AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

DIGS: Differentiable, Incremental, Global, Scalable Pruning for Language Models

Bingcong Li, Junlin Xian, TAO JIANG, jwlin jwlin, Cheng ZOU, Geng-Li Zhang

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 0 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
increase pruning token budget 5.7B -> 24.5BDIGS · Qwen2.5VL-3B, 60% sparsityAvg accuracy: 47.92 -> 51.93Improved
increase pruning token budget 5.7B -> 24.5BL0-Dual · Qwen2.5VL-3B, 60% sparsityAvg accuracy: 44.51 -> 38.97Hurt
increase pruning token budget 5.7B -> 24.5BDIGS · Qwen2.5VL-3B, 30% sparsityAvg accuracy: 56.04 -> 58.75Improved
increase pruning token budget 5.7B -> 24.5BL0-Dual · Qwen2.5VL-3B, 30% sparsityAvg accuracy: 58.26 -> 56.85Hurt
replace dual ascent with per-step gradient-norm control (DIGS vs L0-Dual) at small budgetL0-Dual · Qwen2.5VL-3B, 5.7B tokensAvg accuracy: 30%: 58.26 -> 56.04; 60%: 44.51 -> 47.92Mixed
replace dual ascent with gradient-norm control at 24.5B tokensL0-Dual · Qwen2.5VL-3B, 60% sparsityAvg accuracy: 38.97 -> 51.93Improved
add gradient clipping to L0-Dual (ablation)L0-Dual · 60% sparsity, ~12B tokensAvg3: 46.32 -> 45.69No clear effect
cap lambda at 50 in L0-Dual (ablation)L0-Dual · 60% sparsityAvg3: 46.32 -> 52.74 (but ~12B -> ~19B tokens)Improved
use DIGS instead of stabilized L0-Dual (ablation)L0-Dual+lambda cap · 60% sparsityAvg3: 52.74 -> 56.29Improved
lower c_g (larger pruning budget) (ablation)DIGS · 60% sparsityAvg3: 45.73 -> 57.37Improved
add post-pruning distillation (ablation)DIGS 114B pruned · Qwen2.5VL-3B (33% sparsity)Avg accuracy: 61.52 -> 63.94Improved
add FineVision SFT (ablation)DIGS 114B pruned + distill · Qwen2.5VL-3B (33% sparsity)Avg accuracy: 63.94 -> 64.32Improved

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