PaperlayerView exact revision on arXiv

Source-linked AI summary

SAGE: Sense-Aligned Multimodal Grounding Evaluation with Gloss-Augmented Reasoning

Sunjae Kwon, Suhaas Kiran DG, Zonghai Yao, Talha Chafekar, Deepali Singh, Hong Yu

arXiv:2610.31838v1cs.CLcs.CV

TL;DR

Lexical grounding requires identifying which meaning of a target expression is intended in context and visually instantiated, yet performance can also rely on other compatible visual cues. This paper introduces a shared benchmark and a training-free reasoning framework to examine grounding across lexical settings. Definition augmentation improves all three benchmark subsets, while constraint-guided reasoning has setting-dependent results.

  • Problem

    Lexical grounding must distinguish a target's intended contextual meaning, but correct image selection can reflect other compatible visual cues instead.

    Source: §1 · p. 1

  • Method

    SAGE standardizes VWSD, VerSe, and WikiDiverse as 10-candidate image selection, while GAR compares definition grounding with constraint-guided visual verification.

    Source: §1 · pp. 1–2

  • Results

    Definition augmentation improves all three SAGE subsets, while constraint-guided reasoning helps VWSD and particularly VerSe but is less effective on WikiDiverse.

    Source: §1 · p. 2, §7 · p. 8

  • Takeaways & Limitations

    SAGE shows that noun-centered VWSD performance alone does not fully characterize behavior in verb- and entity-centered grounding.

    Source: §7 · p. 8

    The shared image-selection format does not equalize difficulty, since candidate confusability, context, and semantic granularity may vary across datasets.

    Source: Limitations · p. 9

Abstract

from arXiv · show

This paper introduces Sense-Aligned Grounding Evaluation (SAGE), a unified 10-candidate image-selection benchmark that reformulates VWSD, VerSe, and WikiDiverse into a shared term-centric protocol covering noun-, verb-, and entity-centered grounding. SAGE includes confusable hard negatives and a context-swapped VerSe diagnostic to probe target-term grounding beyond direct caption--image compatibility. We further propose Gloss-Augmented Reasoning (GAR), a training-free framework combining definition grounding with constraint-guided semantic verification. Experiments show that strong VWSD performance does not transfer uniformly across grounding settings. With Gemma3-12B, definition augmentation improves all three subsets, while constraint-guided reasoning improves VWSD and yields particularly strong gains on VerSe, but is less effective on WikiDiverse. Combining the two signals achieves the best macro-average performance, improving Hits@1 from 65.90 to 71.22. These results show that SAGE supports both unified evaluation and fine-grained analysis of explicit semantic reasoning across grounding settings.

1 Introduction

Lexical grounding requires identifying the intended meaning of a target in context and whether it is visually instantiated, while existing noun-, verb-, and entity-centered benchmarks use different task formulations. SAGE standardizes evaluation across these settings, and GAR examines how definitions and visual constraints support discrimination.

1 Introduction

  • Correct image selection can rely on compatible visual cues rather than discrimination of the intended meaning, a concern heightened for VerSe's naturally paired images.VWSD is mainly noun-centered, VerSe targets verb senses, and WikiDiverse focuses on named entities.Source: §1 · p. 1
  • SAGE reformulates VWSD, VerSe, and WikiDiverse as 10-candidate image selection with semantically confusable alternatives.The shared format enables comparison across noun-, verb-, and entity-centered grounding without requiring a common sense inventory or knowledge base.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Similar original VerSe scores of 91–92 Hits@1 diverge to 43.7–74.3 under SPIS, while target-verb neutralization drops performance by 3.4–33.0 points.These results reveal setting-dependent grounding behavior and verb sensitivity obscured by original image–caption pairings.Source: §1 · pp. 1–2
  • GAR is a training-free framework comparing context-generated and lexical-resource-grounded definitions, plus holistic ranking and constraint-guided visual verification.Constraint-guided reasoning decomposes intended meanings into object, attribute, and state constraints.Source: §1 · p. 2
  • Definition augmentation improves all three SAGE subsets, while constraint-guided reasoning helps VWSD and particularly VerSe but is less effective on WikiDiverse.Combining definition ranking and constraint-guided reasoning achieves the highest macro-average, though the best strategy varies by subset.Source: §1 · p. 2

2 Related Work

Prior multimodal ambiguity research spans visual puns, VQA-style tasks, and multilingual or utterance-level reasoning, while lexical grounding benchmarks differ in their targets and outputs. Related semantic-enrichment methods use generated or auxiliary text, but do not systematically compare information sources and uses across grounding settings.

Multimodal Lexical Grounding and Evaluation

  • Multimodal ambiguity has also been studied through visual puns, VQA-style tasks, and multilingual or utterance-level reasoning (Chung et al., 2024; Wang et al., 2025a,c).Source: §2 · p. 2
  • VWSD, VerSe, and WikiDiverse cover contextualized ambiguous terms, verb senses, and multimodal entity linking, respectively, but differ in task formulation and output space.SAGE adopts shared candidate-image selection to compare these settings.Source: §2 · p. 2

Semantic Enrichment and Visual Verification

  • DCLIP, CuPL, and WaffleCLIP augment class names with generated or auxiliary text, while VWSD methods use glosses, lexical knowledge, and generated definitions.These approaches demonstrate semantic enrichment's utility but do not systematically examine its sources and uses across grounding settings; GAR compares definition sources and ranking versus constraint-guided verification.Source: §2 · pp. 2–3

3 SAGE Benchmark

SAGE makes noun-, verb-, and entity-centered grounding comparable through a shared 10-image selection task. It reconstructs the source datasets for that format and adds a VerSe diagnostic that weakens caption–image pairing cues while preserving the target verb sense.

3.1 Task Setup

  • Each instance presents context, a target term, and 10 candidate images, and asks the model to select the image matching the term’s intended contextual meaning.Hits@1 measures the fraction of gold images ranked first; MRR averages their reciprocal ranks.Source: §3.1 · p. 3

3.2 Benchmark Datasets

  • VerSe instances pair a caption and target verb with nine distractors, preferentially drawn from images annotated with alternative OntoNotes senses of that verb.When fewer than nine hard negatives are available, random nonmatching MS-COCO images fill the candidate set.Source: §3.2 · pp. 3–4
  • SPIS replaces the caption-paired gold image with another image carrying the same OntoNotes verb sense, varying scene details to probe grounding beyond pair-specific compatibility.Combined with target-verb neutralization, it tests sensitivity to verb information when caption–image compatibility is weakened.Source: Figure 2 · p. 3, §3.2 · p. 4
  • SAGE standardizes VWSD, VerSe, and WikiDiverse as 10-candidate tasks with one gold image, while retaining their noun-, verb-, and entity-centered grounding settings.VWSD is used without modification; WikiDiverse retains its original split.Source: §3.2 · pp. 3–4, Table 1 · p. 4
  • Human validation found majority-vote agreement with SAGE gold labels in 91.2% of sampled VerSe and 94.9% of sampled WikiDiverse instances after excluding unanimous three-way disagreements.The 200-instance samples had Fleiss’ κ of 0.801 for VerSe and 0.823 for WikiDiverse.Source: §3.2 · p. 4

4 Gloss-Augmented Reasoning

Gloss-Augmented Reasoning represents a contextualized meaning both as a holistic definition and as visually verifiable constraints. It grounds definitions in lexical resources, expands them into checkable cues, and uses constraint verification to rank candidate images.

Context-Guided Definition Synthesis

  • GAR retrieves candidate senses or entities and canonical glosses, then synthesizes a concise definition of the contextually intended meaning without using candidate images or benchmark labels.The retrieved glosses guide synthesis but do not select a lexical-resource sense index.Source: §4 · p. 5

Definition Expansion

  • GAR expands the synthesized definition into a compact description that makes visually checkable cues, such as parts, shape, material, participants, actions, or states, more explicit.The expanded definition serves as a semantic specification for image ranking.Source: §4 · p. 5

Constraint-Guided Semantic Scoring

  • CGR decomposes the expanded definition into applicable object, attribute, and state constraints, then scores each candidate using the VLM’s affirmative probabilities for those constraints.Object constraints identify referents, attributes capture visible properties or relations, and states describe actions or conditions; the highest-scoring candidate is selected.Source: §4 · p. 5, Figure 3 · p. 5

5 Experiments

Across SAGE, model performance varies by grounding setting, and GAR’s signals do not help all subsets in the same way. Combining lexical-definition ranking with constraint-guided reasoning gives the strongest macro-average, while the SPIS diagnostic exposes verb sensitivity obscured by original image–caption pairing.

5 Experiments

  • The experiments compare image–text matching baselines and open- and closed-source VLMs on the same 10-image candidate sets, using significance tests for Hits@1 and MRR.McNemar’s test is used for Hits@1, and a paired t-test for MRR.Source: §5 · pp. 5–6

5.1 Experimental Results

  • Model-ranking agreement is strongest between VWSD and WikiDiverse (Kendall’s τb = 0.74), compared with 0.46 for VWSD–VerSe and 0.47 for VerSe–WikiDiverse.All off-diagonal correlations are significant at p < 0.05.Source: §5.1 · p. 6, Table 3 · p. 7
  • Model rankings and strengths differ across subsets: InternVL 3.5-2B excels on VerSe, while Qwen3-VL-8B is stronger on VWSD and WikiDiverse.The strongest model reaches only 74.3 Hits@1 on VerSe and 73.6 Hits@1 on WikiDiverse; InternVL 3.5-8B significantly underperforms its 2B counterpart on VerSe.Source: §5.1 · p. 6
  • Under SPIS, CLIP, Gemma3-12B, and GPT-5.6-Sol score 43.7, 59.3, and 74.3 Hits@1, respectively, while their original-pairing scores cluster at 91–92 Hits@1.Under SPIS-Neutralized, scores converge to 40.3–41.3 Hits@1; the corresponding drops are 3.4, 18.2, and 33.0 points.Source: §5.1 · p. 6, Table 4 · p. 7
  • Combining Lexical Def. and CGR achieves the strongest macro-average and best VWSD result, while lexical definitions improve all subsets and CGR helps VWSD and especially VerSe but degrades WikiDiverse.Generated definitions provide limited and inconsistent gains; the combined approach uses reciprocal rank fusion.Source: §5.1 · pp. 6–7
  • Object constraints are the strongest single semantic axis across all subsets; adding attribute and state constraints helps VWSD and VerSe but not WikiDiverse.On WikiDiverse, attribute and especially state constraints are substantially weaker than object-only reasoning.Source: §5.1 · p. 7
  • GAR-full reaches 81.2 Hits@1 and 88.6 MRR on VWSD without task-specific training, outperforming compared training-free approaches and exceeding Qwen2.5-VL-72B’s 79.1 Hits@1.It still trails the strongest fine-tuned systems.Source: §5.1 · p. 7

6 Discussion

Hard negatives are substantially more confusable than random alternatives, while CGR failure profiles vary across grounding settings. The discussion points to candidate-aware constraints, adaptive semantic-axis weighting, and more flexible integration as future directions.

6 Discussion

  • CGR failures motivate candidate-aware contrastive constraints and adaptive weighting of object, attribute, and state axes by discriminative value.The complementary strengths of lexical definitions and CGR also motivate integration beyond GAR’s fixed rank fusion.Source: §6 · p. 8
  • 26.38% vs. 11.59% in VerSe and 14.56% vs. 2.98% in WikiDiverse: hard negatives more often outrank the gold image than random negatives.Each negative is compared individually against the gold image.Source: §6 · p. 7, Table 8 · p. 8
  • 76% of WikiDiverse errors involve non-discriminative constraints, compared with 48% of VerSe errors; VWSD errors span upstream errors, decomposition drift, and non-discriminative constraints.Cross-axis conflicts occur in 6% of errors across all three subsets.Source: §6 · pp. 7–8, Table 9 · p. 8

7 Conclusion

The conclusion presents SAGE as a benchmark for noun-, verb-, and entity-centered grounding, alongside GAR, a training-free reasoning framework. Its results show that definition augmentation helps across subsets, whereas constraint-guided reasoning varies in effectiveness by setting.

7 Conclusion

  • SAGE and GAR address multimodal grounding across noun-, verb-, and entity-centered settings through a benchmark and a training-free reasoning framework.GAR combines definition grounding with structured semantic verification.Source: §7 · p. 8
  • Definition augmentation provides consistent gains across all three SAGE subsets.Source: §7 · p. 8
  • Constraint-guided reasoning improves VWSD, yields strong gains on VerSe, and is less effective for fine-grained entity grounding on WikiDiverse.Source: §7 · p. 8

Ethical Considerations

The paper describes source-specific image licensing, human evaluation and review of generated definitions, and the authors’ use of AI-assisted manuscript tools.

Ethical Considerations

  • SAGE uses publicly available resources, but image files will not be redistributed and remain subject to their original source-specific licenses and terms.Source: Ethical Considerations · p. 9
  • Unvetted LLM-generated definitions may still contain harmful or offensive content, although sampled definitions showed no violations of the ACL anti-harassment policy.Three annotators were recruited and paid at or above the local minimum wage.Source: Ethical Considerations · p. 9
  • AI-assisted tools supported manuscript drafting, editing, and refinement; the authors reviewed the content and retained responsibility for the study and final manuscript.Source: Ethical Considerations · p. 9

Limitations

This section outlines limitations of SAGE’s shared evaluation protocol, including factors that constrain direct comparisons of grounding difficulty and coverage. It also describes ambiguity in VerSe sense-level matching, alongside the diagnostic and inference-cost limitations of the framework.

Limitations

  • SAGE standardizes 10-candidate selection, but dataset differences prevent interpreting cross-setting performance gaps as intrinsic differences in grounding difficulty.Potential representation ambiguity was found in 4.0% of sampled WikiDiverse instances and multiple plausible same-sense realizations in 3.0% of sampled VerSe instances.Source: Limitations · p. 9
  • GAR depends on generated semantic representations and visual constraints, while its constraint-guided verification takes approximately 3.7× as long as base pointwise scoring.Overly generic constraints can hinder fine-grained entity discrimination; SAGE is also limited to English, and SPIS is not a fully controlled causal evaluation because replacing the gold image changes objects, participants, and scene context.Source: Limitations · p. 9

A Lexical Category Distribution in VWSD

B Evaluation Metrics

  • Hits@1 and MRR are the primary metrics because VWSD ranks candidate images; Hits@1 is equivalent to Accuracy.Hits@1 measures whether the correct image ranks first, while MRR captures its average inverse rank across instances.Source: B Evaluation Metrics · p. 12

C.1 Examples on VerSe

  • VerSe examples show that verb grounding requires sense disambiguation and integrated reasoning over event roles, affordances, motion, and scene context.Hard negatives can reuse the target verb in a different sense or suggest a related action while violating the intended event constraints.Source: §C.1 · pp. 12–13

C.2 Examples on WikiDiverse

  • WikiDiverse examples show that entity grounding requires world knowledge beyond keyword matching or low-level visual cues.Relevant distinctions include hurricane features and identity, UAV airframes, naming variants, and cultural or media meanings.Source: §C.2 · p. 17

D.1 Annotation Protocol and Interface

  • Inter-annotator agreement was strong for VerSe (κ = 0.801) and WikiDiverse (κ = 0.823), and majority-vote labels matched SAGE gold for 91.2% and 94.9%, respectively.Fleiss’ κ used original annotations over 200 instances per dataset; human–gold agreement excluded all-different cases and covered 176 of 193 VerSe and 185 of 195 WikiDiverse instances.Source: Table 11 · p. 21, §D.1 · p. 21

D.2 VerSe Mismatch Analysis

  • Annotators matched the VerSe gold in 88.5% of sampled instances; remaining disagreements primarily reflected event-level ambiguity under SPIS rather than exact caption–image compatibility.The qualitative mismatch categories include caption-specific versus sense-level match, competing sense or event cues, multiple plausible same-sense realizations, and weak or ambiguous visual evidence.Source: §D.2 · p. 22
  • In the “lean” example, the alternative may better match caption details even though the gold preserves the target verb sense.Both candidates evoke leaning, but their visual realizations and participants differ.Source: Figure 8 · p. 23
  • For “cover,” a candidate depicting protection or shielding can compete with the intended sense of being overlaid.The example makes the visual distinction between the two senses ambiguous.Source: Figure 9 · p. 24
  • For “catch,” several images make the target sense of seizing or stopping movement plausible despite differences in participants and settings.This illustrates how multiple same-sense event realizations can complicate image selection.Source: Figure 10 · p. 25

D.3 WikiDiverse Mismatch Analysis

The WikiDiverse reformulation achieved 94.0% human–gold agreement, with remaining mismatches reflecting several forms of entity-grounding ambiguity. Hard-negative composition varies by subset because VerSe and WikiDiverse provide different sources of confusable alternatives.

D.3 WikiDiverse Mismatch Analysis

  • The 12 mismatches comprised fine-grained knowledge requirements, benchmark label ambiguity, and representation ambiguity.Examples include a geographically specific county referent, competing circuit-map and race-scene readings, and physical versus institutional Parliament representations.Source: §D.3 · p. 26, Table 13 · p. 26, Figure 11 · p. 27
  • 94.0% of sampled WikiDiverse instances matched the SAGE gold labels, suggesting the reformulation is broadly reliable despite residual ambiguity.The 12 inspected mismatches fell into three broad categories.Source: §D.3 · p. 26
  • Residual mismatches reflect expected referent-level ambiguity and the limits of representing an entity with a single image.Named-entity selection can require choosing among plausible visualizations, and Wikipedia page images may not be uniquely representative.Source: §D.3 · p. 26

E Hard-Negative Composition

  • Each SAGE instance contains one gold image and nine negatives, with eligible hard negatives included before random negatives fill remaining positions.Table 14 reports average hard and random negatives per query.Source: E Hard-Negative Composition · p. 30, Table 14 · p. 30
  • VerSe hard negatives come from alternative senses of the same verb, so their availability is limited by the source data’s alternative senses and examples.Source: E Hard-Negative Composition · p. 30
  • WikiDiverse uses retrieved entity candidates with available images as hard negatives, yielding denser sets; these subset differences reflect candidate-generation mechanisms, not a fixed proportion.Source: E Hard-Negative Composition · p. 30

F Details on the Lexical Knowledge Base

The paper uses BabelNet for definition augmentation because it offers broad coverage and rich sense inventories. Coverage and ambiguity vary across lexical resources and datasets, affecting the available grounding information and the size of candidate-sense inventories.

F Details on the Lexical Knowledge Base

G Human Assessment of Synthesized Definition Plausibility

This section documents the assessment and procedures used to evaluate synthesized definitions and visually grounded semantic reasoning. It covers model-specific candidate-ranking interfaces and scoring sensitivity, definition generation, semantic decomposition and verification, and VWSD comparison and diagnostic settings.

G Human Assessment of Synthesized Definition Plausibility

  • 138/150 and 139/150 definitions were judged contextually plausible by annotators A and B, who agreed on 143/150 instances (95.3%; Gwet’s AC1 0.946).AC1 is reported because judgments were concentrated in the plausible category and it is less affected by category-prevalence imbalance than Cohen’s κ.Source: G Human Assessment of Synthesized… · p. 32
  • Eight outputs were jointly judged implausible, with failures involving domain disambiguation, action or posture senses, and fine-grained entity identification or uninformative definitions.The errors differed across VWSD, VerSe, and WikiDiverse; in three WikiDiverse cases, both annotators judged a synthesized definition plausible despite the intended meaning being absent from retrieved glosses.Source: G Human Assessment of Synthesized… · p. 32

H.1 Embedding-Similarity Ranking for Image–Text Matching Models

  • Image–text matching models rank candidates by pointwise similarity between a context-and-target text query and each candidate image, then evaluate rankings with Hits@1 and MRR.Scores are used only for within-instance ranking, so cross-example calibration is unnecessary.Source: §H.1 · p. 33, H Model-Specific Ranking Approach · p. 33

H.2 Token-Likelihood Ranking for Open-Source Generative VLMs

  • Open-source generative VLMs receive a binary YES/NO prompt, and their normalized affirmative-versus-negative token likelihoods provide the candidate ranking score.Deterministic inference uses first-position logits and sums capitalization variants to reduce sensitivity to response-label surface form and tokenization.Source: §H.2 · pp. 34–35
  • 76.5 and 75.6 Hits@1 for Qwen3-VL-8B and Gemma3-12B, respectively, accompany MRR scores of 85.4 and 84.7 under independent candidate ranking.The comparison uses pointwise token-likelihood ranking on VWSD; previously reported open-source results use joint multiple-choice prediction instead.Source: §H.2 · p. 35, Table 19 · p. 35

H.3 Scalar Compatibility Ranking for Proprietary VLMs

  • Proprietary VLMs rank candidates using generated integer compatibility scores from 0 to 100 because token-level log probabilities were unavailable for some models through the experimental API configuration.The same scalar interface is applied uniformly within the proprietary-model family, with Hits@1 and MRR used for evaluation.Source: §H.3 · pp. 35–36

H.4 Sensitivity to the Scoring Interface

  • YES/NO token-likelihood scoring matched or exceeded scalar scoring for both tested models across all three subsets, while scalar scoring substantially lowered Gemma3-12B performance.The authors retain model-native interfaces, noting that uniform scalar scoring may underestimate open-source VLM performance; aggregate conclusions remain consistent across interfaces.Source: §H.4 · p. 37, Table 21 · p. 37

I.1 Context-Guided Definition Synthesis; I.2 Definition Expansion

  • Definition synthesis produces one concise dictionary-style definition of the contextually intended sense from candidate glosses, then expands it with clarifying details constrained to preserve that sense.Expansion is instructed to keep the target meaning as the visual anchor and avoid sense drift or replacing it with surrounding objects or background facts.Source: §I.1 · p. 38, §I.2 · p. 39

I.3 Semantic Decomposition

  • Expanded definitions are decomposed into object, attribute, and state/event axes, preserving target-specific meaning while using concise, visually checkable cues and avoiding unsupported details.The definition-expansion prompt supplies a target word and synthesized definition; axes that cannot be visually checked are marked “not clearly specified.”Source: §I.2 · p. 39, §I.3 · pp. 40–41

I.4 Verification Question Generation

  • Each decomposed component becomes a concise yes/no question evaluated using directly visible evidence in a single image.Source: §I.4 · p. 42

I.5 Checklist Filling

  • For each candidate image, the probability of YES on each object, attribute, and state/event question serves as that axis’s compatibility score.Source: §I.5 · p. 44

J.1 Unsupervised VWSD

  • In zero-shot multiple-choice evaluation with the target expression and all 10 candidate images jointly, Qwen2.5-VL-72B-Instruct achieves 79.05 Hits@1 on original English VWSD without task-specific fine-tuning; MRR is unavailable because it outputs one top-1 prediction.Source: §J.1 · p. 45

J.2 Supervised VWSD

  • Supervised VWSD comparisons include fine-tuned CLIP on the official test split and FCLL, ARPA, and PolCLIP methods that use supervised learning for sense-aware multimodal alignment.Source: §J.2 · p. 45

K Reciprocal Rank Fusion Details

L Target Verb Neutralization

  • The VerSe neutralized-caption diagnostic masks the target verb and replaces it with a generic placeholder, using validation checks and fallback replacements to prevent reintroducing the verb.This is intended as a diagnostic ablation, not a fluency-preserving paraphrase procedure.Source: L Target Verb Neutralization · p. 46

M Detailed on Qualitative Analysis on Error Cases

  • CGR failures are organized into five recurring categories with two representative examples per subsection; reported scores average valid object, attribute, and state questions.Invalid or unspecified axes are omitted from the arithmetic mean.Source: M Detailed on Qualitative Analysis on… · p. 48

M.1 Upstream Semantic or Referent Errors

The appendix traces GAR errors to upstream sense selection, question decomposition, conflicting constraints, and visual verification, while also documenting data-release conditions and runtime costs.

M.1 Upstream Semantic or Referent Errors

  • Upstream definition errors can make generated questions evaluate the wrong sense or referent, causing distractors to outrank the gold image.For VWSD #129, the flashlight distractor receives P(YES) = 0.9997 versus 0.0000 for the gold image; for WikiDiverse #604, the deity distractor mean is 0.999702 versus 0.001587 for the airline image.Source: §M.1 · p. 48, Figure 15 · p. 49

M.2 Decomposition Drift

  • Decomposition drift can redirect questions from a correctly identified sense toward another sense or visual prototype, reversing candidate support.For Milk River, a dairy-specific Attribute cue reverses the valid-axis mean; for facelift, all three questions shift to physical lifting.Source: §M.2 · p. 50, Figure 16 · p. 50, Figure 17 · p. 51

M.3 Cross-Axis Conflict

  • Individually plausible constraints can favor different candidates, and averaging them can reverse the ranking or leave near-tied means.This occurs for both the Phnom Penh Airport logo case and VerSe #956 (fly).Source: §M.3 · p. 52, Figure 18 · p. 52, Figure 19 · p. 53

M.4 Non-Discriminative Constraints

  • Broad constraints may fit multiple candidates, as when generic fighter-jet and gray cues fit both a Su-27 and an F-16.The questions omit engine-count and tail-configuration distinctions, leaving the distractor with a slightly higher mean.Source: §M.4 · p. 54, Figure 21 · p. 55

M.5 Visual Verification Errors

  • Visual verification can accept evidence contradicted by the image, including a standing dog as biting and signage naming Moorgreen as Stobhill Hospital.For VerSe #494, the distractor receives P(YES) = 0.9047; in WikiDiverse #1013, the distractor mean reaches 0.989622 despite contradictory signage.Source: §M.5 · p. 56, Figure 22 · p. 56, Figure 23 · p. 57

N Data Release Plan

  • The planned benchmark release includes fixed instances, ordered candidates, labels, source mappings, collection-time digests, and reconstruction and validation scripts.Because externally hosted WikiDiverse images may change or disappear, future reconstructions may not be pixel-identical; users are asked to report reconstruction dates and image discrepancies.Source: N Data Release Plan · p. 57

O Runtime Comparison

  • On VWSD, CGR required 2,311 seconds, approximately 3.7× the 627-second baseline runtime.The complete definition-based setting took 796 seconds, and candidate-wise constraint evaluation accounts for most of GAR’s computational overhead.Source: O Runtime Comparison · p. 57, Figure 24 · p. 57
Loading 2610.31838v1…