AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Sparsity Forcing: Reinforcing Token Sparsity of MLLMs

Feng Chen, Yefei He, Lequan Lin, Jing Liu, Chenhui Gou, Bohan Zhuang, Qi Wu

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 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
post-train with Sparsity Forcing (RL) on top of ZipVLQwen2.5-VL-7B · 7 image benchmarks avgAvg. score: ratio 79.5% -> 24.7%; avg 72.9 -> 73.6Improved
Sparsity Forcing vs full attentionQwen2.5-VL-7B · 7 image benchmarks avgAvg. score: 73.8 -> 73.6 at 24.7% tokensNo clear effect
Sparsity Forcing instead of FastV/VisionZipQwen2.5-VL-3B · 7 image benchmarks avgAvg. score: 65.1/65.8 -> 68.7Improved
Sparsity Forcing vs full attention on videoLLaVA-Video-7B · 6 video benchmarks avgAvg. score: 61.4 -> 61.3 at 29.6% tokensNo clear effect
RL with joint reward vs SFT sparsity baselines (MOBA, sharpness loss)Qwen2.5-VL-7B · Post-training comparison avgAvg. score: 66.2 / 66.8 -> 70.0Improved
top-p sparse attention instead of top-k (ablation)Sparsity Forcing · VideoMMEVideoMME: 60.2 -> 64.0Improved
top-p instead of threshold pruning (ablation)Sparsity Forcing · VideoMMEVideoMME: 61.6 -> 64.0Improved
lower training p range [0.80,0.95] (ablation)Sparsity Forcing · VideoMMEVideoMME: 64.0 -> 63.6No clear effect
group size 12 instead of 8 (ablation)Sparsity Forcing · MME/VideoMMEVideoMME: 64.0 -> 63.9No clear effect
add latency to efficiency reward (ablation)Sparsity Forcing · 6 video benchmarks avgAvg. score: 61.3 -> 61.2No clear effect
replace sharpness-loss with RL-based Sparsity Forcing over temperature adjustmentQwen2.5-VL-7B · Post-training comparison avgAvg. score: 67.4 -> 70.0Improved

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