AI Paper Atlas

Research map / Video / Video-language understanding

Task-Aware Mechanism: Hybrid MoE Vision Tower Towards Holistic Video Understanding

Qishen Yin, Tanghui Jia, Peng Jin, Li Hao, Juntong Wu, Guanlin Lu, Li Yuan

ICML 2026 · Video › Video-language understanding · approach: Long video understanding

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 MoE vision encoder alone (no DFR) (ablation)Qwen2-7B LVLM baseline · ActivityNet-QA, MVBench, MLVU-Testavg relative change: -2.7%Hurt
add Dynamic Frame Number & Resolution alone (ablation)Qwen2-7B LVLM baseline · ActivityNet-QA, MVBench, MLVU-Testavg relative change: +2.8%Improved
combine MoE vision encoder with DFR (ablation)TallVA · ActivityNet-QA, MVBench, MLVU-Testavg relative change: +12.1%Improved
add DFR and Time/Task instruction to Inductor routing (ablation)baseline uniform budget · MLVU-Testaccuracy: 44.8 -> 46.1Improved
train the LLM first in Stage 2 (ablation)TallVA · MVBenchaccuracy: 59.5 -> 61.8Improved
train Projector only or VE and Projector in Stage 2 (ablation)TallVA · MLVU-Testaccuracy: 46.1 -> 42.2 (Projector); 46.1 -> 32.7 (VE and Projector)Hurt
soft-gated MoE vision encoder (top2 in 4) (ablation)TallVA after stage 2 · MVBenchaccuracy: 61.8 -> 63.9Improved
hard-gate vision encoder by task type (ablation)TallVA after stage 2 · MVBenchaccuracy: 61.8 -> 56.4Hurt
increase soft MoE-VE experts from 4 to 8 (top2) (ablation)TallVA · MVBenchaccuracy: 63.9 -> 63.2Hurt
hard resolution-based MoE projector (ablation)TallVA after stage 2 · MVBenchaccuracy: 61.8 -> 62.6Improved
soft-gated MoE projector (ablation)TallVA after stage 2 · MVBenchaccuracy: 61.8 -> 61.9 (top2 in 4); 61.8 -> 61.6 (top2 in 8)No clear effect
sparse MoE vision encoder without DFR (Top2 in 8) (ablation)dense baseline · MVBenchaccuracy: 58.6 -> 53.3Hurt

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