PaperlayerView exact revision on arXiv

Source-linked AI summary

Do Words Matter? What Text Contributes to Few and Full-Shot Vision-Language Anomaly Detection

Arda Baris Basaran, Mesut Ceylan, Elgar Fleisch, Filipe Barata

arXiv:2610.34610v1cs.CV

TL;DR

The paper asks whether trained fixed-anchor anomaly detectors benefit from prompt meaning or from anchor geometry, which ordinary prompt comparisons cannot distinguish. It proves and tests scale and joint-geometry equivalences, finding that scale-matched text-free anchors track prompts while category arrangement matters across categories.

  • Problem

    Prompt comparisons change both linguistic meaning and embedding geometry, leaving unclear which anchor properties affect trained fixed-anchor detectors.

    Source: §1 · p. 2

  • Method

    The paper proves map-equivalence conditions for single and shared category anchors, then tests scale-controlled and geometry-controlled replacements across detector settings.

    Source: §4.1 · p. 4, §5 · p. 6

  • Results

    Scale-matched anchors track prompts (0.16%–0.26% mean image AUROC gaps); meaning-free category-axis spreading improves both industrial benchmarks, and retained random frames transfer within 0.5% AUROC.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Within this detector family, prompt curation for semantic meaning is unnecessary when matched or learned effective scale suffices; unified detectors can improve by spreading category axes, while frozen adapters should retain trained anchors.

    Source: §7 · p. 9

    Equivalence claims require fixed anchors, cosine-based binary maps, and an adapter able to absorb the required rotation; violations remove the guarantee but do not ensure different performance.

    Source: §4.2 · p. 5

Abstract

from arXiv · show

Vision-language anomaly detectors describe normal and abnormal appearance with text, and their gains are commonly credited to the semantic content of these descriptions. This attribution is plausible in the zero-shot setting, but few- and full-shot detectors interpose a trained visual adapter, and in that setting it has not been isolated. Changing a description changes both its meaning and its geometry, so accuracy comparisons cannot separate the two. We ask what a fixed pair of text embeddings, or anchors, contributes to such a detector: semantic content, effective scale, or joint anchor geometry. For a single anchor pair, the answer is effective scale alone. The score reads the anchors only through their difference, and a trained adapter can rotate to follow any replacement direction, leaving that direction, and with it the linguistic content, unidentifiable. We prove that replacement anchors with the same difference-vector length permit exactly the same probability maps and attainable losses. Across seven medical datasets, three vision backbones, both supervision regimes, and three adapter families, scale-matched random and orthogonal anchors track correct prompts; breaking the scale match opens a gap that restoring it closes without text. The same pattern holds in MVFA and when MadCLIP's learned prompts are replaced by a text-free learned scale. When one adapter serves many categories, the answer expands to their joint arrangement. Rotating all category axes together changes nothing, as predicted. For the tested CLIP prompts, these axes are nearly collinear, and meaning-free rotations that spread them apart improve performance on both industrial benchmarks. Even under frozen zero-shot transfer, a random anchor frame transfers as well as the prompt frame: coordinate consistency, not meaning. What these detectors read from their prompts is scale and arrangement, not semantic content.

1 INTRODUCTION

The paper asks which properties of fixed text anchors remain observable after a visual adapter is trained, because prompt comparisons change both meaning and geometry. Its results point to effective scale for a single pair and joint geometry across categories, rather than semantic content.

1 INTRODUCTION

  • Vision-language references can support low-sample operation and retargeting by changing a prompt’s noun, unlike per-category models fitted from normal data.Anomaly labels are rare, diverse, and seldom available.Source: §1 · pp. 1–2
  • Ordinary prompt comparisons cannot isolate semantics because changing a description also changes its numerical geometry, while a trained adapter can absorb direction changes.This motivates testing semantic content, effective scale, and joint anchor geometry separately.Source: §1 · p. 2

Single-pair identification

  • 0.16%–0.26% mean gaps in image AUROC separate scale-matched random and orthogonal anchors from correct medical prompts; breaking scale opens a gap that restoring it closes.The scale-restoring controls use no text.Source: §1 · p. 2

Shared multi-category geometry

  • 5.6% image and 4.4% pixel AUROC improvements on VisA, and up to 8.8% and 8.3% across vision backbones, followed meaning-free spreading of category axes.The tested CLIP prompt axes were nearly collinear, and the rotations preserved each category’s scale.Source: §1 · p. 2

Frozen zero-shot transfer boundary

  • A retained random anchor frame averages within 0.5% AUROC of prompts in frozen zero-shot transfer.The frozen adapter cannot rotate, so its anchor frame must be retained, though it need not be linguistic.Source: Figure 1 · p. 2, §1 · p. 3

2 RELATED WORK

Prior work varies in where language and visual adaptation enter anomaly detection, but neither frozen-feature studies nor learned-prompt methods isolate fixed anchors paired with a separately trainable visual branch.

2 RELATED WORK

  • Vision-language anomaly detectors range from frozen handcrafted prompts to learned visual projections and learned prompt tokens, so changed text readouts may or may not be compensable by a trainable visual branch.The passage names WinCLIP, APRIL-GAN, PALADIN, AnomalyCLIP, and medical systems as examples.Source: §2 · p. 3
  • Prompt-based detectors commonly attribute gains to anomaly semantics or vision-language alignment, yet strong visual-only systems mean accuracy alone cannot identify language’s contribution.Learned prompts further complicate attribution because their text representations are fitted rather than prescribed.Source: §2 · p. 3
  • Studies of semantic descriptors and random words concern frozen visual features, not fixed text anchors with a separately trainable visual branch.The remaining question is which properties of fixed text embeddings remain identifiable after visual adaptation.Source: §2 · p. 3

3 PRELIMINARIES

The studied detector maps frozen image-patch features through trainable adapters into a fixed two-anchor space, then produces anomaly maps and evaluates them at pixel and image levels.

3 PRELIMINARIES

  • Frozen vision-backbone patch features from retained layers are processed by layer-specific trainable adapters into the anchor space.The section frames this as the detector family under analysis.Source: §3 · p. 3
  • Fixed unit normal and abnormal anchors define cosine logits, with a multiplicative similarity scale that acts as an inverse temperature under conventional softmax parameterization.Text anchors are normalized means of fixed prompt ensembles; random-coordinate controls construct vectors directly.Source: §3 · p. 3
  • Patch logits are reshaped and bilinearly upsampled before a two-class softmax, and the abnormal channel yields the anomaly map.Focal and Dice losses train layer maps; inference averages them, using pixel AUROC and the mean of the top 1% pixels for image score.Source: §3 · p. 3

4 THEORETICAL ANALYSIS

The analysis asks when text-derived anchors can be replaced by text-free vectors without changing representable abnormal-probability maps. It treats a single pair through effective scale and shared category pairs through their joint geometry.

4.1 ONE ANCHOR PAIR: SEPARATING DIRECTION FROM SCALE

  • For one anchor pair, only the anchor difference enters the abnormal probability, and its alignment direction can be followed by a retrained adapter while its effective scale controls probability sensitivity.Normalized adapter outputs cannot absorb a change in axis length by rescaling.Source: §4.1 · p. 4
  • Equal effective scales with separately trained rotation-closed adapters permit exactly the same category-agnostic abnormal-probability maps and attainable objective or evaluation values.Equal maps need not yield identical training outcomes because stochasticity and finite optimization can produce different solutions.Source: §4.1 · p. 4

4.2 SEVERAL ANCHOR PAIRS: WHY THEIR ARRANGEMENT MATTERS

  • With one adapter serving several categories, a common rotation preserves their scoring-axis arrangement, whereas independent rotations can alter relative angles that no shared rotation can undo.The Gram matrix records category-axis lengths and alignments; matching it is sufficient for equivalent maps at the same similarity scale.Source: §4.2 · pp. 4–5
  • The equivalence guarantees require fixed anchors, cosine-based binary probability maps, and an adapter able to follow changed anchors; changing the shared Gram matrix removes the guarantee.Removing a guarantee does not itself establish different performance or a benefit from correct text.Source: §4.2 · p. 5

5 METHOD: ANCHOR CONTROLS

The method tests which anchor properties matter by comparing text prompts with controlled text-free replacements and extensions to more complex adapters, category banks, and frozen transfer.

5 METHOD: ANCHOR CONTROLS

  • Scale-matched random anchors preserve the correct prompts’ effective scale and permit the same probability maps, making separately trained performance gaps finite-training residuals.The controls are repeated across supervision regimes, vision backbones, and adapter families.Source: §5 · p. 5
  • Orthogonal anchors change effective scale, while scaling their similarities by √1 −ρcorrect restores it without restoring text.The matched orthogonal condition is a second probability-map-level control.Source: §5 · p. 5
  • Donor-bank swaps change anatomy, modality, and state descriptions together, but cannot isolate semantics because the donor may not match ρcorrect.The authors treat this as a complementary stress test, not an exact test of meaning.Source: §5 · p. 6
  • Approximately parameter-matched linear, MLP, and contextual Transformer adapters test whether observed scale dependence extends beyond a linear head.The experiments also retain MVFA’s prescribed-anchor pipeline and replace MadCLIP’s learned prompt path with a text-free learned scale on a random frame.Source: §5 · p. 6
  • For a shared industrial adapter, a common rotation preserves category-axis arrangement, whereas independent rotations preserve each scale but change off-diagonal alignments.These interventions test finite-training invariance and whether joint geometry matters when individual scales are fixed.Source: §5 · p. 6
  • Frozen zero-shot medical transfer compares source and target anchor pairs to test whether transfer depends on linguistic content or consistency with the adapter’s trained coordinates.The target is scored without target-side training.Source: §5 · p. 6

6 EXPERIMENTS

Across the tested medical and industrial settings, performance tracked anchor scale and joint geometry more closely than prompt meaning. Results also show that frozen adapters depend on keeping their anchor frame consistent, while the experiments distinguish that boundary from retraining.

6.1 SETUP AND READING RULE

  • The experiments covered seven medical datasets, three vision backbones, real-anomaly-support and normal-only regimes, and full-shot industrial benchmarks with category-specific anchor pairs sharing category-agnostic adapters.The two medical supervision regimes have different support grids and were not pooled for effect magnitudes; image AUROC was primary, with pixel AUROC reported where valid masks exist.Source: §6.1 · pp. 6–7, Table 1 · p. 7

6.2 FEW-SHOT MEDICAL ANCHOR CONTROLS: EFFECTIVE SCALE, NOT PROMPT CONTENT

  • With real-anomaly support, Gram-matched random and orthogonal anchors had mean absolute image-AUROC gaps of 0.16% and 0.18%; with normal-only training, gaps were 0.22% and 0.26%.The comparisons pooled three vision backbones; correct prompts showed no consistent advantage over these scale-matched text-free controls.Source: §6.2 · p. 7
  • Without scale matching, mean absolute image-AUROC gaps rose to 1.46% with real-anomaly support and 0.93% with normal-only training; restoring scale returned gaps to near zero.The largest reported gap was 11.8% on DINOv3-L/16 OCT2017 at K = 2, with 93.6% versus 81.8%.Source: §6.2 · p. 7

6.3 EXTENSIONS ACROSS ADAPTER FAMILIES, RESIDUAL DETECTORS, AND LEARNED PROMPTS

  • Across linear, MLP, and contextual Transformer adapters, correct-prompt versus Gram-matched-random gaps stayed within 0.5% image AUROC, while scale matching reversed unmatched-anchor losses.Unmatched orthogonal anchors reduced contextual-Transformer performance by 5.4% on CLIP ViT-L/14 and 18.7% on DINOv3-L/16 with real-anomaly support; matching scale brought both losses back to within 0.2%.Source: §6.3 · p. 7
  • At K=16, MVFA mean image AUROC was 94.8% with prompts, 93.9% with Gram-matched random anchors, 73.7% with unmatched orthogonal anchors, and 94.5% after scale restoration.MVFA's residual architecture is outside the exact closure result, so this is empirical evidence beyond the theorem's guaranteed scope.Source: §6.3 · p. 8
  • At K=16, MadCLIP learned prompts and a text-free learned scale reached 96.2% and 96.3% mean image AUROC, while both paths approached antipodal anchors.The fixed prompt-scale arm reached 95.3% at K=16 but was less stable at small K.Source: §6.3 · p. 8

6.4 UNIFIED MULTI-CATEGORY DETECTION: JOINT ANCHOR GEOMETRY

  • Independent rotations that spread category axes improved VisA by 5.6% image and 4.4% pixel AUROC, and MVTec-AD by 1.2% image and 3.0% pixel AUROC.Gains reached 8.8% image and 8.3% pixel AUROC on VisA across the tested vision backbones; shared rotations stayed within 0.1% of prompts on all four AUROC endpoints.Source: §6.4 · p. 8, Table 3 · p. 9

6.5 FROZEN ZERO-SHOT TRANSFER: THE BOUNDARY OF ANCHOR EQUIVALENCE

  • With a frozen adapter, retained random source pairs differed from prompts by at most 0.5% image AUROC, whereas random-pair swaps cost 9.5%–19.6% across the reported settings.Swapping pairs can change direction and scale together, so the results support keeping the trained adapter and anchor pair coupled, not a specific geometric cause of the swap cost.Source: §6.5 · p. 9

7 CONCLUSION

The conclusion identifies anchor geometry rather than semantic content as the determinant of performance: scale for a single pair and joint arrangement across categories. It draws practical implications, states limits on the guarantees and evidence, and documents reproducibility and AI assistance.

7 CONCLUSION

  • For a single anchor pair, matching the difference-vector scale preserves representable maps, while spreading near-collinear category axes improves both industrial benchmarks.The conclusion also reports that MVFA and MadCLIP matched prompts using text-free anchors, and that a retained random frame transferred as well as prompts under frozen transfer.Source: §7 · p. 9
  • Prompt wording or expert prompt banks cannot improve accuracy through meaning within this detector family; a unified detector may improve by spreading category axes apart.Frozen adapters should be deployed with their training anchors, and claims that prompt semantics drive performance should control for effective scale.Source: §7 · p. 9
  • The guarantees require fixed anchors, rotation-closed adapters, and cosine-based binary maps; evidence beyond that premise is empirical rather than guaranteed.All anchors use one frozen text encoder, the best scale remains unknown because only two scales were compared, the proposed interference mechanism was not measured, and closeness across conditions is descriptive over three seeds.Source: §7 · p. 9

REPRODUCIBILITY STATEMENT

  • The reproducibility statement points to public datasets and appendices containing training, evaluation, prompt, anchor-construction, proof, and result details.It also says the appendices document changes to MVFA and MadCLIP.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

  • Generative AI assisted with derivations, proof checking, methodology, code, result interpretation, and manuscript drafting; the authors say they reviewed and verified its contributions.The authors report executing the experiments and taking responsibility for the final manuscript and its claims.Source: AI USE STATEMENT · p. 10

A THEORETICAL DETAILS AND PROOFS

The appendix checks how the implemented detector, losses, and evaluation metrics depend on abnormal probability maps, then supplies a Gram-matrix lemma used in the shared multi-category proof. Its roadmap also covers prompt and anchor geometry, rotation closure, and finite-training residuals.

A.1 FROM THE IMPLEMENTED DETECTOR TO ABNORMAL PROBABILITY MAPS

  • The layer-wise scorer uses four intermediate layers with a separate adapter per layer but the same fixed unit anchor pair.The similarity-scale multiplier uses τ = 100 in training and τ = 20 at evaluation.Source: §A.1 · p. 14
  • Bilinear upsampling with class-independent weights preserves the logit difference, and the two-class softmax makes the abnormal probability a function of that difference.Consequently, the normal channel carries no information beyond the abnormal map.Source: §A.1 · p. 14
  • The focal and Dice training terms use target probabilities and the abnormal map, rather than observing the two logits separately.The focal term uses a numerical clamp and γ = 2; the Dice term reads the abnormal map directly.Source: §A.1 · p. 14
  • Pixel and image AUROC both factor through the abnormal probability maps, which are the common outputs consumed by training and evaluation.The image score uses a top-fraction mean of averaged layer maps.Source: §A.1 · p. 14

A.2 GRAM-TO-ROTATION LEMMA

  • Corresponding row vectors with equal Gram matrices can be mapped to one another by an orthogonal transformation.The proof constructs an inner-product-preserving map between their spans and extends it to an orthogonal map on the full space.Source: §A.2 · p. 15

A.3 COMPLETE PROOF OF THEOREM 1

This appendix characterizes when anchor changes preserve detector outputs, extends the analysis to shared multi-category detectors, and specifies the medical prompt-mismatch intervention and clinical and industrial prompt banks. It also reports anchor-scale and cross-category geometry and states the fixed-anchor and rotation-closed-map conditions for exact rotation-symmetry coverage.

A.3 COMPLETE PROOF OF THEOREM 1

  • Equal-length single-pair difference vectors yield identical probability maps, losses, and evaluation scores when the adapter family is closed under rotations.The proof composes a common orthogonal rotation into every adapter layer; the reverse rotation gives the same attainable maps in the opposite direction.Source: §A.3 · p. 15

A.4 UNIFIED MULTI-CATEGORY SPECIALIZATION

  • For a shared multi-category detector, matching the difference-vector Gram matrix is necessary and sufficient for the category vectors to differ only by one global orthogonal rotation.The full anchor Gram is sufficient but not necessary: rotating anchor midpoints while preserving their norms can change that Gram without changing difference vectors or detector outputs. The proof makes no general claim that equal Gram matrices imply equal hypothesis classes for arbitrary adapters and input domains.Source: §A.4 · pp. 15–16

A.5 CLINICAL PROMPT BANKS AND MISMATCHED-DATASET INTERVENTION

  • The medical mismatch intervention replaces each target’s complete prompt bank with a fixed donor bank, while keeping training and evaluation target-specific.The donor prompts jointly change anatomy, modality, normality and abnormality definitions, and templates; because the resulting pair is not scale-matched, this is not an exact probability-map control or a standalone test of semantic necessity. The stated exact test compares correct prompts with text-free controls matched in effective scale.Source: §A.5 · pp. 16–17
  • The mismatched-dataset donor condition uses the donor’s complete normal and abnormal prompt banks unchanged, without rewriting sentences for the target.Source: §A.5 · p. 23

A.6 INDUSTRIAL PROMPT BANKS

  • Industrial banks average and unit-normalize separate normal and abnormal means from 154 template-expanded sentences per state and category, with no trained prompt tokens or anchors.The construction uses seven state phrases and 22 photo templates with frozen OpenAI CLIP ViT-L/14 text encoding.Source: §A.6 · pp. 23–24

A.7 ANCHOR COSINES AND EFFECTIVE SCALES

  • Matched orthogonal anchors rescale similarity by c to recover the reported effective scales; without matching, they inflate the correct-prompt scale by 1/c.The scale calculation uses the anchor-difference norm of √2 and frozen OpenAI CLIP ViT-L/14 anchors; displayed values are rounded.Source: §A.7 · pp. 24–25
  • Independent category rotations raise effective rank from 1.79 to 14.28–14.29 of 15 for MVTec-AD and from 1.78 to about 11.10 of 12 for VisA.Shared rotations preserve the correct-prompt difference Gram matrix and its effective rank; independent rotations alter only cross-category off-diagonal geometry.Source: Table 16 · p. 25

A.8 ADAPTER COVERAGE AND ROTATION CLOSURE

  • Exact theorem coverage requires prescribed, fixed anchor pairs and a visual-map family closed under left multiplication by every orthogonal matrix.Learned-token anchors may not reach arbitrarily rotated pairs, leaving ρ and λeff fitted; visual constraints can also break closure.Source: §A.8 · p. 26

A.9 FINITE-TRAINING RESIDUAL

The theorem equates attainable probability maps and objective values under reparameterization, but finite training can leave residual differences when the procedure is not rotation-equivariant.

A.9 FINITE-TRAINING RESIDUAL

  • Under exact anchor matching, observed finite-training differences can reflect non-equivariant procedure components or stochastic elements that are not fully coupled.The theorem equates attainable maps and objective-value sets, not arbitrary optimization paths.Source: §A.9 · p. 26
  • Frobenius weight decay is invariant, whereas coordinate-aligned uniform initialization is not exactly invariant under right multiplication by O(d).Equivariance requires the initialization distribution to be invariant under that rotation.Source: §A.9 · p. 26
  • Adam's coordinate-tied diagonal preconditioner and initialization are principal algorithmic sources of finite-training discrepancy; bfloat16 and uncoupled randomness may also contribute.The reported residuals are not uniquely attributed to initialization or Adam.Source: §A.9 · p. 26

A.10 PROOF OF PROPOSITION 1

The exact equivalence argument is conditional on prescribed fixed anchors and a rotation-closed visual map family, and its failures do not establish a benefit from linguistic meaning.

A.10 PROOF OF PROPOSITION 1

  • The proof rotates every layer-wise adapter, so it applies only when transformed adapters remain admissible and the adapter is trained against replacement anchors.Exact coverage also requires prescribed anchors held fixed within a run and a visual map family closed under left multiplication by O(d).Source: §A.10 · p. 26, Table 17 · p. 27
  • For category-conditioned adapters, the corollary requires one shared rotation for all category difference vectors; changing their joint arrangement removes that construction.The scope check concerns exact theorem assumptions, not a judgment about a method.Source: §A.10 · pp. 26, 28, Table 17 · p. 27
  • Failure of the sufficient equivalence argument is not a sufficient condition for performance differences and does not establish that linguistic meaning is beneficial.Table 18 concerns rotation closure of the three adapter families.Source: Table 18 · p. 27, §A.10 · p. 28

B INTERNAL TRAINING, CHECKPOINT, AND EVALUATION PROTOCOL

The protocol fixes frozen encoders, prescribed anchors, layer-specific adapters, and evaluation procedures, while distinguishing real-anomaly supervision from online synthetic-anomaly training. Synthetic anomalies differ between medical normal-only and industrial full-shot settings.

B INTERNAL TRAINING, CHECKPOINT, AND EVALUATION PROTOCOL

  • Three frozen visual backbones feed four layer-specific adapters into a 768-dimensional anchor space; prompt ensembles are encoded once and anchors remain fixed.The protocol uses 224×224 inputs and retains features from transformer blocks 5, 11, 17, and 23.Source: B INTERNAL TRAINING, CHECKPOINT, AND… · p. 28, Table 19 · p. 28
  • Training sums focal loss with γ = 2 and Dice loss across four layer maps; evaluation averages abnormal maps and computes pixel AUROC and a top-1% image score.Industrial AUROCs are macro-averaged over object categories, whereas medical AUROC uses the applicable dataset test split.Source: B INTERNAL TRAINING, CHECKPOINT, AND… · p. 28
  • Real-anomaly support uses K normal and K abnormal images with dense masks for four medical datasets, while normal-only anomalies are generated online.OCT2017 is evaluated at image level only; BUSI pixel AUROC is omitted from normal-only anchor tables for a common image-only endpoint, despite available evaluation masks.Source: B INTERNAL TRAINING, CHECKPOINT, AND… · p. 28

B.1 SYNTHETIC-ANOMALY GENERATION

  • Medical normal-only synthesis combines deterministic image views with CutPaste, Gaussian intensity-change, or identity tasks, and draws targets from task-specific rules.The nonidentity edits are confined to one or two Perlin-noise regions and do not consult foreground masks.Source: §B.1 · p. 29
  • Industrial synthesis Poisson-blends category-matched crops within object foregrounds; calibrated intensity-change targets and synthesis settings are held constant across anchor conditions.Syntheses leaving the destination effectively unchanged are dropped from the synthetic half of the batch.Source: §B.1 · p. 29
  • Internal results use checkpoints from the completed final epoch rather than validation- or test-selected epochs; external MVFA and MadCLIP replications follow their released selection rules.The latter selection rules are documented separately in Appendix C.Source: §B.1 · p. 29
  • Medical anchor-substitution tables average three runs per condition without seed-paired testing, while industrial randomized seeds represent both geometry draws and training runs.Correct-prompt anchors are deterministic; the architecture ablation pairs its three seeds across conditions.Source: §B.1 · p. 29

C EXTERNAL VALIDATION PROTOCOLS

External validations preserve the released visual and scoring pipelines while changing the anchor path. In MadCLIP, text-free learned and fixed scales approach or deliberately differ from the learned prompt scale.

MVFA architecture and controls

  • MVFA retains its released visual, optimization, scoring, and evaluation pipelines, changing only the anchor path.Its residual parameterization and MadCLIP’s jointly learned prompts lie outside the exact fixed-anchor, rotation-closed premise of Theorem 1.Source: C EXTERNAL VALIDATION PROTOCOLS · p. 29
  • MVFA encodes its released prompt ensemble as unit anchors with cosine ρ and compares correct prompts with seeded random and orthogonal frames.Random-frame seeds match training seeds; matched orthogonal controls restore the effective scale, with cosine and difference norms checked to 10−6.Source: C EXTERNAL VALIDATION PROTOCOLS · pp. 29–30

Checkpoint provenance and aggregation

MadCLIP architecture and controls

  • MadCLIP’s text-free learned-scale arm freezes a seeded orthonormal frame and trains only scalar θ, while its prompt-scale arm fixes ρ = +0.986.Both arms retain the visual branches, losses, data, schedules, and image and pixel scoring rules.Source: C EXTERNAL VALIDATION PROTOCOLS · p. 30
  • Across 48 MadCLIP dataset–support–seed cells, learned prompt and text-free cosines averaged −0.984 and −0.994, while the fixed prompt-scale arm remained at +0.986.Both fitted paths approached the maximum attainable effective scale, with or without language; the prompt-scale arm deliberately supplied a much smaller scale.Source: C EXTERNAL VALIDATION PROTOCOLS · p. 30

D FULL RESULT TABLES

The full result tables cover industrial, transfer, medical anchor-substitution, and adapter-family experiments. They distinguish supervision regimes and specify how metrics and aggregates are formed.

Industrial experiment

  • Table 22 reports full-shot industrial results across three vision backbones using linear adapters and synthetic-anomaly training.AUROC columns are category-macro means, while PRO, F1max, and AUPR are pooled over categories.Source: Table 22 · p. 31

Anchor transfer

  • Tables 23 and 24 measure anchor-swap cost under frozen transferred adapters, with Table 24 distinguishing coordinate-frame disruption from linguistic meaning.The source and target evaluations share domain shift and differ only in scoring anchors; the random condition changes readout direction without co-rotating the adapter.Source: Table 23 · p. 32, Table 24 · p. 32

Adapter architectures

  • Tables 25 and 26 report approximately parameter-matched linear, MLP, and Transformer adapters at K=16 under real-anomaly support and normal-only synthetic-anomaly training, with respective parameter counts of 3,148,800, 3,216,128, and 3,159,552.Table 25 uses CLIP ViT-L/14, and Table 26 replicates it on DINOv3-L/16; their normal-only blocks report image AUROC only.Source: Table 25 · p. 33, Table 26 · p. 34

Anchor substitution

  • Tables 27 and 28 compare anchor substitutions separately under real-anomaly support and normal-only training, with rows within each block sharing support, training, and scoring.Table 28 reports image-level AUROC; the supervision regimes are never pooled.Source: Table 27 · p. 35, Table 28 · p. 36

Complete result tables

  • The supplied DINOv3-L/16, DINOv2-L/14, and CLIP ViT-L/14 rows list correct prompts, mismatched-dataset bank, Gram-matched random, unmatched orthogonal, and matched orthogonal conditions.Source: D FULL RESULT TABLES · p. 35

Related papers

Loading 2610.34610v1…