Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models

Chao Gou; Jiayu Tang; Shuo Yang; Tat-Seng Chua; Wenhao Yang; Xiaobo Xia; Xiaoyan Xiao; Yuchen Zhou; Yuqin Dai

REVIEW 4 major objections 4 minor 2 cited by

Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Logic-aware training lifts CLIP by 40+ points on logic tests

desk verdict A broad new logic benchmark and a plausible fine-tuning recipe whose headline gains are undermined by a shared LLM-perturbation pipeline between training and test data. read the letter →

arxiv 2508.11317 v1 pith:6P2WH2BW submitted 2025-08-15 cs.CV cs.MM

Yuchen Zhou , Jiayu Tang , Shuo Yang , Xiaoyan Xiao , Yuqin Dai , Wenhao Yang , Chao Gou , Xiaobo Xia

show 1 more author

Tat-Seng Chua

This is my paper · ORCID

classification cs.CVcs.MM

keywords vision-languagemodelslogicalreasoningbenchmarkCLIPcontrastivelearninghardnegativeminingmultiple-choicemultimodalretrievallogic-awaretraining

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that mainstream vision-language models such as CLIP largely ignore logical structure, and that the gap can be measured and largely closed. The authors build LogicBench, a 51,360-pair benchmark spanning images, videos, traffic anomalies, and medical reports across nine logical relations, and find that existing models run more than 40 accuracy points below human scores, especially on causality, conditionality, temporality, and negation. They then train LogicCLIP on hard negative captions generated by five LLMs from MSCOCO captions, using a loss that combines standard contrastive alignment, multiple-choice discrimination, and explicit logical-category classification. On LogicBench, LogicCLIP raises Image MCQ from roughly 38% to 82% for CLIP-B and from 37% to 84% for CLIP-L, while COCO and Flickr30K retrieval either hold or improve. The paper's central claim is that logical blindspots are a trainable defect, not an inherent limitation of contrastive vision-language models.

What carries the argument

LogicCLIP's optimization objective is a weighted sum of three losses: the standard CLIP symmetric contrastive loss, a multiple-choice cross-entropy loss over one positive and three LLM-generated hard negative captions, and a binary cross-entropy loss that trains a text-side classifier to predict which of nine logical categories a caption instantiates. The supporting machinery is the data pipeline: spaCy and regular-expression filtering harvest human captions rich in logical structure from MSCOCO, and five LLMs generate three logically perturbed negatives per positive, with human review. The multiple-choice objective teaches fine-grained discrimination; the structure-aware loss pushes text em

What would settle it

Compare models on LogicBench against a control set in which the negative options are generated by a different process—for example, human editors rewriting the same captions, or LLMs from a family not used in training—matched for length and lexical overlap but with the logical relation preserved or broken. If LogicCLIP's margin over CLIP collapses to near zero on such controls, the reported logical understanding is largely an artifact of the specific LLM perturbation style.

Watch

Extended reading notes

Core claim

The paper's central discovery is that CLIP-style vision-language models have a specific, measurable weakness: they encode surface semantics—objects and attributes—but not the logical relations that reorganize a sentence's meaning. Using LogicBench, the authors show that current models sit more than 40 accuracy points below human performance on logical multiple-choice questions, with the largest deficits in causality, temporality, conditionality, and negation. The paper further claims that this weakness is correctable: fine-tuning with LogicCLIP, which adds hard negative captions generated by five LLMs and a logical structure-aware classification loss to standard contrastive training, raises

Load-bearing premise

The claims depend on the assumption that the LLM-generated negative captions differ from positive captions only in logical structure, not in surface cues like word frequency, phrasing artifacts, or plausibility; if that fails, the measured blindspot and the training gains partly reflect detection of those cues.

Editorial extensions

If this is right

If LogicCLIP is right, contrastive VLMs can learn logic without a new architecture; fine-tuning with hard negatives and a structure-aware loss is sufficient.
LogicBench gives the field a standardized diagnostic across four application domains and two task formats, so logical blindspots can be tracked rather than assumed.
Because image-only training transfers to video, anomaly, and medical scenarios, logical understanding learned from everyday captions appears to generalize across visual domains.
General retrieval also improves, so logic-aware training can be added to CLIP-style pretraining without the usual alignment-versus-reasoning trade-off.
The category-level results suggest the hardest relations—causality, temporality, conditionality, negation—are exactly where targeted logical data yields the largest gains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Going beyond the paper: a decisive test is whether LogicCLIP's gains survive when evaluation negatives come from a different generator family or from human-written violations; if performance collapses, part of the measured 'logic' may be recognition of LLM perturbation patterns rather than general logical understanding.
Going beyond the paper: the structure-aware loss implies text embeddings can be organized by logical relation, so one could probe CLIP-style models by checking whether captions with identical vocabulary but different logical connectives form separated clusters.
Going beyond the paper: the medical results, while improved, remain far below image results, suggesting that everyday-image transfer alone is insufficient for clinical logical reasoning and that domain-specific logical data may be needed.
Going beyond the paper: LogicBench's MCQ format with LLM-generated hard negatives could be adapted to calibrate instruction-tuned multimodal LLMs, not just contrastive retrievers.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. The paper builds a large benchmark (LogicBench) spanning 9 logical categories and 4 scenarios, plus a fine-tuning recipe (LogicCLIP) with a structure-aware classification loss. The benchmark is broader than prior negation/compositionality benchmarks, and the ablations are reasonably thorough. But the central claim — that VLMs are 40+ points below humans and that LogicCLIP closes most of that gap — is not convincingly supported. The evaluation and training data share the same LLM-perturbation distribution, and that confound is strong enough to put the magnitude of the improvement in doubt.

What's good: LogicBench goes beyond CREPE/CC-Neg/NegBench by covering nine categories and four domains, and the authors document the data construction pipeline in detail. The structure-aware objective (multi-hot classification of logical categories) is new and the ablation study shows each loss component matters. They also test LLaVA and include qualitative examples. That's real work.

The main problem is the shared pipeline. In LogicBench MCQs, the correct option is always a human-written caption while the three negatives are LLM rewrites. LogicCLIP is trained on exactly the same distinction: 25% positive human captions, 75% LLM negatives. A model can learn to prefer human phrasing or detect LLM-typical rewrite patterns without doing any image-grounded logical reasoning. The paper does not check for this — no style-matched negatives, no analysis of text-only heuristics. The transfer to video, anomaly, and medical scenarios is particularly unconvincing, because those test sets use the same instruction templates and the same five LLMs as the training data. The general COCO/Flickr gains are mostly explained by fine-tuning on MSCOCO captions; NegCLIP, which also fine-tunes on MSCOCO with hard negatives, gets nearly the same COCO R@1 (41.52 vs 42.54). The human evaluation is under-specified (10% sample, no rater count or variance) and there's a garbled artifact in Section V. The 'data, code, and models will be publicly released' promise is not kept — no link anywhere.

These are fixable, but they're not minor. The headline comparison against human performance and the 80–90% post-training accuracies likely overstate the model's true logical understanding. If the authors could add a control condition where negatives are human-written logically incorrect captions (or at least a different LLM family/template for test), the benchmark would be much stronger.

Who should read it: someone working on compositional reasoning benchmarks or contrastive fine-tuning. It's a good starting point, but not yet a reliable resource. I'd send it to peer review — the benchmark idea and the confound are worth referee time — but I'd expect major revisions focused on breaking the shared-generation link. For my own work, I wouldn't cite it yet.

Referee Report

4 major / 4 minor

Summary. The paper introduces LogicBench, a benchmark of 51,360 vision-language pairs spanning nine logical categories (conjunction, disjunction, negation, contrast, comparison, condition, causality, temporality, inclusion) and four scenarios (images, videos, anomaly detection, medical), with retrieval and multiple-choice tasks. It reports that CLIP-style VLMs score far below human performance (the paper claims a >40-point gap on hard categories) and proposes LogicCLIP, a contrastive fine-tuning framework that generates LLM-based logical negatives from MSCOCO captions and trains with three objectives: standard CLIP loss, a multiple-choice hard-negative loss, and a logical-structure classification loss. Experiments report large LogicBench gains (e.g., OpenAI CLIP-B Image MCQ 38.56% to 81.91%) while retaining or improving COCO/Flickr30K retrieval.

Significance. If the validity concerns are resolved, LogicBench would be a useful large-scale diagnostic for a real weakness of joint-embedding VLMs, and LogicCLIP would demonstrate a practical mitigation. The paper ships a concrete benchmark, multiple baselines, and a training recipe, and the authors commit to releasing data, code, and models. The main contribution at present is the benchmark; the method is largely a combination of existing hard-negative fine-tuning (NegCLIP, NegFull) with an added text-side logic classifier. The central claims are, however, not yet supported by an evaluation that is independent of the training data-generation process, as detailed below.

major comments (4)

[§V, Experimental Settings; Table 1] The test and training distributions are not independent. LogicBench negatives and LogicCLIP training negatives are generated by the same five LLMs, the same nine logical categories, the same instruction templates, and overlapping source data (MSCOCO/CC12M validation captions for LogicBench; MSCOCO training captions for LogicCLIP). In this setting, the large gains (Table 1: CLIP-B Image MCQ 38.56→81.91, Video 35.57→77.92) may reflect learning the LLM perturbation style or the original-vs-LLM-rewrite distinction rather than image-grounded logical understanding. Please add control experiments: (a) evaluate LogicCLIP on LogicBench-style negatives produced by a different protocol (e.g., human-authored logical perturbations or template perturbations from different source sentences); (b) train LogicCLIP with negatives generated by different LLMs or templates; (c) report the accuracy of a text-o
[§V, Experimental Settings; Table 1] The human evaluation is described only as 'randomly selecting 10% of the samples' with no rater count, adjudication protocol, or inter-rater agreement. Because the headline result is a comparison to human performance (e.g., 96.32% Image MCQ), the human reference must be reproducible. Please report the number of raters, per-item labeling, agreement statistics, and whether the 10% sample is stratified by scenario and logical category. Also clarify how ties or ambiguous negatives were handled.
[§V, Table 1 vs. §IV training data] The claimed general-benchmark gains are confounded by fine-tuning on the evaluation distribution. LogicCLIP is fine-tuned on MSCOCO training captions, and COCO test retrieval (Table 1: R@1 30.44→42.54 for CLIP-B) improves even with the standard CLIP loss alone (Variant 3 in Supp. Table 2: 35.75 R@1). A matched control is needed: fine-tune the base model on the same MSCOCO data with non-logical hard negatives, or with standard caption fine-tuning, to isolate the effect of the logic-aware losses. As written, the statement 'Logic-aware training also improves performance on general benchmarks' is not supported beyond in-distribution fine-tuning effects.
[§III Data Construction Pipeline; Supp. A.2] The benchmark construction may introduce surface-level cues. Positive captions are selected by spaCy/regex keyword filters ('and', 'or', 'not', 'because', etc.), while all negatives are LLM rewrites. In every MCQ, the positive is the original human caption and all negatives are generated by the same five LLMs. This asymmetry alone can allow a model to detect the original/LLM boundary without evaluating the image. The paper's human review ensures semantic plausibility but does not remove the source-style confound. Please quantify this: report word-frequency and perplexity differences between positive and negative options, and provide a no-image control (e.g., an LLM or text classifier that picks the positive option). If a text-only model performs above chance, LogicBench is not yet a clean measurement of visual logical understanding.

minor comments (4)

[Table 1] Header contains typo 'Anomoly' (should be 'Anomaly').
[Section III] Cross-references to figures are inconsistent: the text refers to 'Figure 7' for the construction pipeline and 'Figure 6(a)' for categories, but the provided manuscript does not match these numbers. Please renumber figures and ensure all cross-references in the main text point to the correct figures.
[Supp. B.1] The training data is 25% positive and 75% negative captions, and the objective weights are α=4, β=2, γ=1. No sensitivity analysis is reported for the positive/negative ratio or for α, β, γ. At least a small grid or ablation over these choices would help establish robustness.
[References] The DADA-2000 reference contains garbled text ('asƒ analyzed'); please correct. Also the phrase 'abnormal video' in Figure 8 should be 'abnormal video' for grammar.

Circularity Check

1 steps flagged · score 6.0 of 10

LogicBench and LogicCLIP share the same LLM perturbation pipeline; LogicCLIP's MCQ gains largely reflect fitting the benchmark's own negative-generation distribution.

fitted input called prediction [Section IV (Logic-Aware Training Data Generation, LMC objective) and Section III / Supp. A.2 (LogicBench construction and instruction templates)]
"LLMs perturb c to generate a set of three negative captions, denoted as Cneg = {cneg,1, cneg,2, cneg,3}. Each caption cneg,m ∈ Cneg is semantically plausible but logically incorrect or misleading for image I ... The task of LLMs is to create three logically incorrect captions by perturbing the given logical structure."

The LMC objective is a 4-way cross-entropy that selects cpos among {cpos} ∪ Cneg. LogicBench MCQ is the same 4-option task: one original human caption and three LLM-generated logical perturbations. Training positives are MSCOCO captions filtered by the same spaCy/regex rules, and training negatives are generated by the same five LLMs (Qwen 2.5-max, DeepSeek-V3, Gemini-2.5-pro, GPT-4.1, LLaMA 3.3 70B) used for LogicBench. Thus the 'prediction' on LogicBench-Image is the training objective applied to samples from the same distribution; the large gains (38.56→81.91) are expected from fitting the human-vs-LLM style distinction, not from acquiring image-grounded logical structure. Video, anomaly, and medical scenarios use the same instruction template and LLM set, so text-side artifacts transfe

full rationale

No load-bearing self-citation chain or imported uniqueness theorem is present; the paper's core issue is benchmark/training coupling. LogicCLIP is trained on exactly the positive/negative distinction that LogicBench uses to measure 'logical understanding': original human-written captions versus LLM-perturbed hard negatives produced with the same instruction templates and same LLMs. Consequently, LogicCLIP's headline MCQ improvements on LogicBench are partly a measure of in-distribution discrimination rather than independent logical generalization. The general COCO/Flickr retrieval results and the medical/video source variation provide some external checks, which is why this is partial circularity (6) rather than a completely forced derivation (8–10).

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical or theoretical entities are introduced. The free parameters are the hand-chosen loss weights, the 25/75 positive/negative data ratio, and the three-negatives-per-positive design. The domain assumptions concern the validity of the automated filtering, LLM-generated negatives, human review, and cross-domain transfer of logical structure.

free parameters (3)

Loss weights alpha, beta, gamma = 4, 2, 1
Chosen by hand to balance alignment, multiple-choice, and logic classification; not swept or justified in the paper.
Positive/negative caption ratio in LogicCLIP training data = 25% positive, 75% negative
Design choice for the 475,624-sample training set; not ablated except via loss variants.
Number of negative captions per positive sample = 3
Each positive gets three LLM-generated hard negatives in training; MCQ evaluation uses 3 negatives (4 options) or 4 negatives (5 options in medicine).

assumptions (5)

domain assumption SpaCy plus regular expressions can reliably identify captions containing specific logical structures
Used to filter positive samples from CC12M/MSCOCO/MSRVTT/DADA-2000/Open-i; if the filters are noisy, positive samples may be mislabeled. Section III Data Construction Pipeline.
domain assumption LLM-generated perturbed captions are fluent, plausible, and logically incorrect
Negative sample quality is the entire supervisory signal for LogicCLIP and the benchmark discrimination task. Section IV and Appendix A.2.
domain assumption Human expert review guarantees logical consistency and correctness of final benchmark labels
Human review is mentioned but no protocol, inter-annotator agreement, or rater counts are given. Section III Data Construction Pipeline.
domain assumption Logical structures learned from daily images transfer to video, anomaly, and medical domains
LogicCLIP is trained only on MSCOCO daily images yet evaluated on LogicBench-Video/Anomaly/Medicine; the paper treats logical structure as domain-invariant. Section V Result 4.
standard math Cross-entropy and contrastive losses are the correct optimization primitives
Standard ML background; not in question.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models." pith.science (2026). https://pith.science/paper/6P2WH2BW

@misc{pith2026250811317,
  author       = {Pith},
  title        = {Pith review of: Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6P2WH2BW}},
  note         = {Machine review of arXiv:2508.11317}
}

read the original abstract

Vision-Language Models (VLMs), exemplified by CLIP, have emerged as foundational for multimodal intelligence. However, their capacity for logical understanding remains significantly underexplored, resulting in critical ''logical blindspots'' that limit their reliability in practical applications. To systematically diagnose this, we introduce LogicBench, a comprehensive benchmark with over 50,000 vision-language pairs across 9 logical categories and 4 diverse scenarios: images, videos, anomaly detection, and medical diagnostics. Our evaluation reveals that existing VLMs, even the state-of-the-art ones, fall at over 40 accuracy points below human performance, particularly in challenging tasks like Causality and Conditionality, highlighting their reliance on surface semantics over critical logical structures. To bridge this gap, we propose LogicCLIP, a novel training framework designed to boost VLMs' logical sensitivity through advancements in both data generation and optimization objectives. LogicCLIP utilizes logic-aware data generation and a contrastive learning strategy that combines coarse-grained alignment, a fine-grained multiple-choice objective, and a novel logical structure-aware objective. Extensive experiments demonstrate LogicCLIP's substantial improvements in logical comprehension across all LogicBench domains, significantly outperforming baselines. Moreover, LogicCLIP retains, and often surpasses, competitive performance on general vision-language benchmarks, demonstrating that the enhanced logical understanding does not come at the expense of general alignment. We believe that LogicBench and LogicCLIP will be important resources for advancing VLM logical capabilities.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models
cs.CV 2026-07 conditional novelty 6.0 of 10

Dual-encoder VLMs behave like bag-of-concepts scorers; a training-free score edit (complement for negation, min/max for AND/OR) restores truth-functional constraint execution on frozen encoders.
BalDRO: A Distributionally Robust Optimization based Framework for Large Language Model Unlearning
cs.LG 2026-01 conditional novelty 5.0 of 10

BalDRO makes LLM unlearning more balanced by updating against a worst-case-weighted forget distribution, improving forget quality on TOFU/MUSE at stable utility.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages · cited by 2 Pith papers

[3]

Figure 6: Instruction Templates for LogicBench-Image

<hard negative caption> Placeholder Explanation: {logic type} refers to one of the following logical structures: conjunction, disjunction, nega- tion, contrast, comparison, condition, causality, temporality, and inclusion. Figure 6: Instruction Templates for LogicBench-Image. Instruction Templates for the LogicBench-Video You are given a caption that desc...

work page
[4]

and”, “both

Iterated Learning Improves Compositionality in Large Vision-Language Models. In CVPR, 13785–13795. Zhou, Y .; Liu, L.; and Gou, C. 2024. Learning from observer gaze: Zero-shot attention prediction oriented by human- object interaction recognition. In CVPR, 28390–28400. Zhou, Y .; Tan, G.; and Gou, C. 2024. Hierarchical home action understanding with impli...

work page 2024
[5]

The family enjoyed the colorful frisbee, tasty sand- wiches, a comfortable blanket, and playing with their lively dog

<Hard negative caption> Figure 9: Instruction Templates for LogicBench-Medicine. • TripletCLIP-CC12M (Patel et al. 2024): This vari- ant is accessible via https://huggingface.co/TripletCLIP/ CC12M TripletCLIP ViTB12. • MetaCLIP (Xu et al. 2024): The model and code are available at https://github.com/facebookresearch/ MetaCLIP. • ConCLIP (Singh et al. 2025...

work page 2024
[10]

Figure 7: Instruction Templates for LogicBench-Video

<hard negative caption> Placeholder Explanation: {logic type} refers to one of the following logical structures: conjunction, disjunction, nega- tion, contrast, comparison, condition, causality, temporality, and inclusion. Figure 7: Instruction Templates for LogicBench-Video. Instruction Templates for the LogicBench-Anomaly You are given a caption that de...

work page
[12]

<hard negative caption>

work page
[13]

Figure 8: Instruction Templates for LogicBench-Anomaly

<hard negative caption> Placeholder Explanation: {logic type} refers to one of the following logical structures: conjunction, disjunction, nega- tion, contrast, comparison, condition, causality, temporality, and inclusion. Figure 8: Instruction Templates for LogicBench-Anomaly. Instruction Templates for the LogicBench-Medicine You are given a caption that...

work page
[14]

no pleural effusion

Extract positive findings (statements without negation) and negative findings (statements with negation, e.g., “no pleural effusion”)

work page
[15]

Because <positive finding> and <negative finding>, the impression is <IMPRESSION>

Construct ONE correct option using this rule-based format: “Because <positive finding> and <negative finding>, the impression is <IMPRESSION>.”

work page

Show all 14 references

[16]

no effusion

Generate FOUR hard negative options using logical perturbations: - Negation Flip: Reverse negation (e.g., “no effusion” → “effusion present”). - Conjunction Trap: Combine correct and incorrect statements with “and”. - Disjunction Confusion: Use “or” with one correct and one in...
[17]

Only the first option should be correct

Return exactly five options, labeled 1 to 5. Only the first option should be correct. Please use the following exact format:
[20]

<Hard negative caption>
[2019]

In ITSC, 4303– 4309

Dada-2000: Can driving accident be predicted by driver attentionƒ analyzed by a benchmark. In ITSC, 4303– 4309. Gadre, S. Y .; Ilharco, G.; Fang, A.; Hayase, J.; Smyrnis, G.; Nguyen, T.; Marten, R.; Wortsman, M.; Ghosh, D.; Zhang, J.; et al. 2023. Datacomp: In search of the ne...

2000 arXiv
[2023]

In NeurIPS, 35544–35575

Improving clip training with language rewrites. In NeurIPS, 35544–35575. Fang, J.; Yan, D.; Qiao, J.; Xue, J.; Wang, H.; and Li, S
[2024]

In CVPR, 9568–9578

Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 9568–9578. Wang, J.; Chan, K. C.; and Loy, C. C. 2023. Exploring clip for assessing the look and feel of images. In AAAI, vol- ume 37, 2555–2563. Wang, Z.; Zhou, S.; He, S.; Huang, H.; Yang, L.; Zha...

2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [3]

Figure 6: Instruction Templates for LogicBench-Image

<hard negative caption> Placeholder Explanation: {logic type} refers to one of the following logical structures: conjunction, disjunction, nega- tion, contrast, comparison, condition, causality, temporality, and inclusion. Figure 6: Instruction Templates for LogicBench-Image. Instruction Templates for the LogicBench-Video You are given a caption that desc...

work page

[2] [4]

and”, “both

Iterated Learning Improves Compositionality in Large Vision-Language Models. In CVPR, 13785–13795. Zhou, Y .; Liu, L.; and Gou, C. 2024. Learning from observer gaze: Zero-shot attention prediction oriented by human- object interaction recognition. In CVPR, 28390–28400. Zhou, Y .; Tan, G.; and Gou, C. 2024. Hierarchical home action understanding with impli...

work page 2024

[3] [5]

The family enjoyed the colorful frisbee, tasty sand- wiches, a comfortable blanket, and playing with their lively dog

<Hard negative caption> Figure 9: Instruction Templates for LogicBench-Medicine. • TripletCLIP-CC12M (Patel et al. 2024): This vari- ant is accessible via https://huggingface.co/TripletCLIP/ CC12M TripletCLIP ViTB12. • MetaCLIP (Xu et al. 2024): The model and code are available at https://github.com/facebookresearch/ MetaCLIP. • ConCLIP (Singh et al. 2025...

work page 2024

[4] [10]

Figure 7: Instruction Templates for LogicBench-Video

<hard negative caption> Placeholder Explanation: {logic type} refers to one of the following logical structures: conjunction, disjunction, nega- tion, contrast, comparison, condition, causality, temporality, and inclusion. Figure 7: Instruction Templates for LogicBench-Video. Instruction Templates for the LogicBench-Anomaly You are given a caption that de...

work page

[5] [12]

<hard negative caption>

work page

[6] [13]

Figure 8: Instruction Templates for LogicBench-Anomaly

<hard negative caption> Placeholder Explanation: {logic type} refers to one of the following logical structures: conjunction, disjunction, nega- tion, contrast, comparison, condition, causality, temporality, and inclusion. Figure 8: Instruction Templates for LogicBench-Anomaly. Instruction Templates for the LogicBench-Medicine You are given a caption that...

work page

[7] [14]

no pleural effusion

Extract positive findings (statements without negation) and negative findings (statements with negation, e.g., “no pleural effusion”)

work page

[8] [15]

Because <positive finding> and <negative finding>, the impression is <IMPRESSION>

Construct ONE correct option using this rule-based format: “Because <positive finding> and <negative finding>, the impression is <IMPRESSION>.”

work page