Query-adaptive Recurrent Tokenization
We introduce the Query-Adaptive Recurrent Tokenizer (QART), a backbone-agnostic module that makes visual representations both query-steerable and budget-adaptive.
ProofPaper ↗
Key points
- Visual information worth extracting from an image and the token budget it warrants depend on the text query.
- Yet modern vision and vision-language models largely encode images independently of the query and allocate a fixed visual token budget to every input.
- QART steers the visual representation toward query-relevant evidence while adapting its token budget to the amount of information required.
- Starting from frozen encoder features, QART recurrently constructs a compact representation and uses reconstruction quality over the query-relevant visual support to control token allocation.
Sources (1)
- [1]Query-adaptive Recurrent TokenizationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 05:28 AM
We introduce the Query-Adaptive Recurrent Tokenizer (QART), a backbone-agnostic module that makes visual representations both query-steerable and budget-adaptive.
Visual information worth extracting from an image and the token budget it warrants depend on the text query.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026DICL-Qwen3VL: Decoupled Confidence Calibration for Interpretable Workwear Detection
- Oct 16, 2026Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design
- Oct 16, 2026HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
- Oct 15, 2026Each Reward to Its Relevant Tokens: Reward-Specific Token-Level Credit Assignment for Multi-Reward RL
- Oct 15, 2026Learning to Think and Explore in 3D World
- Oct 15, 2026CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness