, multi-class, and few-shot settings, reporting small average improvements over the baselines.","tokens_in":19640,"tokens_out":6358,"duration_ms":62746,"significance":"If the central analysis were correct, this would be one of the first works to explicitly diagnose and mitigate cross-modal fusion bias in multimodal anomaly detection, and the plug-and-play nature of UCFB would make the contribution practically attractive. The empirical comparisons do show gains on two standard benchmarks, and the paper includes useful ablations and visualizations. However, the conceptual contribution rests on two unvalidated premises: that the FIM-trace gap is a faithful measure of task-relevant information suppression, and that input blurring replicates the effect of that suppression. Because these premises motivate the UAA design, the 'understanding' contribution is not established. The empirical gains are also mostly small and reported without statistical significance, making the strength of the performance claim uncertain.","major_comments":[{"comment":"The quantity Tr(F_kappa) defined in Eq. (1) is the expected squared norm of the gradient of the total loss with respect to the modality encoder's parameters. Since the total loss is computed after fusion, this gradient contains contributions backpropagated through the fusion module and depends on the other modality's branch. Thus a small trace does not necessarily mean the modality is suppressed; it may indicate that the features are already predictable from the other modality or that the encoder has a different effective parameterization. Observation 1 compares unimodal training (separate heads) with per-modality traces in a multimodal model, so the observed gap can reflect architectural change rather than inter-modal competition. The UAA design and the paper's 'understanding' claim rest on interpreting this gap as suppression. Please validate the measure against an independent modality","section":"Section 2.1, Eq. (1), Fig. 1"},{"comment":"The ablative cases replace one modality with a blurred input. Blurring removes information at the input level; it does not reproduce the gradient-level out-competition that UAA later modulates. These experiments show that early input corruption is harmful and irreversible, but they do not establish that the FIM-trace imbalance observed in Fig. 1 is the causal mechanism of suboptimal fusion. The design of UAA follows directly from this causal premise, so a direct manipulation of the hypothesized mechanism is needed (e.g., scaling one modality's gradient norm in early training, or ablating the UAA module and measuring both the FIM-trace gap and performance jointly). As presented, the connection between the blur interventions and the method is asserted rather than validated.","section":"Section 2.2, Table 1, Cases 1–5"},{"comment":"All quantitative comparisons report a single run per method/setting. Several improvements are small (0.1–0.3% I-AUROC in Tables 4–5). Without multiple seeds, confidence intervals, or significance tests, the claim of 'consistent improvements' is not supported. Please provide mean ± std over at least 3 seeds (or paired statistical tests) for the main comparisons, including the few-shot results.","section":"Tables 2–5"},{"comment":"UCFB has two hyperparameters, Υ (Eq. 4) and μ (Eq. 12), and Table 6 shows strong sensitivity on EasyNet (I-AUROC ranging from 91.7 to 93.8). The main experiments (Tables 2–5) do not state which values were used, nor whether the same values were applied across all baselines, datasets, and settings. If Υ and μ were tuned on the benchmark itself, the reported gains may partly reflect hyperparameter selection rather than the method. Report the chosen values and evaluate sensitivity on a validation split.","section":"Section 4.3, Table 6"},{"comment":"The approximation Γ^t_{κ,b} ≈ (δξ²/2) Σ_i ||g^t_{ψκ,i}||² relies on Lemma 1, which assumes gradients are random vectors drawn uniformly from a high-dimensional sphere. Real mini-batch gradients are correlated through the data and the optimization trajectory, and they are not uniformly distributed on the sphere. The paper's Remark acknowledges this is idealized, but an empirical check of the cross-term magnitude is not provided. If the cross terms are not negligible, Γ^t_{κ,b} does not approximate Tr(F^t_{κ,b}) as claimed. Please add empirical evidence or an alternative justification.","section":"Section 3.3, Eq. (6) and Lemma 1"}],"minor_comments":[{"comment":"The notation is inconsistent: the equation defines g_φκ but the trace uses g_ψκ. Please unify the symbol for the modality encoder parameters.","section":"Section 2.1, Eq. (1)"},{"comment":"The proof of Lemma 2 is said to be in the Supplementary Materials. Since this is a convergence claim used to justify the method, either include the proof or state the exact assumptions and a proof sketch in the main text.","section":"Section 3.3, Lemma 2"},{"comment":"The table lists 'Baseline+UCFB (Ours)' for CFM and EasyNet, but the corresponding numbers (96.1 and 93.8) appear in later tables. It would be clearer to refer to the main experimental section and avoid duplicating results without context.","section":"Table 1"},{"comment":"The captions say 'FIM trace' while the y-axis labels say 'Information acquisition amount.' Please align the terminology and state the units (e.g., squared gradient norm) for reproducibility.","section":"Figure 4"},{"comment":"The paper claims 'theoretical results' in the contributions, but the only theoretical content is Lemma 1 (an idealized orthogonality claim) and Lemma 2 (deferred). If the authors intend a theoretical contribution, they should clarify which precise statements are proven and under what assumptions.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The paper's main value is likely empirical: UCFB is a simple plug-in that shows modest gains on two MAD benchmarks. The theoretical 'Fisher information' framing is not yet supported: the measure in Eq. (1) is a gradient norm after fusion, not a per-modality Fisher information about the task, and the blur ablations do not establish causality. The editor may consider asking the authors to either substantially strengthen the validation of the FIM measure and the blur-surrogate equivalence, or reframe the contribution as an empirical regularizer with significance testing, rather than as a validated understanding of cross-modal fusion bias."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a decent engineering paper wrapped in an overclaimed theoretical narrative. The UCFB plug-in—gradient-norm-based regularization plus CCA—gives small but consistent I-AUROC gains (roughly 0.1–1.9 points) across four baselines on MVTec 3D-AD and Eyecandies, in single-class, multi-class, and few-shot settings. That part looks real. The 'understanding' part, where the gains are attributed to Fisher Information traces and a critical learning epoch, is much weaker.\n\nWhat is new: nobody has applied this FIM-trace imbalance diagnostic to MAD before, and the plug-in design is clean and easy to adopt. The authors tested across multiple backbones and settings, and the ablations in Table 6 show both components (UAA and CSA) contribute, with the combination beating either alone. That is useful evidence.\n\nThe soft spots: Eq. 1 defines Tr(F) as the expected squared norm of the gradient with respect to the encoder parameters. That is not Fisher information about the task; it is a gradient norm that depends on the rest of the network and the loss scaling. Comparing these traces across modalities inside a fused model does not measure independent information acquisition. Observation 1 compares unimodal models to per-modality traces inside a multimodal model, so the 'gap' may simply reflect a change in architecture. The blur ablations in Table 1 show that degrading input quality early hurts, but that is not the same as one modality suppressing another in gradient space. So the causal story—suppression during a critical epoch drives the degradation—is not established. It might be true, but this paper does not prove it.\n\nAlso: there are no error bars or multiple-seed statistics anywhere. Table 6 looks like hyperparameter selection on the same test set. Lemma 2's proof is deferred to the supplementary, and Lemma 1 is standard concentration-of-measure, not novel. The consistent small gains could come from generic gradient balancing rather than the Fisher logic.\n\nBottom line: as a plug-and-play empirical contribution, it deserves a serious referee. The authors should add seeds, use validation splits for tuning, and reframe the theory as heuristic motivation rather than mechanism. The paper is honest about the idealized assumption in Lemma 1, but the central causal claim needs reining in.","headline":"A useful plug-in for MAD with a shaky Fisher-information story; the empirical gains are modest but real, the causal claims need reining in.","tokens_in":20108,"tokens_out":2763,"would_cite":true,"duration_ms":27865,"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":"This paper claims that cross-modal fusion bias in RGB-depth anomaly detection is a measurable imbalance in each encoder's Fisher information trace, and that rebalancing it early with a plug-in module (UCFB) improves existing MAD methods.","keywords":["multimodal anomaly detection","cross-modal fusion bias","Fisher information matrix","modality suppression","RGB-depth fusion","unimodal adaptive adjustment","canonical similarity analysis","few-shot anomaly detection"],"falsifier":"Construct a variant of MVTec 3D-AD where the depth channel is replaced by random noise while RGB carries all anomaly signal. If the Fisher trace gap between RGB and depth still appears in the early epochs and equalizing it with UCFB still improves I-AUROC, then the gap is not specifically about task-relevant suppression; if the gain disappears, the suppression story is confirmed.","tokens_in":19184,"feed_emoji":"🔍","tokens_out":10734,"duration_ms":101936,"temperature":0.7,"pith_summary":"This paper sets out to show that cross-modal fusion bias—one sensor's encoder out-learning and silencing the other—is a real, diagnosable bottleneck in multimodal anomaly detection, not a footnote. Using the trace of the Fisher Information Matrix as a per-modality measure of information acquisition, the authors find that RGB and depth encoders acquire information at comparable rates when trained alone but diverge sharply when trained jointly, with the imbalance locked in during the earliest epochs. Five blur-based ablation cases show that disturbing this early phase permanently lowers fusion performance, while later disturbances barely matter. The proposed UCFB module detects the dominant modality and critical window, slows the dominant encoder with a Fisher-guided regularizer, and strengthens the suppressed encoder through canonical similarity analysis. Plugged into four existing MAD baselines, it yields consistent I-AUROC gains on MVTec 3D-AD and Eyecandies across single-class, multi-class, and few-shot settings while keeping memory and speed comparable.","feed_headline":"Plug-in fix lifts anomaly detection up to 1.9%","feed_subtitle":"UCFB slows the dominant sensor during the critical early epochs and improves four MAD baselines on two benchmarks","key_machinery":"The central object is the per-modality Fisher information trace, defined as $Tr(F^\\kappa) = \\mathbb{E}_{x_\\kappa \\sim \\chi_\\kappa}\\left[\\|\\nabla_{w_\\kappa} \\mathcal{L}(w_\\kappa, x_\\kappa)\\|^2\\right]$, which the paper treats as each modality's information-acquisition level. It is used three times: to compute the dominance gap $\\Delta_\\kappa$ that names the dominant modality, to detect the critical learning epoch via the relative drop of the trace between consecutive epochs against a threshold $\\Upsilon$, and to set the regularization strength $\\delta = \\exp(\\mu \\tanh(\\Delta_\\kappa))$ for a squared-parameter-change penalty $\\Gamma^t_{\\kappa,b}$ that slows the dominant encoder. The second mecha","core_discovery":"The paper's central claim is that cross-modal fusion bias in RGB-depth anomaly detection is visible in the trace of each modality's Fisher Information Matrix (the expected squared norm of that modality's encoder gradients). Under joint training, one trace rises much faster than the other in the first epochs; once this dominance is established, the suppressed modality never recovers, and blur experiments indicate that early interference is irreversible. From this, the paper derives a design principle: interventions must occur in the critical early learning epoch, slowing the dominant modality while boosting the suppressed one. UCFB implements this with unimodal adaptive adjustment—a Fisher-gu","pith_inferences":["Editorial extension: the same Fisher trace diagnostic could be applied to other multimodal tasks such as RGB-thermal tracking or video-audio recognition, where early modality suppression is also suspected; the paper tests only RGB-depth anomaly detection.","Editorial extension: the paper's critical-epoch finding implies that simple curriculum choices—like warming up the weaker modality first or ordering training data—might reproduce part of UCFB's gain; this is a cheap experiment the paper does not run.","Editorial extension: because the reported gains are averages over categories, an informative follow-up would be to check whether UCFB's benefit concentrates in categories with intrinsically weak depth signal, which would sharpen the causal story."],"forward_implications":["Existing MAD methods can be upgraded by inserting UCFB without redesigning their fusion modules, since it operates on per-modality gradients and features.","The first training epochs are decisive for fusion quality; damage done there is not repaired by simply training longer, so future methods should schedule any balancing intervention early.","The Fisher trace gap can serve as a live, cheap diagnostic for detecting when one modality is being silenced, because it is computable from gradients already available during training.","Re-balancing information acquisition—slowing the dominant modality and strengthening the suppressed one—is a viable alternative to inventing new fusion architectures for anomaly detection.","The gains hold even with very limited data (5/10/50-shot), so the module does not require large training sets to be useful."],"supporting_citations":[{"why":"Supplies the MVTec 3D-AD real-world benchmark used in the fusion-bias analysis, ablations, and main evaluation.","marker":"[2]"},{"why":"Supplies the Eyecandies synthetic benchmark used alongside MVTec 3D-AD to test the generality of the fusion-bias findings.","marker":"[4]"},{"why":"Baseline MAD method whose unimodal and multimodal Fisher traces are analyzed and whose performance UCFB improves.","marker":"[6]"},{"why":"Baseline cross-modal feature-mapping MAD method used in the empirical study of fusion bias and in the evaluation table.","marker":"[10]"},{"why":"Introduces the use of Fisher information traces to measure per-modality information acquisition, the diagnostic foundation of the paper.","marker":"[21]"},{"why":"Provides the canonical similarity analysis module that UCFB uses to enhance inter-modal interaction.","marker":"[29]"},{"why":"Gives the squared-gradient approximation of the Fisher information trace that makes the per-batch monitoring in Eq. (1) and the regularizer practical.","marker":"[36]"},{"why":"Recent architecture-oriented MAD baseline that UCFB is plugged into and improves in the evaluation.","marker":"[47]"},{"why":"Supplies the on-the-fly gradient modulation mechanism that the unimodal adaptive adjustment module adapts to slow the dominant modality.","marker":"[53]"},{"why":"Baseline incremental unified framework whose RGB-depth fusion is analyzed for bias and improved by UCFB.","marker":"[59]"}],"fun_headline_variants":["Fisher info guides fusion-bias fix for RGB-depth anomaly detection","Plug-and-play UCFB curbs cross-modal bias in anomaly detection","Slowing dominant sensor early lifts anomaly detection","Dynamic calibration overcomes fusion bias in anomaly detection","UCFB: Fisher-guided fix for cross-modal bias in anomaly detection"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the Fisher information trace faithfully measures task-relevant information acquisition, so a large RGB-depth trace gap during joint training reflects one modality genuinely suppressing the other (and that blurring an input simulates that suppression); if the gap is a generic training artifact, the diagnosis and the UCFB design do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Fisher info guides fusion-bias fix for RGB-depth anomaly detection","Plug-and-play UCFB curbs cross-modal bias in anomaly detection","Slowing dominant sensor early lifts anomaly detection","Dynamic calibration overcomes fusion bias in anomaly detection","UCFB: Fisher-guided fix for cross-modal bias in anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000363,"raw_usage":{"total_tokens":1768,"prompt_tokens":692,"completion_tokens":1076,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":993}},"tokens_in":436,"tokens_out":1076,"duration_ms":10132,"temperature":1.0,"reasoning_tokens":993,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:34:48.091395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a variant of MVTec 3D-AD where the depth channel is replaced by random noise while RGB carries all anomaly signal. If the Fisher trace gap between RGB and depth still appears in the early epochs and equalizing it with UCFB still improves I-AUROC, then the gap is not specifically about task-relevant suppression; if the gain disappears, the suppression story is confirmed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline cross-modal feature-mapping MAD method used in the empirical study of fusion bias and in the evaluation table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the use of Fisher information traces to measure per-modality information acquisition, the diagnostic foundation of the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the canonical similarity analysis module that UCFB uses to enhance inter-modal interaction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the squared-gradient approximation of the Fisher information trace that makes the per-batch monitoring in Eq. (1) and the regularizer practical."},{"cited_title":"InEuropean Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Recent architecture-oriented MAD baseline that UCFB is plugged into and improves in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the on-the-fly gradient modulation mechanism that the unimodal adaptive adjustment module adapts to slow the dominant modality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline incremental unified framework whose RGB-depth fusion is analyzed for bias and improved by UCFB."}],"review_version":1}