AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Why and When Visual Token Pruning Fails? A Study on Relevant Visual Information Shift in MLLMs Decoding

Jiwan Kim, Kibum Kim, Wonjoong Kim, Byung-Kwan Lee, Chanyoung Park

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 9 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add DSTP (detect RVIS, swap in reserved tokens) to FastVQwen3-VL-4B · visual reasoning avg, 33.3% tokensrelative accuracy: 55.68% -> 82.90%Improved
add DSTP to DivPruneInternVL3.5-8B · visual reasoning avg, 33.3% tokensrelative accuracy: 63.55% -> 90.09%Improved
add DSTP to VisionZipQwen3-VL-4B · visual reasoning avg, 22.2% tokensrelative accuracy: 46.40% -> 68.74%Improved
add DSTP to FastVQwen3-VL-4B · ScienceQA (33.3%)accuracy: 87.98 -> 91.84Improved
add DSTP to VisionZipQwen3-VL-4B · VQAT (33.3%)accuracy: 77.10 -> 77.20No clear effect
RISD fixed-threshold detection instead of random trigger (ablation)DSTP · MathVerseaccuracy: 37.69 -> 52.54Improved
fixed threshold instead of running-average threshold (ablation)DSTP · MathVerseaccuracy: 51.51 -> 52.54Improved
CPTS union swap instead of Hard selection (ablation)DSTP · MathVerseaccuracy: 45.05 -> 52.54Improved
CPTS union swap instead of Merge (ablation)DSTP · MMMU-Proaccuracy: 28.38 -> 30.52Improved
CPTS instead of full visual tokens (ablation)DSTP · MathVerseaccuracy: 53.04 -> 52.54No clear effect
add DSTP to FastVQwen3-VL-4B · MathVerselatency per example: 119.35s -> 102.55sImproved
add DSTP to FastVInternVL3.5-8B · MathVerselatency per example: 9.14s -> 9.99sHurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 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