PaperlayerView exact revision on arXiv

Source-linked AI summary

Inductive Visual Logic for Few-Shot Out-Of-Distribution Adaptation in VLMs

Hung-Jen Chen, Yu-Heng Ho, Ting-Yao Huang, Po-Hsiang Hsu, Li-Yu Chen, Chun-Yi Lee, Min Sun

arXiv:2609.38362v1cs.CV

TL;DR

Specialized distant-OOD domains challenge VLMs because required discriminative features may be absent from their learned representations, even when descriptive ability remains. IVL builds class traits from few-shot support images and uses hierarchical filtering to ground query predictions in visual evidence. Across distant-OOD benchmarks, it achieves the highest aggregate accuracy under two VLM backbones while remaining training-free.

  • Problem

    In specialized distant-OOD domains, VLMs may lack the discriminative features needed for classification, while standard adaptation operates within the encoder’s existing feature space.

    Source: §1 · p. 1

  • Method

    IVL uses semantic and primitive visual descriptions of few-shot support images to build per-class trait dictionaries, then filters for spatially grounded trait evidence when classifying queries.

    Source: §1 · p. 3, Dual-Mode Trait Extraction · p. 9

  • Results

    IVL achieves the highest aggregate accuracy across distant-OOD benchmarks under Qwen-VL and LLaVA in 1-shot and 8-shot settings.

    Source: §1 · pp. 3–4

  • Takeaways & Limitations

    The results support trait-based reasoning as a training-free approach to distant-OOD VLM adaptation, with predictions traceable to traits and spatial grounding.

    Source: §5 · p. 15

    IVL depends on classes differing in traits the model can name; when they do not, accuracy remains near chance and trait-based reasoning reaches an applicability boundary.

    Source: §4.3 · p. 14

Abstract

from arXiv · show

Generative vision-language models (VLMs) such as Qwen-VL and LLaVA achieve strong zero-shot performance on tasks overlapping with their pretraining distribution, yet fail on specialized domains where the required discriminative features were never learned, a regime we term distant out-of-distribution (OOD). Standard adaptation methods cannot overcome this representational absence because they operate within the encoder's existing feature space. However, VLMs retain a robust descriptive capacity even when discrimination collapses: a model that cannot classify a medical scan can still articulate its visual patterns. Exploiting this asymmetry, we introduce Inductive Visual Logic (IVL), a training-free framework that constructs classification knowledge from the model's surviving descriptive ability. IVL extracts visual traits from few-shot support images through dual-mode prompting, combining semantic descriptions with primitive visual observations, and organizes them into per-class trait dictionaries. At inference, hierarchical filtering identifies spatially grounded trait evidence for classification. Across multiple distant-OOD benchmarks, IVL achieves the highest aggregate accuracy under two VLM backbones while producing interpretable, trait-traceable predictions.

1 Introduction

In distant-OOD domains, VLMs lack the discriminative features needed for classification, and standard adaptation methods cannot reliably create that missing structure. IVL instead builds trait-based classification from surviving descriptive capacity, achieving the highest aggregate accuracy across distant-OOD benchmarks under two VLM backbones.

1 Introduction

  • Distant-OOD adaptation is limited by representational absence: gradient-based methods refine existing features, while prompt-based and in-context approaches remain anchored to pretrained representations.The paper distinguishes this from ordinary distributional shift, where relevant discriminative representations remain available.Source: §1 · pp. 1–2
  • VLMs can retain descriptive capacity when discrimination collapses, enabling classification rules to be built from observed visual traits rather than opaque internal representations.The paper describes this as an asymmetry between descriptive generation and discriminative representations.Source: §1 · pp. 2–3
  • IVL extracts traits from few-shot support images using semantic descriptions and low-level visual observations, then filters for spatially grounded evidence at inference.Unlike attribute-based methods that generate descriptors solely from pretrained semantic knowledge, IVL grounds traits in support-image observations.Source: §1 · p. 3
  • IVL achieves the highest aggregate accuracy across distant-OOD benchmarks under Qwen-VL and LLaVA in 1-shot and 8-shot settings, with interpretable, evidence-grounded predictions.The framework is training-free.Source: §1 · pp. 3–4

2 Related Work

Generative VLMs perform strongly on familiar visual concepts but degrade on specialized domains, motivating few-shot adaptation. Existing approaches adapt prompts, attributes, or parameters, while IVL’s related-work distinction is to ground traits in direct support-image observations.

2.1 Generative Vision-Language Models

  • Generative VLMs use language-model backbones to produce text conditioned on visual inputs, but their performance degrades on specialized domains where natural-image semantics do not transfer.Parameter-efficient methods such as LoRA have limited effectiveness under large domain gaps.Source: §2.1 · p. 4

2.2 Few-Shot Domain Adaptation with VLMs

  • Few-shot adaptation includes learned prompts, attribute guidance, parameter-efficient fine-tuning, cache-based adaptation, and RL-guided fine-tuning; the paper reports these paradigms fail on distant-OOD tasks.The cited approaches rely on recombining or adapting pretrained representations.Source: §2.2 · p. 4

2.3 Trait-Based and Neuro-Symbolic Approaches

  • Prior attribute approaches use explicit concepts or generated descriptors, but IVL grounds semantic and low-level traits in support images rather than relying on internal concept composition.The related work notes that concept bottleneck models require extensive annotations and that prior VLM descriptor approaches do not address few-shot learning from novel domains.Source: §2.3 · pp. 4–5

3 Methodology

IVL formulates distant-OOD few-shot classification as matching query images to traits extracted from labeled support examples, without updating model parameters. Its method motivates this design through representational limits, diagnoses OOD structure using frozen-encoder metrics, and builds and filters trait evidence for classification.

3.1 Problem Formulation

  • IVL represents each class with natural-language visual traits extracted from labeled support images and classifies queries by matching visually grounded traits to classes.The task is posed without updating VLM parameters.Source: §3.1 · p. 5

3.2 Theoretical Foundations

  • The theoretical framing states that optimization within the encoder representation subspace incurs irreducible excess risk when target-task structure lies outside that subspace.IVL instead uses the decoder’s preserved descriptive capacity to organize and spatially ground primitive attribute descriptions.Source: §3.2 · pp. 6–7

3.3 Characterizing Representational Absence

  • Distant-OOD status is assessed using frozen-encoder feature-space divergence and cluster-label alignment; datasets qualify when MK-MMD is high or NMI is low.The reported thresholds are τMMD = 0.49 and τNMI = 0.42; the paper reports near-OOD 1-shot SFT mean ∆= +18.6% and distant-OOD ∆from −27.5% to +36.9%.Source: §3.3 · pp. 7–8

3.4 Framework Overview

  • The framework has two stages: build a per-class trait database from support images, then classify queries by finding traits in that database that are visually grounded.The pipeline requires no parameter updates and produces predictions traceable to traits and spatial grounding.Source: §3.4 · p. 8

3.5 Trait Building Stage

  • Trait building uses semantic and low-level prompts, then clusters, canonicalizes, and filters the extracted traits into a per-class vocabulary.Semantic prompts use retained domain knowledge; low-level prompts target primitive colors, textures, and shapes while prohibiting category labels. Their combined coverage exceeds either mode alone when the sets are not fully redundant.Source: §3.5 · p. 8, Dual-Mode Trait Extraction · pp. 9–10

3.6 Inference Stage

4 Experimental Results

Across 18 distant-OOD tasks, IVL performs strongly under both VLM backbones, while the experiments also identify pipeline failures and domain-dependent ablation results.

4.1 Experimental Setup

  • Across 18 tasks, IVL is evaluated on Retinal OCT, WM811k, 15 MVTec AD categories, and Pokémon using shared few-shot supports and Qwen2.5-VL or LLaVA backbones.MVTec AD is treated as a stress test for general-purpose VLMs, not compared with specialized anomaly-detection pipelines.Source: §4.1 · pp. 11–12

4.2 Quantitative Results

  • IVL’s trait building uses one 32 GB GPU and 67% fewer GPU-seconds than SFT, while per-image inference adds moderate overhead from VLM prompt processing.The retrieval pipeline itself is not the main source of inference overhead.Source: §4.2 · p. 12
  • IVL achieves the best results on Retinal OCT (20.83%), WM811k (16.51%), and Pokémon (50.30%), and ranks top-two on 12 tasks with Qwen2.5-VL and 10 with LLaVA.Under LLaVA, it leads on Pokémon, WM811k, and MVTec AD mean, and matches the best method on Retinal OCT (33.70% vs. 34.30%) without parameter updates.Source: §4.2 · p. 12, Table 1 · p. 13

4.3 Failure Modes

  • Reference-free vocabulary collapse yields 11.6% accuracy on MVTec AD Transistor, while dictionary coverage failure accounts for 48.4% of Pokémon errors.The first arises when benign within-tolerance variation is read as a defect; the second occurs when filtering drops the true class from the candidate dictionary.Source: §4.3 · p. 13

4.4 Ablation Studies

  • On MVTec AD, low-level prompting beats semantic prompting in 10 of 15 categories, while dual-mode extraction beats the better single mode in 12 of 15.Transistor, Toothbrush, and Zipper are exceptions; the comparison is a 1-shot, three-run mean per mode.Source: Table 2 · p. 14, Dual-mode prompt design · p. 14

4.3 Failure Modes

  • IVL’s trait-based reasoning is limited when classes differ only in traits the model cannot name, leaving accuracy near chance, as with FractalDB’s generative-parameter labels.The authors identify parametric or hybrid methods as still needed for this boundary case.Source: §4.3 · p. 14

4.4 Ablation Studies

  • On Pokémon, semantic-only (32.6%) outperforms low-level-only (27.8%); removing visual grounding costs −4.2 pp, salience filtering −2.6 pp, and all hierarchical filtering −2.1 pp.The component ablation uses a 144-sample balanced subset with Qwen2.5-VL at 1-shot.Source: Table 3 · p. 14, Component ablation · p. 15

5 Conclusion

IVL addresses distant-OOD VLM failures with training-free trait-based reasoning, grounding classifications in spatially localized evidence. The paper reports success across distant-OOD benchmarks and identifies open-set recognition and compositional generalization as future directions.

5 Conclusion

  • IVL constructs classification features through dual-mode trait extraction and hierarchical filtering, grounding predictions in spatially localized evidence.It uses trait-based reasoning rather than parameter adaptation.Source: §5 · p. 15
  • Experiments across distant-OOD benchmarks confirm that IVL succeeds where parametric methods fail.Source: §5 · p. 15
  • Open-set recognition and compositional generalization are identified as promising directions for future work.Source: §5 · p. 15

Supplementary Overview

The supplement collects theoretical derivations, empirical analyses, and implementation details supporting the main paper, from regime measurements to failure modes and reproducibility.

Supplementary Overview

  • The supplement provides theoretical derivations, empirical analyses, and implementation details supporting the main paper.Source: Supplementary Overview · p. 19
  • Its analyses cover per-dataset MK-MMD/NMI measurements, cross-model robustness, trait evidence, component ablations, failure modes, and computational costs.Source: Supplementary Overview · p. 19
  • The supplement also includes dataset and human-evaluation protocols, non-generative VLM comparisons, reproducibility details, and qualitative comparisons with Visual-RFT.Source: Supplementary Overview · p. 19

Notation Reference

The supplement formalizes analyses of distant-OOD VLM behavior under stated assumptions, presenting them as principled analyses rather than novel mathematical theorems. Its Proposition 1 derivation contrasts declining discriminative capacity with surviving descriptive capacity.

Notation Reference

A.1 Proof of Proposition 1: Descriptive-Discriminative Asymmetry

  • Under assumptions linking encoder specialization and feature-space divergence, increasing divergence is associated with greater class-conditional overlap and declining discriminative capacity.The derivation argues that mutual information between encoder representations and target labels approaches zero as overlap increases.Source: §A.1 · pp. 21–22
  • The derivation argues that descriptive capacity survives because primitive attributes such as colors, textures, shapes, and spatial relations are compositional and domain-agnostic.Its stated lower bound depends on visual-vocabulary coverage and visual-to-language grounding, not feature-space divergence.Source: §A.1 · pp. 21, 23
  • The derivation assumes that the divergence–mutual-information relationship is monotonic only on average, with accidental feature alignment permitting dataset-specific exceptions.The Pokémon benchmark is described as a boundary case: low MK-MMD but low NMI.Source: §A.1 · p. 23

A.2 Proof of Proposition 2: Gradient Futility under Representational Absence

The supplementary analyses formalize an excess-risk floor from missing encoder directions, examine complementary semantic and low-level traits, and detail how MK-MMD and NMI characterize distant-OOD datasets. The authors report volatile 1-shot SFT gains in distant-OOD settings, contrasting with positive gains across near-OOD datasets.

A.2 Proof of Proposition 2: Gradient Futility under Representational Absence

  • For updates constrained to the encoder’s pretraining-supported subspace, squared-loss excess risk is lower-bounded in proportion to the missing component of the optimal classifier.The stated bound is proportional to ||w†⊥||²/||w†||²; the 0–1 risk extension requires standard margin and tail conditions.Source: §A.2 · pp. 24–25
  • In few-shot settings, LoRA or SFT updates are poorly constrained: nC training points span at most nC dimensions, far fewer than d for typical settings.The text links this limited signal to unstable adaptation and spurious correlations; the reported 1-shot MVTec gains vary by category.Source: §A.2 · p. 25

A.3 Dual-Mode Complementarity Analysis

  • Semantic trait coverage degrades with divergence, while low-level trait coverage remains approximately stable; their union covers at least as many class pairs as either mode alone.The combined coverage is strictly greater when the modes provide non-redundant discriminating traits for a class pair.Source: §A.3 · p. 26
  • The dual-mode ablation reports that removing either trait mode lowers accuracy relative to the full system on distant-OOD datasets.Neither single mode recovers full dual-mode accuracy on Pokémon.Source: §A.3 · pp. 26–27

B Characterizing Representational Absence: Detailed Analysis

  • The distant-OOD characterization uses high MK-MMD and/or low NMI; the 22 non-reference datasets are assigned using thresholds on these measures.The two-stage thresholding applies MK-MMD first and then NMI on the near-side subset; 1-shot SFT gains are validation, not a defining criterion.Source: B Characterizing Representational… · p. 27, §B.1 · pp. 27–28

B.1 Per-Dataset MK-MMD and NMI Values

  • 1-shot SFT gains range from −27.5% to +36.9% across MVTec categories, while all near-OOD datasets show positive gains.Distant-OOD gains are described as near-zero on average with high variance, indicating unreliable rather than uniformly ineffective adaptation.Source: Table 5 · p. 28, §B.1 · p. 29
  • Retinal OCT is assigned to distant-OOD by MK-MMD = 0.578 despite NMI = 0.564, illustrating that MK-MMD can detect a domain gap NMI alone would miss.Source: §B.1 · p. 29
  • Near-OOD datasets serve as calibration references rather than primary IVL targets, since their low MK-MMD and high NMI indicate preserved encoder discriminative capacity.The text says IVL is designed for cases where this condition fails and is not expected to have a structural advantage when discriminative capacity remains intact.Source: §B.1 · p. 29

B.2 Cross-Model Robustness

The supplementary analyses test the OOD regime assignments across encoders, thresholds, and reference distributions, then examine trait extraction, grounding, and classification evidence. Pets-37’s LLaVA results support its distant-OOD assignment, while trait evidence remains informative across the evaluated domains.

B.2 Cross-Model Robustness

  • 31.34% at 1-shot and 33.18% at 8-shot: IVL outperforms all baselines on Pets-37 under LLaVA, where the framework assigns the dataset to distant-OOD.The assignment follows low NMI (0.379 < τNMI = 0.45); IVL also outperforms baselines under the reported downstream evaluation.Source: Table 6 · p. 30, §B.2 · pp. 30–31

B.4 Threshold Robustness

  • All 22 Qwen2.5-VL dataset assignments remain unchanged when τNMI is swept from 0.30 to 0.60 or when the MK-MMD reference changes to ReLAION-400M.Under the alternative reference, τMMD shifts from 0.49 to 0.45 while τNMI remains 0.42.Source: §B.4 · p. 32, §B.5 · p. 32, Table 8 · p. 32

B.6 Additional Benchmarks

  • 49.0%: IVL exceeds zero-shot (33.8%) on APTOS, while one-shot SFT falls to 15.8%, extending the reported medical-domain pattern beyond Retinal OCT and WM-811K.APTOS has MK-MMD 0.616 and NMI 0.200, placing it in the distant-OOD regime.Source: §B.6 · p. 33

C Trait Extraction Statistics Across the OOD Spectrum

  • Across 22 datasets, extraction yields 11.2–27.1 traits per image and MeanNormIDF of 0.733–0.972, with neither metric systematically declining as MK-MMD increases.Lower MeanNormIDF on WM-811K and Retinal OCT is attributed to inter-class visual ambiguity, while low-level extraction alone has non-trivial accuracy across benchmarks.Source: C Trait Extraction Statistics Across… · p. 34, Table 9 · p. 34

D Trait Grounding and Class Alignment

  • 81.2% of correct Pokémon predictions are explainable through trait evidence, and all MVTec categories except Toothbrush exceed their category-specific random alignment baseline by 1.9–5.7×.MVTec alignment varies with defect distinctiveness; explicit reasoning traces distinguish errors involving overlapping traits from those involving insufficient trait coverage.Source: D Trait Grounding and Class Alignment · pp. 35–37

D.1 Per-Class Trait Signatures and D.2 Per-Trait Grounding Visualization and Comparison to SFT

  • 15/18 representative Pokémon signatures and all five Hazelnut class signatures are human-recognizable; on the Hazelnut normal sample, IVL reasoning rejects defect traits while SFT’s none activation collapses.SFT’s crack/none activation ratio reaches 108×, whereas IVL identifies the region as the nut cap and predicts none.Source: §D.1 · p. 37, §D.2 · pp. 37–38, Figure 7 · p. 38

E Ablation Studies

Ablations on a balanced 144-sample Pokémon test subset show that multiple pipeline choices contribute to performance, while the selected defaults prioritize robustness beyond that subset. The primary evaluations target distant-OOD benchmarks, with few-shot support sets sampled by class.

E Ablation Studies

  • Ablation results are specific to a balanced 144-sample Pokémon test subset and are not directly comparable to full-test-set results; the default configuration prioritizes robustness and consistency.Removing or simplifying modules often changes accuracy, indicating contributions from interacting design choices.Source: E Ablation Studies · pp. 38–39
  • 6.2% lower accuracy with semantic-only extraction and 11.1% lower with low-level-only extraction support the dual-mode strategy’s combination of semantic structure and primitive visual features.The authors attribute semantic-only degradation to generic descriptions in novel domains and low-level-only degradation to insufficient structural coherence.Source: E Ablation Studies · p. 39
  • Removing both filtering stages costs −2.1%, while over-filtering (k2=50) costs −5.6%; global-only filtering gains 0.7% on this subset, and local refinement supplies patch-level attention maps for grounding.Removing grounding entirely costs −4.2%; the reported net contribution of local refinement is +3.5 pp once grounding is included.Source: E Ablation Studies · p. 39
  • K-means gains 2.1% on the balanced subset, but HDBSCAN remains the default because it needs no preset cluster count and preserves rare traits in unequal-density settings.The authors identify unequal trait density as typical of real-world few-shot scenarios beyond the controlled subset.Source: E Ablation Studies · p. 40
  • The specialized Sentence Transformer outperforms the higher-dimensional Qwen VLM encoder for trait clustering, while salience filtering removal lowers accuracy by 2.6%.The paper attributes the embedding result to finer semantic granularity and the filtering result to avoiding semantically inconsistent traits.Source: E Ablation Studies · p. 40

F.1 Public Datasets

  • The primary IVL evaluation targets Retinal OCT, WM-811k, and MVTec AD; MVTec’s 15 categories are evaluated as independent recognition datasets.Classification support sets use 1 or 8 images per class; each MVTec category uses one image per defect type for one-shot evaluation.Source: §F.1 · p. 41

F.2 Pokémon Dataset

The Pokémon benchmark compares human and model performance on normal and specially designed cases, while the accompanying analyses expose limits of IVL when traits are ambiguous or not class-separable. The computational analysis evaluates adaptation costs on a controlled Pokémon setup.

F.2 Pokémon Dataset

  • The 1,025-species benchmark uses 144 support images and 881 test images; support examples are single-type, while dual-type test answers allow either or both correct types.The 144-image support pool provides eight images for each of 18 types.Source: §F.2 · pp. 41–42
  • Human accuracy was 56.1%, 61.9%, and 85.2% for beginner, intermediate, and expert groups; it fell nearly 20% on special cases while model performance stayed approximately stable.The special cases are Ultra Beast and Paradox Pokémon, designed to decouple appearance from type.Source: §F.2 · p. 43

G.1 MVTec AD Transistor Category: Vocabulary Collapse

  • IVL predicts bent_lead for several transistor states, including none, cut_lead, damaged_case, and misplaced.Naturally curved leads also appear in normal and other defect samples; analyzing images independently without a normal reference makes curvature appear sufficient evidence for bent_lead.Source: §G.1 · p. 43

G.2 FractalDB: Applicability Boundary

  • On 20-class FractalDB under Qwen2.5-VL (1-shot), zero-shot reaches 0.3%, SFT 5.5%, and IVL 4.9%, with no IVL advantage.IFS-family labels can yield visually similar patterns and non-discriminative language descriptions; distant-OOD assignment does not establish trait-separability.Source: §G.2 · p. 44

H Computational Analysis

  • The efficiency analysis compares IVL, SFT, and ICL on Pokémon (18 classes, 1-shot) using Qwen2.5-VL-7B on H100 GPUs, separating offline and online costs.Offline cost is a one-time adaptation expense, while online cost is measured per image.Source: H Computational Analysis · p. 45

H.1 Efficiency Comparison Across Adaptation Methods

IVL is compared with adaptation and in-context baselines for compute, latency, and accuracy, and its class-count scalability and prompting and evaluation protocols are described. The efficiency results show accuracy gains alongside reduced Pokémon adaptation compute, with moderate per-image latency overhead.

H.1 Efficiency Comparison Across Adaptation Methods

  • IVL exceeds SFT by +4.8 pp on Pokémon and MVTec AD, and Visual-RFT by +8.0 pp on MVTec AD; Pokémon adaptation uses 67% fewer GPU-seconds than SFT.IVL's offline trait-building pipeline uses 603 GPU-seconds on one 32 GB GPU, while SFT uses 1,854 GPU-seconds and requires at least three H100 GPUs; IVL inference takes 7.06 s/image versus 4.33 s for SFT.Source: §H.1 · pp. 45, 47, Table 20 · p. 47

H.2 Scalability with Number of Classes

  • As classes grow from 20 to 100, online latency remains nearly constant, with ratios of 0.90–1.00; per-class offline cost is stable or slightly lower, at ratios of 0.86–0.93.Filtering caps the traits entering the final VLM pass independently of database size; the higher absolute latency reported for Aircraft reflects its larger trait list and higher-resolution inputs.Source: §H.2 · pp. 47–48, Table 21 · p. 48

I Non-Generative Baselines

  • The evaluation includes CLIP ViT-B/32 adapter and prompt baselines, plus DIFO on ResNet features, for distant-OOD benchmarks.The supplied passage identifies the methods and evaluation scope but does not report their outcomes.Source: I Non-Generative Baselines · p. 48, Table 22 · p. 48

J.1 IVL Pipeline Hyperparameters

  • Hyperparameter defaults were fixed before experiments and not tuned per dataset; the adaptive k1 rule retains 80% of canonicalized traits within minimum and maximum bounds.The bounds prevent very small databases from being over-filtered and very large ones from saturating attention computation.Source: §J.1 · pp. 48–49

J.2 IVL Prompts

  • IVL's accuracy advantage is attributed to support-image-grounded traits rather than prompt structure alone.Comparable structured-descriptor baselines trail by more than 13 pp on MVTec AD mean, while removing localized evidence costs −4.2 pp.Source: §J.2 · pp. 49–50

J.3 Baseline

  • Baseline comparisons use the same backbone and unified training and evaluation protocols so performance differences reflect methodological design rather than backbone variance.Menon et al. and CuPL components were replaced with the corresponding unified experimental backbone.Source: §J.3 · pp. 51–52

J.4 In-Context Learning Evaluation

  • The ICL baseline uses fixed, class-balanced few-shot demonstrations in canonical order and excludes test images from demonstrations.The protocol uses the same deterministic support sampling as other few-shot methods and requires no parameter updates.Source: §J.4 · p. 52

K Qualitative Results

The qualitative examples contrast Visual-RFT's fluent but sometimes evidence-inconsistent explanations with IVL's trait-based reasoning. IVL links its predictions to explicit visual observations and cited evidence.

K Qualitative Results

  • In a hazelnut example, Visual-RFT interprets a shell fissure as a natural opening, predicts none, and misses the crack-like anomaly.Its explanation is fluent but inconsistent with the visual evidence.Source: K Qualitative Results · p. 53
  • For the same example, IVL identifies the fissure as evidence for the crack trait and predicts the corresponding anomaly class.Its process compares explicit visual observations against a predefined trait dictionary.Source: K Qualitative Results · p. 53
  • The qualitative examples show IVL linking final predictions to cited visual evidence, providing a transparent basis when visual evidence is subtle.Source: K Qualitative Results · p. 53

Related papers

Loading 2609.38362v1…