PARCEL: Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding

Researchers introduced PARCEL, a vision-language tokenization method that combines pool tokens and conditioned query tokens for elastic visual-token compression.

By using pool tokens as spatial anchors, PARCEL aims to cut compute while preserving more spatial detail than pooling-only methods and better grounding than query-only methods.

Across 27 benchmarks, PARCEL outperformed prior matryoshka-style baselines over a wide range of token budgets.

The result improves the accuracy-compute tradeoff for vision-language models that need to operate under changing image-token budgets.