Revisiting Compositionality in Dual-Encoder Vision-Language Models: The Role of Inference

Dual-encoder vision-language models (VLMs) expose a similarity interface that enables zero-shot retrieval but fails compositional constraints: queries like "umbrella and no person" retrieve images containing both, even when concept detection is reliable. We trace this to an interface-level Bag-of-Concepts effect, where similarity scores approximate mean pooling of concept evidence regardless of operators. Although operator-dependent signals exist in text embeddings, they are too weak or misaligned to affect rankings. Fine-tuning does not reliably resolve this failure because the dominant bottleneck is how similarity aggregates evidence rather than what encoders represent. We propose factored inference, which separates evidence extraction from constraint execution, and introduce LCSE (Logic-Constrained Score Editing), a training-free method that executes constraints externally using concept scores from frozen encoders. We also introduce FACTOR-Bench, where LCSE achieves 85.5% accuracy versus 73.2% for the best fine-tuned baseline, 90.7% when applied to SigLIP 2, and improves NegBench COCO MCQ accuracy from 27.2% to 65.2% while preserving retrieval performance.

Recent vision-language models (VLMs) typically rely on a single vision encoder trained with contrastive image-text objectives, such as CLIP-style pretraining. While contrastive encoders are effective for cross-modal alignment and retrieval, self-supervised visual encoders often capture richer dense semantics and exhibit stronger robustness on recognition and understanding tasks. In this work, we investigate how to scale the fusion of these complementary visual representations for vision-language modeling. We propose CoME-VL: Complementary Multi-Encoder Vision-Language, a modular fusion framework that integrates a contrastively trained vision encoder with a self-supervised DINO encoder. Our approach performs representation-level fusion by (i) entropy-guided multi-layer aggregation with orthogonality-constrained projections to reduce redundancy, and (ii) RoPE-enhanced cross-attention to align heterogeneous token grids and produce compact fused visual tokens. The fused tokens can be injected into a decoder-only LLM with minimal changes to standard VLM pipelines. Extensive experiments across diverse vision-language benchmarks demonstrate that CoME-VL consistently outperforms single-encoder baselines. In particular, we observe an average improvement of 4.9% on visual understanding tasks and 5.4% on grounding tasks. Our method achieves state-of-the-art performance on RefCOCO for detection while improving over the baseline by a large margin. Finally, we conduct ablation studies on layer merging, non-redundant feature mixing, and fusion capacity to evaluate how complementary contrastive and self-supervised signals affect VLM performance.

Contrastive vision-language (V&L) models remain a popular choice for various applications. However, several limitations have emerged, most notably the limited ability of V&L models to learn compositional representations. Prior methods often addressed this limitation by generating custom training data to obtain hard negative samples. Hard negatives have been shown to improve performance on compositionality tasks, but are often specific to a single benchmark, do not generalize, and can cause substantial degradation of basic V&L capabilities such as zero-shot or retrieval performance, rendering them impractical. In this work we follow a different approach. We identify two root causes that limit compositionality performance of V&Ls: 1) Long training captions do not require a compositional representation; and 2) The final global pooling in the text and image encoders lead to a complete loss of the necessary information to learn binding in the first place. As a remedy, we propose two simple solutions: 1) We obtain short concept centric caption parts using standard NLP software and align those with the image; and 2) We introduce a parameter-free cross-modal attention-pooling to obtain concept centric visual embeddings from the image encoder. With these two changes and simple auxiliary contrastive losses, we obtain SOTA performance on standard compositionality benchmarks, while maintaining or improving strong zero-shot and retrieval capabilities. This is achieved without increasing inference cost. We release the code for this work at https://github.com/saic-fi/concept_centric_clip.