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

2026-05-28 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision, Natural Language Processing · Depth: Expert, quick

Summary

PARCEL (Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding) is a new visual tokenization architecture designed to address the quadratic computational bottleneck in Large Vision-Language Models (LVLMs) caused by dense token sequences. Existing elastic visual-token compression methods, like spatial-only or query-only compression, struggle under aggressive compression, leading to issues like spectral aliasing or degraded spatial grounding. PARCEL resolves this by dynamically partitioning feature extraction, establishing spatial pool tokens as low-frequency layout anchors, and conditioning elastic query tokens on these anchors through Pool-Conditioned Query Resampling. This approach encourages query tokens to focus on complementary visual features. Evaluated across 27 benchmarks, PARCEL consistently outperforms existing matryoshka baselines, improving the performance-efficiency Pareto frontier while maintaining a "train once, deploy anywhere" paradigm.

Key takeaway

For Machine Learning Engineers optimizing Large Vision-Language Model inference, PARCEL offers a superior approach to visual token compression. You can achieve better performance-efficiency trade-offs across various visual-token budgets compared to existing methods. Consider integrating PARCEL to deploy a single model that adapts efficiently to different computational constraints, preserving fine-grained visual detail without sacrificing spatial grounding. This improves model utility and reduces operational complexity.

Key insights

PARCEL dynamically partitions visual feature extraction using anchored resampling to overcome LVLM compression limitations.

Principles

• Spatial-only compression causes spectral aliasing.
• Query-only compression degrades spatial grounding.
• Anchor tokens can guide elastic queries.

Method

PARCEL establishes spatial pool tokens as low-frequency layout anchors. It then conditions elastic query tokens on these anchors via Pool-Conditioned Query Resampling, directing queries to complementary visual features.

In practice

• Improve LVLM efficiency with PARCEL.
• Deploy a single model for multiple budgets.
• Enhance fine-grained detail in compressed vision.

Topics

• Vision-Language Models
• Visual Tokenization
• Model Compression
• Efficient Inference
• PARCEL
• Performance-Efficiency

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• OccamToken: Efficient VLM Inference with Training-Free and Budget-Adaptive Token Pruning — OccamToken uses register-anchored relative evidence testing for training-free, budget-adaptive VLM token pruning, improving efficiency.
• Elastic Attention Cores for Scalable Vision Transformers [R] — Elastic Attention Cores offer a scalable, efficient alternative to dense self-attention in Vision Transformers.
• UHR-BAT: Budget-Aware Token Compression Vision-Language model for Ultra-High-Resolution Remote Sensing — UHR-BAT efficiently compresses visual tokens in ultra-high-resolution remote sensing imagery using query-guided, region-faithful selection.
• ST-SimDiff: Balancing Spatiotemporal Similarity and Difference for Efficient Video Understanding with MLLMs — Efficient video understanding balances compressing redundant content with preserving dynamic, event-driven changes.
• KVCapsule: Efficient Sequential KV Cache Compression for Vision-Language Models with Asymmetric Redundancy — KVCapsule efficiently compresses VLM KV caches by leveraging asymmetric redundancy and dynamic attention patterns in vision tokens.
• PERL: Parameter Efficient Reasoning in CLIP Latent Space — Iterative latent reasoning offers a parameter-efficient adaptation mechanism for discriminative vision-language models like CLIP.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.