AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Semantic Impact–Driven Visual Scheduling in Vision-Language Models

Xuan Wang, Yilin Liu, Fangxiang Feng, Caixia Yuan, Huixing Jiang, Xiaojie Wang

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a training-free framework that schedules visual tokens in VLMs by measuring their semantic impact on the output distribution, rather than via attention weights, to reduce inference latency.

Problem

VLMs have high inference latency because of long visual token sequences, and existing methods infer visual necessity from intermediate signals such as attention weights.

Approach

Assess visual necessity by its semantic impact on the output distribution, using a token embedding subspace, rather than attention-based interaction signals; applied on demand in a training-free manner.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
score visual tokens by semantic impact instead of attention weight (ablation)SIVS · TextVQA (no OCR)accuracy (%): 37.16 -> 38.40Improved
score visual tokens by semantic impact instead of attention-weighted value norm (ablation)SIVS · TextVQA (no OCR)accuracy (%): 36.90 -> 38.40Improved
SIVS versus FastV, SparseVLM, VisPruner at higher compressionvisual KV compression · LLaVA-1.5-7B seven benchmarksrelative accuracy: 99.95% at 87.66% compressionImproved
SIVS versus vanilla on POPE and MileBenchSIVS · LLaVA-1.5-7Bscore: 85.9 -> 86.2 (POPE); 38.08 -> 38.94 (MileBench)Improved
SIVS versus vanilla on GQA and TextVQASIVS · LLaVA-1.5-7Bscore: 61.9 -> 61.7 (GQA); 58.3 -> 56.85 (TextVQA)Hurt
apply SIVS to Qwen2.5-VL 7BSIVS · TextVQAscore: 84.50 -> 81.92Hurt
apply SIVS to Qwen2.5-VL 3B-72BSIVS · GQAscore: 58.89 -> 58.86 (3B); 62.16 -> 62.26 (72B)No clear effect
increase Top-K size K (ablation)SIVS · TextVQAaccuracyImproved
increase threshold coefficient alpha beyond 0.5 (ablation)SIVS · TextVQAaccuracyHurt
refresh visual selection every step (smaller T) (ablation)SIVS · LLaVA-Benchaccuracy and latencyMixed
use mid-to-late layer for semantic impact estimation (ablation)SIVS · TextVQAaccuracyImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy