COMPOSITIONALITY IN CONTRASTIVE VISION-LANGUAGE MODELS: A SURVEY OF METHODS AND BENCHMARKS

Contrastive vision language models (CVLMs), such as CLIP, have achieved impressive results in zeroshot tasks. Despite their progress, CVLMs continue to struggle with compositional reasoning-the ability to correctly interpret novel combinations of familiar concepts. In vision-language learning, compositionality involves several core abilities: linking attributes to objects, reasoning about spatial relations between entities, reasoning about temporal or event relations, understanding logical operators such as negation and quantifiers, and handling role-sensitive interpretation. Research has shown that a common failure mode in CLIP is a bag-of-words behavior, where images and captions are treated as unordered sets of concepts. This undermines attribute-object binding, relational reasoning, and sensitivity to word order or role assignment. More broadly, these models often exploit shortcut signals, such as word co-occurrence statistics or dataset biases, rather than learning systematic, compositional rules that generalize to unseen combinations. This survey reviews recent efforts to strengthen compositionality in CLIP-style vision-language models. We propose a taxonomy with four primary method families and a complementary line of work on diagnostic probes. Data-centric approaches reshape the training signal via rule-based edits, synthetic data generation, graph-based construction, and targeted sampling to expose models to diverse compositional structures. Objective-centric approaches revise learning criteria so that optimization explicitly rewards correct compositional bindings rather than shortcut cues. Architecture-centric approaches modify model architectures to enable finer-grained visual grounding and similarity refinement. Inference-time approaches that are training-free, enhance frozen models by decomposing captions, grounding parts in visual regions, and fusing local signals into the global image-text scores to enforce compositional constraints. Finally, diagnostic analysis identifies where and why compositional failures occur, attributing errors to specific components and evaluating the reliability of existing benchmarks. Our goal is to provide a clear and organized overview of this rapidly growing field. The taxonomy offers newcomers an accessible entry point while serving as a reference for experts; we highlight key open challenges alongside promising directions for future research.