AI Paper Atlas

Research map / Video / Video-language understanding

Decomposed Attention Fusion in MLLMs for Training-free Video Reasoning Segmentation

Su Ho Han, Jeongseok Hyun, Pilhyeon Lee, Minho Shim, Dongyoon Wee, Seon Joo Kim

ICLR 2026 · Video › Video-language understanding · approach: Long video understanding · 5 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 contrastive object-background fusion (ablation)DecAF · Ref-DAVIS attention mask, Qwen2.5VL-7BJ&F: 14.5 -> 25.3Improved
add contrastive object-background fusion (ablation)DecAF · ReasonVOS SAM mask, Qwen2.5VL-7BJ&F: 58.4 -> 63.9Improved
combine video and frame attention (ablation)DecAF · Ref-DAVIS, Qwen2.5VL-7BJ&F: 65.9 / 67.4 -> 75.2Improved
add multi-scale frame input (ablation)DecAF · Ref-DAVIS, InternVL3-8BJ&F: 54.0 -> 62.8Improved
use V-Max normalized rollout instead of plain rollout (ablation)DecAF · Ref-DAVIS, Qwen2.5VL-7BJ&F: 68.4 -> 75.2Improved
add attention consistency score (ablation)DecAF SAM prompting · Ref-DAVIS, Qwen2.5VL-7BJ&F: 60.0 -> 75.2Improved
add penalty term to consistency score with Otsu threshold (ablation)DecAF · Ref-DAVIS, Qwen2.5VL-7BJ&F: 68.3 -> 67.9Hurt
start rollout from layer 21 instead of 14 (ablation)DecAF · Ref-DAVIS / ReasonVOS, Qwen2.5VL-7BJ&F: Ref-D 75.2 -> 73.6; ReasV 63.9 -> 64.1Mixed
raise tau_trk and tau_pq to 0.9 (ablation)DecAF SAM prompting · ReasonVOS, Qwen2.5VL-7BJ&F: 63.9 -> 66.4 (ReasonVOS); 75.2 -> 74.2 (Ref-DAVIS)Mixed
scale InternVL3 from 8B to 14BDecAF · ReVOS overallJ&F: 47.4 -> 47.0No clear effect
scale Qwen2.5VL from 3B to 7BDecAF · Ref-DAVISJ&F: 58.1 -> 75.2Improved
use DecAF instead of Loc-Head with InternVL3-8BSAM2 prompting · Ref-DAVISJ&F: 66.3 -> 62.8Hurt

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