AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

SPLIT-VLM: Salience-Guided Partitioning towards Local Coverage for Importance-Aware Token Dropping in Vision-Language Models

Seungil Lee, Gilha lee, Hyun Kim

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 SPLIT-VLM, a salience-guided partitioning method for importance-aware visual token dropping in vision-language models to reduce compute and memory.

Problem

Vision tokens dominate compute and memory in VLMs; prior token-reduction methods trade off salience (position bias, extra compute) against diversity (under-covers salient regions, hyperparameter-sensitive).

Approach

Combines salience and local coverage via salience-guided partitioning to choose which vision tokens to drop, aiming to get benefits of both salience and diversity without their stated drawbacks. Full details unavailable beyond the abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use Temporal Shift for allocation and Diversity Score for selection instead of Diversity Score for both (ablation)SPLIT · Qwen2-VL-7B, 88.9% reductionAverage accuracy (%): 85.0 -> 85.6Improved
use Temporal Shift for allocation instead of Diversity Score (selection by Diversity Score) (ablation)SPLIT · Qwen2-VL-7B, 88.9% reductionAverage accuracy (%): 85.0 -> 85.6Improved
use Diversity Score for selection instead of Temporal Shift (allocation by Temporal Shift) (ablation)SPLIT · Qwen2-VL-7B, 88.9% reductionAverage accuracy (%): 85.2 -> 85.6Improved
add region-wise budget allocation to VisPruner (ablation)VisPruner · LLaVA-1.5-7B, 64 tokensAverage relative accuracy (%): 91.4 -> 91.7Improved
add region-wise budget allocation to VisionZip (ablation)VisionZip · LLaVA-1.5-7B, 64 tokensAverage relative accuracy (%): 90.5 -> 91.0Improved
add region-wise budget allocation to CDPruner (ablation)CDPruner · LLaVA-1.5-7B, 64 tokensAverage relative accuracy (%): 92.4 -> 93.0Improved
add region-wise budget allocation to CDPruner (imbalance Gini) (ablation)CDPruner · LLaVA-1.5-7B, 64 tokensGini coefficient of retained tokens: 0.325 -> 0.269Improved
use importance-based adaptive allocation instead of uniform or random allocation (ablation)region budget · LLaVA-1.5-7B (POPE, MME)POPE / MMEImproved
use 16 regions instead of finer partitioning (e.g., 6x6) at 64 tokens (ablation)SPLIT · LLaVA-1.5-7B, 64 tokensaccuracyImproved
compute temporal shift across all ViT layers instead of a subset (ablation)SPLIT · LLaVA-1.5-7BaccuracyImproved
add cross-modal importance to Diversity Score selection after an early LLM layer (SPLIT-LLM B) (ablation)SPLIT · Qwen2-VL-7B, 88.9% reductionAverage accuracy (%): 85.6 -> 86.0Improved
use cross-modal similarity for region budget allocation (SPLIT-LLM A) (ablation)SPLIT · Qwen2-VL-7B, 88.9% reductionAverage accuracy (%): 85.6 -> 85.6No clear effect

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