ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Hugging Face Daily Papers Papers

Summary

ReToken introduces a single learnable retrieval token that selects sparse query-relevant visual tokens from a pre-filled visual KV cache, improving long-context vision-language model performance on visual retrieval tasks. It yields consistent gains across image and video benchmarks while enabling efficient long-video inference on a single H100.

Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. We present ReToken, a single learnable embedding trained as an explicit retrieval target that selects a sparse set of query-relevant visual tokens from a pre-filled visual KV cache. Trained on only a small image-QA dataset, ReToken yields consistent gains across image and video benchmarks: on Visual Haystacks it improves Qwen3VL-8B by 13.4 points and InternVL3.5 by 12.4 points (>20% relative), and on LVBench it transfers zero-shot to long video for an 8.0-point gain with Qwen3VL-8B. Thanks to its lightweight design, both training and long-video inference fit on a single H100. Code is available at: https://github.com/avaxiao/ReToken
Original Article
View Cached Full Text

Cached at: 07/31/26, 05:53 AM

Paper page - ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Source: https://huggingface.co/papers/2607.28627

Abstract

Longvisualcontextposesachallengeforvision-languagemodels:performancedegradesasthenumberofdistractorsgrows,andprocessingalltokensatonceiscomputationallyinfeasibleunderGPUmemoryconstraints.WepresentReToken,asinglelearnableembeddingtrainedasanexplicitretrievaltargetthatselectsasparsesetofquery-relevantvisualtokensfromapre-filledvisualKVcache.Trainedononlyasmallimage-QAdataset,ReTokenyieldsconsistentgainsacrossimageandvideobenchmarks:onVisualHaystacksitimprovesQwen3VL-8Bby13.4pointsandInternVL3.5by12.4points(>20%relative),andonLVBenchittransferszero-shottolongvideoforan8.0-pointgainwithQwen3VL-8B.Thankstoitslightweightdesign,bothtrainingandlong-videoinferencefitonasingleH100.Codeisavailableat:https://github.com/avaxiao/ReToken

View arXiv pageView PDFGitHub1Add to collection

Get this paper in your agent:

hf papers read 2607\.28627

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.28627 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.28627 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.28627 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

EasyVideoR1: Easier RL for Video Understanding

Hugging Face Daily Papers

EasyVideoR1 is an efficient reinforcement learning framework for training large vision-language models on video understanding tasks, featuring offline preprocessing with tensor caching for 1.47x throughput improvement, a task-aware reward system covering 11 problem types, and evaluation across 22 video benchmarks. It also supports joint image-video training and a mixed offline-online data training paradigm.

Video2LoRA: Parametric Video Internalization for Vision-Language Models

Hugging Face Daily Papers

This paper introduces Video2LoRA, a method that predicts Low-Rank Adaptation (LoRA) weights directly from video representations, enabling efficient video processing in frozen vision-language models. It reduces visual token load by up to 1500x and query TTFT by 6-80x while maintaining performance on video summarization and captioning benchmarks.