AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Stem: Rethinking Causal Information Flow in Sparse Attention

Lin Niu, Xin Luo, Linchuan Xie, Yifu Sun, Guanghua Yu, Jianchen Zhu, S Kevin Zhou

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse and structured attention · 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
add Token Position-Decay to uniform top-k at equal total budget (ablation)Stem on Qwen3-8B · LongBenchAVG accuracy: 29.41 -> 31.43Improved
add Token Position-Decay to uniform top-k at equal total budget (ablation)Stem on Llama-3.1-8B-Instruct · LongBenchAVG accuracy: 37.48 -> 40.85Improved
add Output-Aware Metric on top of TPD (ablation)Stem on Qwen3-8B · LongBenchAVG accuracy: 31.43 -> 31.64Improved
add Output-Aware Metric on top of TPD (ablation)Stem on Llama-3.1-8B-Instruct · LongBenchAVG accuracy: 40.85 -> 41.48Improved
replace score-aware metric with output-aware metric for token selection (ablation)sparse attention top-k 1024 · Qwen3-8B, 30 LongBench 8K sampleshead logits MSE: 0.3368 -> 0.3126Improved
decay ratio mu = 0.7 instead of uniform budget (mu = 1.0) (ablation)Stem on Qwen3-8B · LongBenchAVG accuracy: 31.64 vs 31.67No clear effect
lower decay ratio mu below 0.6 (ablation)Stem · LongBenchAVG accuracyHurt
raise OAM coefficient beta above 0.3 (ablation)Stem · LongBenchAVG accuracyHurt
add value magnitude term (beta 0 to 0.2) (ablation)Stem · LongBenchAVG accuracyImproved
apply Stem on top of DeepSeek-V3.2 DSADeepSeek-V3.2 · LongBenchAVG accuracy: 42.84 -> 43.16No clear effect
apply Stem decay schedule on top of MiniCPM-4.1 InfLLMv2MiniCPM-4.1 · LongBenchAVG accuracy: 42.75 -> 42.76No clear effect
use Stem instead of dense attention for prefillLlama-3.1-8B-Instruct · 128K prefill latencytotal latency: 1540ms -> 420msImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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