AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

From Drop-off to Recovery: A Mechanistic Analysis of Segmentation in MLLMs

Boyong Wu, Sanghwan Kim, Zeynep Akata

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 2 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
project vision features through the adapter into the LLM spaceLLaVA-1.5 (CLIP, DINOv2, SigLIP) · ADE20K linear probemIoU: DINOv2 adapter 42.90 vs encoder baseline above peak 44.50Hurt
pass image tokens through LLM layersLLaVA-1.5 CLIP · ADE20K linear probemIoU: adapter 33.22 -> peak 40.74Improved
pass image tokens through LLM layersLLaVA-1.5 DINOv2 · ADE20K linear probemIoU: adapter 42.90 -> peak 44.50; peak stays below encoder (-1.08)Mixed
block attention to correctly classified tokens (ablation)LLaVA-1.5 CLIP · ADE20K conflicting classessegmentation qualityHurt
block attention to incorrectly classified tokens (ablation)LLaVA-1.5 CLIP · ADE20K conflicting classessegmentation qualityImproved
add bidirectional attention among image tokens (ablation)LLaVA-1.5 CLIP · ADE20K patch 0pixel accuracy: 0.6195 -> 0.7630Improved
add bidirectional attention among image tokens (ablation)LLaVA-1.5 CLIP · ADE20K linear probepeak mIoU: 40.74 -> 41.16Improved
add bidirectional attention among image tokens (ablation)LLaVA-1.5 DINOv2 · ADE20K linear probepeak mIoU: 44.50 -> 44.82Improved
add bidirectional attention among image tokens (ablation)LLaVA-1.5 SigLIP · ADE20K linear probepeak mIoU: 35.39 -> 41.26Improved
add bidirectional attention among image tokens (ablation)LLaVA-1.5 DINOv2 · VQA benchmarksaccuracy: GQA 62.1 -> 60.5; MMBench 57.7 -> 55.3; VizWiz 51.4 -> 45.8Hurt
add bidirectional attention among image tokens (ablation)LLaVA-1.5 CLIP · VQA benchmarksaccuracy: GQA 62.6 -> 62.7; MMBench 66.4 -> 65.5No clear effect
add bidirectional attention among image tokens (ablation)LLaVA-1.5 SigLIP · VQA benchmarksaccuracy: MMBench 63.7 -> 66.9; VizWiz 58.2 -> 53.3Mixed

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