{"id":"bd862c22-751f-4ec5-b8eb-2a14aca22ee5","arxiv_id":"2608.07958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new multimodal anomaly detection benchmark for battery electrode manufacturing shows that existing methods have high false-positive rates, and a density-aware coreset method modestly improves FPR95.","lead":"This paper introduces LIBAD, a dataset of visible-light and X-ray images from real battery electrode production lines, labeled for 11 defect types. It also presents DA-Core, a method that lowers false alarms in anomaly detection on this data by sampling normal features more evenly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LIBAD's validity as a defect benchmark rests on unverified exchangeability of normal and anomalous patches across materials and batches; pooled evaluation may reward domain-shift detection.","rationale":"The reader's weakest assumption identifies the same load-bearing risk: normal and anomalous patches may differ systematically in material composition, production batch, or process settings, so anomaly scores can capture domain shift rather than defects. The paper's dataset construction paragraph acknowledges this risk but offers only a qualitative mitigation ('whenever possible') and a stratified splitting scheme; it does not provide the metadata or controlled analysis needed to verify exchangeability. Because this concern affects the validity of the entire benchmark, not just the DA-Core method, it is more fundamental than the secondary issue of lambda selection on the test set. The dataset does have genuine strengths: real roll-to-roll collection, three aligned modalities, explicit registration verification, and repeated splits. However, the central claim 'the first multimodal anomaly detection benchmark for Li-ion battery electrode manufacturing' can only be accepted as conditional until the material/batch confound is either ruled out empirically or controlled in the released protocol. The proposed concrete test—group-matched versus cross-group evaluation, plus within-material retraining if metadata can be shared—would settle whether the concern lands. If the check shows no material gap, the benchmark and the DA-Core result stand; if it shows a large gap, the pooled evaluations would need to be re-interpreted or the dataset rebalanced.","tokens_in":13722,"tokens_out":8101,"duration_ms":97220,"concrete_test":"For each of the 10 official splits, evaluate PatchCore and DA-Core separately on each defect group's anomalous test patches against (a) that group's matched normal test patches and (b) all other groups' normal test patches, reporting per-group AUROC and FPR95. If group-matched scores are materially better than cross-group scores (e.g., more than 5 percentage points in FPR95), part of the pooled result is attributable to cross-group domain shift. Additionally, if per-sample material/batch metadata can be released, retrain and test within a single material category and compare to the pooled result; a large drop in within-material AUROC or rise in FPR95 would confirm confounding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central benchmark claim requires that anomaly scores reflect defect-related evidence, not material, batch, or process differences. The only stated safeguard is the Dataset Construction paragraph: normal patches are collected 'near the defective samples or under the same manufacturing conditions whenever possible' and are grouped for stratified splitting. However, Appendix Table S1 shows the 744 patches span at least six approximate material categories (anode; NMC/NCM, LFP, LCO, and unspecified cathode), and Note 1 states exact formulations are commercially confidential. No per-sample material or batch metadata is released, no cross-tabulation of material versus defect group is provided, and no analysis demonstrates that normal and anomalous test patches within each group are matched in composition and process conditions. Because all normal samples are pooled during training, a defect category that is disproportionately associated with one material or one run can be detected largely by global material appearance rather than by localized defect evidence. Such domain-shift detection would inflate AUROC and deflate FPR95, undermining both the benchmark's purpose and the DA-Core comparison. The exchangeability assumption is therefore load-bearing and currently unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents LIBAD, a multimodal anomaly detection benchmark for Li-ion battery electrode manufacturing, containing 744 registered electrode patches (VIS, high-resolution X-ray, inline low-resolution X-ray) with 11 defect categories and image-level labels. The authors benchmark several existing multimodal methods under a VIS+X-rayL protocol and propose DA-Core, a PatchCore-style memory-bank method whose coreset selection adds a local-density term to farthest point sampling. They report that at a 0.05 coreset ratio DA-Core lowers FPR95 from 60.4% to 54.3% relative to standard FPS, and that it reaches lower FPR95 than the best standard FPS result while reducing inference time. They also quantify cross-modal anomaly inconsistency, with 29.5% of anomalous samples visible in only one of the two analyzed modalities.","tokens_in":13964,"tokens_out":9561,"duration_ms":101698,"significance":"If the dataset is valid, LIBAD is a useful contribution: it extends industrial anomaly detection to continuous process manufacturing and to weakly correlated modalities, using real production data with aligned imaging and an inline-compatible evaluation protocol. The paper is careful in several ways: 10 official splits, registration-quality verification, full acquisition details, and an honest acknowledgment that absolute FPR95 remains too high for deployment. DA-Core is a modest but plausible improvement in false-positive control, and the density-aware coreset idea is simple and reproducible. The main caveat is that the benchmark's validity, and therefore the meaningfulness of the method comparison, depends on controlling material and batch confounds that are not yet demonstrated.","major_comments":[{"comment":"The central validity claim of the benchmark—that anomaly scores reflect defect-related evidence rather than material, batch, or process differences—rests on an exchangeability assumption that is stated but not verified. The only safeguard offered is that normal patches were collected 'near the defective samples or under the same manufacturing conditions whenever possible,' yet Table S1 shows the 744 samples span anode, NMC/NCM, LFP, LCO, and unspecified cathode categories, and Note 1 states that exact formulations are commercially confidential. No per-sample material or batch metadata is released, no cross-tabulation of material versus defect group is provided, and no analysis demonstrates that normal and anomalous test patches within each group are matched in composition and process conditions. Because all normal samples are pooled during training, a defect category that is disproportionately associated with one material or one production run could be detected largely by global material appearance rather than by localized defect evidence, inflating AUROC and deflating FPR95. This is load-bearing for both the benchmark and the DA-Core comparison. Please add group-level or material-matched results (e.g., per-group AUROC/FPR95, a matched-pair analysis, or anonymized batch/material metadata) to support the exchangeability claim.","section":"Dataset Construction / Appendix Note 1 / Table S1"},{"comment":"The headline improvement (FPR95 60.4% to 54.3% at coreset ratio 0.05) is obtained with lambda = 0.7, selected as the best value from a sensitivity scan evaluated on the same 10 test splits; no validation-based selection is described. Choosing the density weight from the test data is a form of test-set fitting and makes the reported gain optimistically biased. Although the improvement is not isolated to lambda = 0.7 (lambda = 0.3 already gives 57.1%), the central quantitative claim should either use a pre-specified or nested selection rule, or report a scalar summary over the sensitivity table. Please clarify how lambda (and k, alpha, beta) were set and, if possible, use the official validation splits for model selection.","section":"Experiments (Ablation, Table 3)"},{"comment":"The comparison between DA-Core and the PatchCore VIS+X-rayL baseline is not fully controlled, because the manuscript does not state how the multimodal PatchCore baseline fuses VIS and X-rayL, while DA-Core adds an OCSVM late-fusion module with scaling factors alpha and beta. If PatchCore uses simple score averaging or feature concatenation, part of DA-Core's improvement over PatchCore in Table 2 could come from the fusion module rather than from density-aware coreset selection. The ablation in Fig. 3(a,b) appears to compare standard FPS with density-aware FPS while keeping the rest of the pipeline fixed, which is the correct controlled comparison; please state explicitly that this is the case and specify the PatchCore fusion procedure.","section":"Experiments (Table 2, Eq. (10))"}],"minor_comments":[{"comment":"The value of k and the choice of tau are not specified; report them in the implementation details.","section":"Method (Eq. (6))"},{"comment":"The values of alpha and beta and the OCSVM hyperparameters are not specified; they should be listed to make the fusion module reproducible.","section":"Method (Eq. (10))"},{"comment":"The paper reports standard deviations but no paired significance tests or confidence intervals for the FPR95 differences; given the overlap in standard deviations (e.g., 54.3 +/- 5.7 vs. 60.4 +/- 3.9), a paired test would strengthen the claim that the improvement is not noise.","section":"Experiments (Table 2)"},{"comment":"The caption of Fig. 3 should state explicitly whether the 'standard FPS' curve uses the same OCSVM fusion and the same anomaly-scoring pipeline as DA-Core, so that the controlled comparison is unambiguous.","section":"Experiments (Fig. 3)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The primary risk is dataset-level confounding; the paper's own appendix confirms material diversity without releasing metadata. If the authors can supply matched-group analyses or anonymized batch/material information, the contribution is solid. The lambda-on-test issue is fixable by a nested selection procedure. I support major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"LIBAD is worth knowing about: it's the first multimodal anomaly detection dataset for roll-to-roll battery electrode manufacturing, with aligned visible-light and X-ray images, real production data, and 11 defect classes. The authors also quantify cross-modal inconsistency (29.5% of anomalies visible in only one modality), which is a real and understudied phenomenon. The benchmark protocol is careful—10 stratified splits, FPR95 as a primary metric, inference time reported—and DA-Core, a density-aware extension of farthest-point sampling, gives a modest but consistent FPR95 reduction at small coreset ratios. The paper is honest that absolute FPR95 is still too high for deployment.\n\nThe main soft spot is the exchangeability assumption. Normal patches are collected \"near the defective samples or under the same manufacturing conditions whenever possible,\" but all normals are pooled into one training set. Appendix Table S1 shows at least six material categories (anode, NMC/NCM, LFP, LCO, unspecified cathode), and the exact formulations are confidential. Without per-sample material or batch metadata, or a cross-tabulation of materials vs. defect groups, we can't rule out that anomaly scores reflect batch or composition differences rather than defect evidence. The paper acknowledges this concern but doesn't test it. This is load-bearing for the benchmark's validity; it should be addressed with per-material or per-batch evaluation, or at least an analysis showing the defect groups are balanced on material.\n\nA smaller issue: lambda=0.7 is chosen from a sensitivity table on the same test splits, without a validation-based selection protocol. It's a mild overfitting, but the results are fairly flat across lambda in 0.6–0.9, so this is minor. Also, the dataset release and reproducibility are unverified; the link exists but I couldn't check it.\n\nThe method itself is an incremental but legitimate extension of PatchCore, and the experiments are sound in structure. The dataset fills a real gap in process-manufacturing IAD, and the authors did a good job of framing the problem. The confounding issue is fixable, not fatal, but it needs to be addressed before the benchmark can be trusted as measuring defect-related evidence. I'd send this to serious peer review; it deserves referee time. I'd also suggest the authors add material-matched evaluation and fix lambda selection on validation.","headline":"LIBAD fills a real gap with a useful multimodal battery-manufacturing benchmark, but the benchmark's core validity claim rests on an untested material/batch exchangeability assumption and the headline DA-Core result is selected from a test-set sweep.","tokens_in":14447,"tokens_out":3128,"would_cite":false,"duration_ms":31812,"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":"A new benchmark for Li-ion electrode manufacturing shows defects are often visible in only one modality, and a density-aware memory bank cuts false positives from 60.4% to 54.3%.","keywords":["anomaly detection","multimodal benchmark","Li-ion battery electrodes","roll-to-roll manufacturing","X-ray radiography","visible-light imaging","density-aware coreset","false-positive rate"],"falsifier":"Train DA-Core separately on anode and cathode material groups (or on one production batch) and test on the other; if the FPR95 on normal patches from the unseen group jumps far above the within-group value while defect detection holds, then the memory bank is encoding material- or batch-specific texture rather than defect-related evidence, and the benchmark's anomaly scores partly measure domain shift.","tokens_in":13554,"feed_emoji":"🔋","tokens_out":7494,"duration_ms":75039,"temperature":0.7,"pith_summary":"LIBAD is a dataset of 744 electrode patches physically cut from real roll-to-roll Li-ion battery production lines and imaged with double-sided visible light, high-resolution X-ray, and inline-compatible low-resolution X-ray. The paper's central claim is that this is the first multimodal anomaly detection benchmark for continuous process manufacturing, and that it exposes a failure mode missing from object-centric RGB-3D benchmarks: defect evidence is often strong in one modality and weak or absent in another. On the practical inline setting of visible light plus low-resolution X-ray, representative existing methods transfer poorly, with false-positive rates (FPR95, the share of normal samples flagged when 95% of defects are caught) between 59.5% and 69.7%. The paper proposes DA-Core, a memory-based detector whose density-aware coreset selection keeps fine-grained normal variations in a compact memory bank; at a 5% memory ratio it lowers FPR95 from 60.4% to 54.3% and also cuts inference time. If correct, this gives the community a realistic testbed for an underexplored industrial setting and a faster baseline for it.","feed_headline":"Density-aware coresets cut false alarms in electrode inspection","feed_subtitle":"LIBAD, the first multimodal roll-to-roll electrode benchmark, shows why existing detectors fail and how a 5% memory bank fixes it.","key_machinery":"The load-bearing mechanism is DA-Core's density-aware farthest point sampling (DAFPS). Standard coreset selection picks the next memory feature farthest from the current bank, which covers the feature space but under-represents dense clusters of common normal textures; DAFPS replaces that score with $q_i = \\tilde{r}_i^M (1 + \\lambda \\tilde{\\rho}_i)$, where $\\tilde{r}_i^M$ is the normalized distance to the current memory bank and $\\tilde{\\rho}_i$ is a log-transformed, quantile-normalized local density estimate from a Gaussian kernel over the $k$ nearest neighbors. This keeps global coverage while favoring representative variations inside dense normal regions, so normal test patches are more likely to find close memory neighbors. The benchmark itself is also a named central object: LIBAD, a set of 744 physically cropped electrode patches with aligned visible-light, high-resolution X-ray, and low-resolution X-ray images, registered by optimizing a correlation-ratio similarity measure and organized into 10 stratified splits.","core_discovery":"The paper establishes two related results. First, electrode patches in LIBAD are visually homogeneous, and among the 383 anomalous samples 29.5% are visible in only one modality (11.0% visible only in visible light and 18.5% only in X-ray), so anomaly evidence is modality-selective rather than shared. Second, under the practical inline setting of visible light plus low-resolution X-ray, existing multimodal methods that assume strong cross-modal correspondence or RGB-3D priors transfer poorly, with FPR95 across baselines ranging from 59.5% to 69.7%. DA-Core, which selects memory features by balancing global coverage with local density of normal features, reaches 54.3% FPR95 at a 5% coreset ratio, outperforming the best standard farthest-point-sampling result obtained at a 20% ratio while reducing inference time by 43.9%. The authors conclude that both the distribution of normal features and the relationship between modalities need explicit treatment in process-manufacturing anomaly detection.","pith_inferences":["Beyond the paper's claims, an unresolved question is how much of the measured anomaly signal is defect evidence versus material- or batch-specific texture; a held-out-batch evaluation, training on one production batch or material and testing on another, would isolate this.","The modality-selective subset (42 visible-only, 71 X-ray-only anomalies) suggests a direct test for fusion rules: a method that knows when to down-weight a modality should beat one that always trusts both, and per-defect-category scores would show where false positives concentrate.","Density-aware selection is a generic trick that could plausibly transfer to any memory-based anomaly detector on homogeneous textured surfaces, including other continuous processes such as foil, film, or textile manufacturing."],"forward_implications":["If LIBAD is a fair testbed, then multimodal anomaly detection for continuous processes cannot be evaluated on object-centric RGB-3D benchmarks alone.","If DA-Core's improvement is real, density-aware coreset selection should become the default starting point for memory-based industrial anomaly detection with non-uniform normal feature distributions.","At the reported 5% memory ratio, DA-Core's FPR95 (54.3%) beats the best standard FPS result (56.6% at 20% ratio) while cutting inference time by 43.9%, so compact memory banks need not sacrifice false-positive control.","Because the absolute FPR95 remains around 54%, the paper's own conclusion is that LIBAD is still an open benchmark and no tested method is deployment-ready."],"supporting_citations":[{"why":"Provides the MVTec 3D-AD dataset and the normal-only train/validation protocol that LIBAD follows.","marker":"[3]"},{"why":"Defines the PatchCore memory-bank pipeline and standard farthest point sampling that DA-Core modifies.","marker":"[18]"},{"why":"Supplies the correlation-ratio similarity measure used to register the visible-light and X-ray images.","marker":"[17]"},{"why":"DINOv3 is the frozen feature backbone for DA-Core and all adapted baselines.","marker":"[19]"},{"why":"M3DM contributes the OCSVM late-fusion strategy that DA-Core uses and is a benchmark baseline.","marker":"[22]"},{"why":"Real-IAD D3 represents the object-centric multimodal datasets that LIBAD extends beyond.","marker":"[24]"},{"why":"MulSen-AD is a prior multimodal benchmark unifying appearance, geometry, and internal properties, used as comparison context.","marker":"[11]"},{"why":"CFM, a cross-modal reconstruction baseline, demonstrates poor transfer under weakly correlated modalities.","marker":"[6]"},{"why":"CFR, another cross-modal reconstruction baseline, is benchmarked as a representative method that fails on LIBAD.","marker":"[8]"},{"why":"G2SF supplies a geometry-guided fusion baseline adapted from RGB-3D methods.","marker":"[21]"}],"fun_headline_variants":["LIBAD: first multimodal benchmark for battery electrode defects","DA-Core cuts false positives with density-aware memory","5% memory bank outperforms 20% for electrode defects","Modality-selective defects: new benchmark for battery QC","New benchmark reveals modality inconsistency in electrode lines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark's validity rests on the assumption that normal patches cut near defective regions or under the same production conditions differ from anomalous patches only by the defect, so that after pooling, anomaly scores measure defect evidence rather than material, batch, or process-setting differences.","fun_headline_variants_meta":{"raw":{"variants":["LIBAD: first multimodal benchmark for battery electrode defects","DA-Core cuts false positives with density-aware memory","5% memory bank outperforms 20% for electrode defects","Modality-selective defects: new benchmark for battery QC","New benchmark reveals modality inconsistency in electrode lines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001153,"raw_usage":{"total_tokens":4824,"prompt_tokens":1035,"completion_tokens":3789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":3712}},"tokens_in":651,"tokens_out":3789,"duration_ms":32169,"temperature":1.0,"reasoning_tokens":3712,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:35:49.783733+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DA-Core separately on anode and cathode material groups (or on one production batch) and test on the other; if the FPR95 on normal patches from the unseen group jumps far above the within-group value while defect detection holds, then the memory bank is encoding material- or batch-specific texture rather than defect-related evidence, and the benchmark's anomaly scores partly measure domain shift.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MVTec 3D-AD dataset and the normal-only train/validation protocol that LIBAD follows."},{"cited_title":"InProceedings of the IEEE/CVF conferenceoncomputervisionandpatternrecognition, 14318–14328","cited_arxiv_id":null,"evidence_quote":"Defines the PatchCore memory-bank pipeline and standard farthest point sampling that DA-Core modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the correlation-ratio similarity measure used to register the visible-light and X-ray images."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"M3DM contributes the OCSVM late-fusion strategy that DA-Core uses and is a benchmark baseline."},{"cited_title":"Real-iadd3:Areal-world2d/pseudo-3d/3d dataset for industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"Real-IAD D3 represents the object-centric multimodal datasets that LIBAD extends beyond."},{"cited_title":"A.; Widmer, J","cited_arxiv_id":null,"evidence_quote":"MulSen-AD is a prior multimodal benchmark unifying appearance, geometry, and internal properties, used as comparison context."},{"cited_title":"InProceedings of the IEEE/CVFConferenceonComputerVisionandPattern Recognition, 17234–17243","cited_arxiv_id":null,"evidence_quote":"CFM, a cross-modal reconstruction baseline, demonstrates poor transfer under weakly correlated modalities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CFR, another cross-modal reconstruction baseline, is benchmarked as a representative method that fails on LIBAD."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"G2SF supplies a geometry-guided fusion baseline adapted from RGB-3D methods."}],"review_version":1}