Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval
arXiv:2607.04605v1 Announce Type: cross Abstract: Multi-vector vision-language retrieval preserves fine-grained visual evidence through maximum-similarity late interaction, but dense image-side tokens make storage and scoring expensive. Existing token compression methods reduce this cost, yet they can remove or collapse object- and region-level evidence that future query tokens may need to select. We propose SaMer, an object-aware token merging framework that compresses image-side post-projector tokens into $K$ representative centroids while preserving the original late-interaction interface.
The proliferation of multi-vector vision-language models necessitates more efficient token processing, making innovations in compression crucial for scaling these systems effectively.
Efficient visual token handling directly impacts the cost and performance of advanced AI models, determining their commercial viability and wider adoption in various applications.
This research introduces an object-aware token merging framework that reduces computational costs for vision-language retrieval without sacrificing fine-grained visual evidence, improving scalability.
- · AI developers
- · Cloud computing providers
- · Companies using vision-language AI
- · Researchers in computer vision
- · Inefficient vision-language models
- · High-cost data storage solutions
Reduced computational overhead for vision-language AI enhances model deployment and accessibility.
More efficient and cost-effective AI systems accelerate the development of complex AI agents and augmented reality applications.
Improved scalability of vision-language AI could enable new forms of human-computer interaction and intelligence amplification across industries.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI