# The Hard Positive Truth about Vision-Language Compositionality

arXiv: https://arxiv.org/abs/2409.17958 · HTML: https://mlsanity.com/paper/2409.17958/

## Key facts

- Cited by 22 papers in the map (data through 14 Oct 2026).
- Placed in Prompt learning and adaptation of CLIP, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 40 of its 52 references that are in the map.
- First posted on arXiv 26 Sep 2024.

## Placed in

- Multimodal LLMs / vision-language models / [Prompt learning and adaptation of CLIP](https://mlsanity.com/mllm/prompt-learning-and-adaptation-of-clip/)

## Builds on (most cited first, 30 of 40)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Segment Anything](https://mlsanity.com/paper/2304.02643/) (2023)
- [Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding](https://mlsanity.com/paper/2205.11487/) (2022)
- [Efficient Estimation of Word Representations in Vector Space](https://mlsanity.com/paper/1301.3781/) (2013)
- [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://mlsanity.com/paper/2201.12086/) (2022)
- [LAION-5B: An open large-scale dataset for training next generation image-text models](https://mlsanity.com/paper/2210.08402/) (2022)
- [Distributed Representations of Words and Phrases and their Compositionality](https://mlsanity.com/paper/1310.4546/) (2013)
- [Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations](https://mlsanity.com/paper/1602.07332/) (2016)
- [CLIPScore: A Reference-free Evaluation Metric for Image Captioning](https://mlsanity.com/paper/2104.08718/) (2021)
- [BLOOM: A 176B-Parameter Open-Access Multilingual Language Model](https://mlsanity.com/paper/2211.05100/) (2022)
- [Imagen Video: High Definition Video Generation with Diffusion Models](https://mlsanity.com/paper/2210.02303/) (2022)
- [Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time](https://mlsanity.com/paper/2203.05482/) (2022)
- [Annotation Artifacts in Natural Language Inference Data](https://mlsanity.com/paper/1803.02324/) (2018)
- [FLAVA: A Foundational Language And Vision Alignment Model](https://mlsanity.com/paper/2112.04482/) (2021)
- [DataComp: In search of the next generation of multimodal datasets](https://mlsanity.com/paper/2304.14108/) (2023)
- [When and why vision-language models behave like bags-of-words, and what to do about it?](https://mlsanity.com/paper/2210.01936/) (2022)
- [LiT: Zero-Shot Transfer with Locked-image text Tuning](https://mlsanity.com/paper/2111.07991/) (2021)
- [Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality](https://mlsanity.com/paper/2204.03162/) (2022)
- [A Corpus for Reasoning About Natural Language Grounded in Photographs](https://mlsanity.com/paper/1811.00491/) (2018)
- [TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering](https://mlsanity.com/paper/2303.11897/) (2023)
- [A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark](https://mlsanity.com/paper/1910.04867/) (2019)
- [Visual Relationship Detection with Language Priors](https://mlsanity.com/paper/1608.00187/) (2016)
- [SWAG: A Large-Scale Adversarial Dataset for Grounded Commonsense Inference](https://mlsanity.com/paper/1808.05326/) (2018)
- [Compositionality decomposed: how do neural networks generalise?](https://mlsanity.com/paper/1908.08351/) (2019)
- [Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts](https://mlsanity.com/paper/2111.08276/) (2021)
- [What's "up" with vision-language models? Investigating their struggle with spatial reasoning](https://mlsanity.com/paper/2310.19785/) (2023)
- [Action Genome: Actions as Composition of Spatio-temporal Scene Graphs](https://mlsanity.com/paper/1912.06992/) (2019)
- [SugarCrepe: Fixing Hackable Benchmarks for Vision-Language Compositionality](https://mlsanity.com/paper/2306.14610/) (2023)
- [Improving CLIP Training with Language Rewrites](https://mlsanity.com/paper/2305.20088/) (2023)
- [CREPE: Can Vision-Language Foundation Models Reason Compositionally?](https://mlsanity.com/paper/2212.07796/) (2022)

## Built on by (most cited first, 22 of 22)

- [CLIP Behaves like a Bag-of-Words Model Cross-modally but not Uni-modally](https://mlsanity.com/paper/2502.03566/) (2025)
- [Decoupled Global-Local Alignment for Improving Compositional Understanding](https://mlsanity.com/paper/2504.16801/) (2025)
- [Natural Language Inference Improves Compositionality in Vision-Language Models](https://mlsanity.com/paper/2410.22315/) (2024)
- [AQAScore: Evaluating Semantic Alignment in Text-to-Audio Generation via Audio Question Answering](https://mlsanity.com/paper/2601.14728/) (2026)
- [New Dataset and Methods for Fine-Grained Compositional Referring Expression Comprehension via Specialist-MLLM Collaboration](https://mlsanity.com/paper/2502.20104/) (2025)
- [A Good CREPE needs more than just Sugar: Investigating Biases in Compositional Vision-Language Benchmarks](https://mlsanity.com/paper/2506.08227/) (2025)
- [Importance Sampling for Multi-Negative Multimodal Direct Preference Optimization](https://mlsanity.com/paper/2509.25717/) (2025)
- [OpenVoxel: Training-Free Grouping and Captioning Voxels for Open-Vocabulary 3D Scene Understanding](https://mlsanity.com/paper/2601.09575/) (2026)
- [VAQUUM: Are Vague Quantifiers Grounded in Visual Data?](https://mlsanity.com/paper/2502.11874/) (2025)
- [The Artificial Intelligence Cognitive Examination: A Survey on the Evolution of Multimodal Evaluation from Recognition to Reasoning](https://mlsanity.com/paper/2510.04141/) (2025)
- [Global Geometry Is Not Enough for Vision Representations](https://mlsanity.com/paper/2602.03282/) (2026)
- [Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models](https://mlsanity.com/paper/2607.23052/) (2026)
- [Omni-NegCLIP: Enhancing CLIP with Front-Layer Contrastive Fine-Tuning for Comprehensive Negation Understanding](https://mlsanity.com/paper/2603.29258/) (2026)
- [Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality](https://mlsanity.com/paper/2606.13288/) (2026)
- [Explicit Logic Channel for Validation and Enhancement of MLLMs on Zero-Shot Tasks](https://mlsanity.com/paper/2603.11689/) (2026)
- [Concrete Jungle: Towards Concreteness Paved Contrastive Negative Mining for Compositional Understanding](https://mlsanity.com/paper/2604.13313/) (2026)
- [Foveated Probes Recover Localized Binding Information in Vision Foundation Models](https://mlsanity.com/paper/2608.00726/) (2026)
- [PragMatch: Separating Pragmatic Incongruity from Cross-Modal Mismatch in Large Vision-Language Models](https://mlsanity.com/paper/2608.09772/) (2026)
- [RePair: Turning Retrieval Failures into Counterfactual Hard Pairs](https://mlsanity.com/paper/2608.29604/) (2026)
- [BindCLIP: One Balanced Coupling For Compositional Vision Language Scoring](https://mlsanity.com/paper/2609.23717/) (2026)
- [Diagnosing the Sources of Compositional Failure in Vision-Language Models: A Controlled Analysis](https://mlsanity.com/paper/2609.31456/) (2026)
- [The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors](https://mlsanity.com/paper/2610.16977/) (2026)

## Abstract (arXiv)

Several benchmarks have concluded that our best vision-language models (e.g., CLIP) are lacking in compositionality. Given an image, these benchmarks probe a model's ability to identify its associated caption amongst a set of compositional distractors. In response, a surge of recent proposals show improvements by finetuning CLIP with distractors as hard negatives. Our investigations reveal that these improvements have, in fact, been significantly overstated -- because existing benchmarks do not probe whether finetuned vision-language models remain invariant to hard positives. By curating an evaluation dataset with 112,382 hard negatives and hard positives, we uncover that including hard positives decreases CLIP's performance by 12.9%, while humans perform effortlessly at 99%. CLIP finetuned with hard negatives results in an even larger decrease, up to 38.7%. With this finding, we then produce a 1,775,259 image-text training set with both hard negative and hard positive captions. By training with both, we see improvements on existing benchmarks while simultaneously improving performance on hard positives, indicating a more robust improvement in compositionality. Our work suggests the need for future research to rigorously test and improve CLIP's understanding of semantic relationships between related "positive" concepts.