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’. |
Similar Papers
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. |
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 . |
Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality (2026.acl-long)
Copied to clipboard
| Challenge: | a contrastive learning approach for vision-language models is needed to capture compositional information. |
| Approach: | They propose a framework that masks compositional concepts in one modality and reconstructs them conditioned on full contextual information from the other . |
| Outcome: | The proposed framework enhances compositionality in visual language models and improves their ability to capture syntactic structure and linguistic information. |
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. |
Do Neural Language Models Inferentially Compose Concepts the Way Humans Can? (2024.lrec-main)
Copied to clipboard
| Challenge: | a new study shows that language models and humans may rely on different approaches to represent and compose lexical items across sentence structure. |
| Approach: | They propose to use a dataset to test the performance of neural language models and humans on inferentially driven conceptual compositions. |
| Outcome: | The proposed model elicits probability estimates for a noun in a minimally composed phrase . RoBERTa, BERT-large, and GPT-2 exhibited the closest resemblance to human responses . |
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. |
Combine to Describe: Evaluating Compositional Generalization in Image Captioning (2022.acl-srw)
Copied to clipboard
| Challenge: | Recent work on compositionality has focused on the ability to combine simpler concepts to understand & generate arbitrarily more complex conceptual structures. |
| Approach: | They propose to use a set of image captioning models to benchmark their compositional generalization properties. |
| Outcome: | The proposed models do not generalize in terms of systematicity and productivity, but are robust to synonym substitutions. |
Are representations built from the ground up? An empirical examination of local composition in language models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Compositionality is a hallmark of human language, but many phrases are non-compositional . a study by a team of researchers shows that LMs may not be able to distinguish between compositional and non-composable phrases. |
| Approach: | They propose to predict LM-internal representations of longer phrases given their constituents . they find that the representation of a parent phrase can be predicted with some accuracy . |
| Outcome: | The proposed model can predict a parent phrase with some accuracy given its children's transformations, but this is not the case. |
Quantifying Compositionality of Classic and State-of-the-Art Embeddings (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Static word embeddings make strong claims about compositionality, but the SOTA generative models go too far in the other direction. |
| Approach: | a new study evaluates the compositionality of word embeddings by canonical correlation analysis . strong compositional signals are observed in later training stages across data modalities . |
| Outcome: | a new evaluation of compositional models shows that they exploit access meanings when justified . strong compositional signals are observed in later training stages and in deeper layers of the transformer-based model before a decline at the top layer. |