Text encoders bottleneck compositionality in contrastive vision-language models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing multimodal models are often unable to reason about simple spatial relations or attribute attachments. |
| Approach: | They first curate CompPrompts, a set of increasingly compositional image captions that VL models should be able to capture . then train text-only recovery probes that aim to reconstruct captions from single-vector text representations produced by several VL model. |
| Outcome: | The proposed model can reconstruct captions from single-vector text representations produced by several models on a broader range of scenes compared to previous models. |
Similar Papers
Vision-Free Retrieval: Rethinking Multimodal Search with Textual Scene Descriptions (2025.emnlp-main)
Copied to clipboard
| Challenge: | Contrastively trained Vision-Language Models exhibit shallow language understanding, manifesting bag-of-words behaviour. |
| Approach: | They propose a vision-free, single-encoder retrieval pipeline to replace traditional text-to-image retrieval paradigm with structured image descriptions. |
| Outcome: | The proposed approach reduces the modality gap and improves compositionality and performance on short and long caption queries. |
Long Story Short: Disentangling Compositionality and Long-Caption Understanding in Contrastive VLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for vision-language models treat compositionality and long-caption understanding in isolation. |
| Approach: | They analyze when compositional reasoning and long-caption understanding transfer across tasks and when this relationship fails. |
| Outcome: | The proposed model can generalize on poorly grounded captions and with strong visual grounding, while architectural choices can limit compositional learning. |
ComCLIP: Training-Free Compositional Image and Text Matching (2024.naacl-long)
Copied to clipboard
| Challenge: | erroneous semantics of individual entities are essentially confounders that cause the matching failure. |
| Approach: | They propose a training-free compositional CLIP model which disentangles input images into subjects, objects, and action subimages and composes CLIP’s vision encoder and text encoder to perform evolving matching over compositional text embedding and subimage embeddments. |
| Outcome: | The proposed model mitigates spurious correlations introduced by the pretrained CLIP models and dynamically evaluates the importance of each component. |
Exploring How Generative MLLMs Perceive More Than CLIP with the Same Vision Encoder (2025.acl-long)
Copied to clipboard
| Challenge: | Recent studies show that CLIP models struggle with visual reasoning tasks . despite the success of Contrastive Language-Image Pretraining, there are still limitations . |
| Approach: | They propose to use a visual encoder to train CLIP-like models for fine-grained visual reasoning tasks. |
| Outcome: | The proposed models outperform CLIP-like encoders in visual reasoning tasks . the study highlights the importance of VLM architectural choices . |
Does CLIP Bind Concepts? Probing Compositionality in Large Image Models (2024.findings-eacl)
Copied to clipboard
| Challenge: | Large-scale neural network models combining text and images have made incredible progress in recent years, but to what extent they encode compositional representations of the concepts over which they operate remains an open question . |
| Approach: | They compare the performance of a large pretrained vision and language model (CLIP) to a set of three synthetic datasets designed to test concept binding. |
| Outcome: | The proposed model can encode compositional concepts and bind variables in a structure-sensitive way, e.g., differentiating ‘cube behind sphere’ from ‘cub behind cube’. |
Preserving Multi-Modal Capabilities of Pre-trained VLMs for Improving Vision-Linguistic Compositionality (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing fine-tuning approaches for compositional understanding compromise performance in zero-shot multi-modal tasks. |
| Approach: | They propose a method to enhance compositional understanding in pre-trained vision and language models without sacrificing performance in zero-shot multi-modal tasks. |
| Outcome: | The proposed method achieves compositionality on par with state-of-the-art models and retains strong multi-modal capabilities. |
On the Language Encoder of Contrastive Cross-modal Models (2024.findings-acl)
Copied to clipboard
Mengjie Zhao, Junya Ono, Zhi Zhong, Chieh-Hsin Lai, Yuhta Takida, Naoki Murata, Wei-Hsiang Liao, Takashi Shibuya, Hiromi Wakaki, Yuki Mitsufuji
| Challenge: | Pretrained audio-language models such as AudioCLIP and AudioCLAP have shown promising results on vision-language (VL) tasks. |
| Approach: | They extensively evaluate how unsupervised and supervised sentence embedding training affect language encoder quality and cross-modal task performance. |
| Outcome: | The proposed model improves on visual-language (VL) and audio-language tasks when the amount of training data is large. |
Do Vision-Language Models Understand Compound Nouns? (2024.naacl-short)
Copied to clipboard
| Challenge: | Open-vocabulary vision-language models (CLIP) are emerging as a promising new paradigm for text-to-image retrieval. |
| Approach: | They propose a benchmark to evaluate the effectiveness of open-vocabulary vision-language models (CLIP) for text-to-image retrieval using contrastive loss. |
| Outcome: | The proposed framework improves CN understanding of CLIP by 8.25% on Compun. |
Seeing Through Words, Speaking Through Pixels: Deep Representational Alignment Between Vision and Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies show that deep vision-only and language-only models project inputs into a partially aligned representational space. |
| Approach: | They investigate whether a model's representational code is semantically shared . they find that alignment peaks in mid-to-late layers of both model types . |
| Outcome: | a forced-choice "Pick-a-Pic" task shows human preferences for image-caption matches are mirrored in embedding spaces across vision-language model pairs. |
Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates (2025.acl-long)
Copied to clipboard
| Challenge: | Recent advances in multimodal systems have demonstrated remarkable capabilities in generating multimodal content from multimodal inputs. |
| Approach: | They propose a benchmark that leverages large language models to generate deceptive text samples to exploit compositional vulnerabilities across different modalities. |
| Outcome: | The proposed approach exploits compositional vulnerabilities across images, videos, and audios. |