cs.CVNov 18, 2026

Let a Small Model Read First: Faster Prefill for Long Interleaved Image-Text Inputs

Authors: Xiaolin Lin, Kai-Chiang Wu, Bing Li, Grace Li Zhang

Organizations: Technische Universität Darmstadt · National Yang Ming Chiao Tung University · TU Ilmenau

Abstract

Vision-language models (VLMs) can answer questions about long interleaved inputs, such as documents of many pages and collections of hundreds of images, but they are slow to start: on inputs of about 100K tokens, the prefill takes most of the wait for the first output token. A common approach prunes the input by scoring its tokens with the target model itself or with a small model of the same family that first writes an answer. However, this scoring takes much of the time that pruning saves. We find that the evidence for a question lies in a few whole units (images or paragraphs), and that a small hybrid model finds them with one attention head of an early layer, even at half image resolution. Based on these findings, we propose SkimPrefill, which lets the small hybrid model read the whole input first, up to that early layer, with images at half resolution. Most layers of the small hybrid model (Qwen3.5-0.8B) run in linear time, so this pass is short. SkimPrefill then ranks every unit with that head and gives the target the top units as an ordinary shorter prompt. One untrained selector serves targets of different model families. Extensive experiments on MMLongBench at 128K tokens, with 19 datasets of five task types, show the effectiveness of SkimPrefill. With 10% of the input tokens kept, SkimPrefill speeds up the prefill of Qwen3.5-9B, Gemma-4-12B and Qwen3-VL-32B by 3.7 times, 11.6 time and 8.2 times while maintaining their average accuracy.

Figures & tables

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 18, 2026

Read Together, Answer Apart: How Earlier Questions Help Vision-Language Models Look

Vision-language models (VLMs) are often asked several questions about the same image in a single prompt (co-asking), yet existing evaluations report conflicting accuracy gains and hallucinations at the prompt level without tracking how individual answers change. By pairing each question's single-question and co-asked answers across 22 open-weight VLMs (1B-35B), we find that erratic net shifts reflect two opposing effects: on Densely Captioned Images (DCI), every model repairs 15.4%-29.8% of its single-question errors under co-asking, while the share of correct answers turned wrong falls from 30.8% on weaker models to 4.7% on stronger ones. Across all 14 models supporting custom attention masks, we show through attention readouts and layer-window interventions that earlier same-image questions read visual information into their question tokens in the early-to-middle layers during prefill, helping later questions concentrate image attention onto their queried targets. This visual help accumulates across the scene even across non-overlapping regions, whereas blocking earlier questions from attending to the image lowers both target localization and accuracy in all 14 models. While cross-question help is established during shared question reading, part of the co-asking damage enters later during answer decoding when the decoder attends to other question stems. We therefore introduce Read Together, Answer Apart, a training-free decoding mask that keeps shared prefill intact and isolates each answer from other stems, improving co-asking accuracy on 12 of 14 models on DCI (up to +4.71 percentage points, closing 70% of the pooled gap to asking alone) and on 10, 8, and 5 of the 14 models on WebUI, FinTabNet, and GQA (up to +11.4, +5.3, and +1.1 points).
Jul 30, 2026

ReToken: Improving Long-Context VLMs with Visual Retrieval Token

Long visual contexts challenge vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once can exceed GPU memory limits. We present RETOKEN, a single learnable embedding that extracts retrieval signals from the VLM's internal representations to select query-relevant visual tokens from the pre-filled KV cache. This enables retrieval within the answering VLM, without a separate retriever or re-encoding. Despite being trained on only a small image-QA dataset, RETOKEN generalizes 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). On LVBench, it transfers zero-shot to long video and improves Qwen3VL-8B by 8.0 points. 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
Nov 18, 2026

SEE: Unlocking the Hidden Perception of Frozen VLMs

Vision-language models (VLMs) grasp the overall content of an image but often miss the fine details a question hinges on, such as a small object in a large photo or a number in a dense infographic. Yet even when their answer is wrong, the region they point to often contains the evidence. We introduce Self-Elicited Evidence (SEE), a training-free method that lets a frozen VLM look again where it points. When the initial answer has low confidence, SEE prompts the model to propose regions where the details may appear, answers the question on these enlarged crops, and aggregates the answers with the original prediction through confidence-weighted voting. The model also answers with its own transcript, which helps on text-rich images. SEE needs no detector or OCR engine. Across five backbones, SEE improves every small-object benchmark by 2.3-13.1 points and outperforms five training-free baselines on small objects. With SEE, Qwen3.5-4B surpasses the direct answer of Qwen3.5-9B and matches the 6 times larger Qwen3.5-27B.