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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add Token Position-Decay to uniform top-k at equal total budget (ablation) | Stem on Qwen3-8B · LongBench | AVG accuracy: 29.41 -> 31.43 | Improved |
| add Token Position-Decay to uniform top-k at equal total budget (ablation) | Stem on Llama-3.1-8B-Instruct · LongBench | AVG accuracy: 37.48 -> 40.85 | Improved |
| add Output-Aware Metric on top of TPD (ablation) | Stem on Qwen3-8B · LongBench | AVG accuracy: 31.43 -> 31.64 | Improved |
| add Output-Aware Metric on top of TPD (ablation) | Stem on Llama-3.1-8B-Instruct · LongBench | AVG accuracy: 40.85 -> 41.48 | Improved |
| replace score-aware metric with output-aware metric for token selection (ablation) | sparse attention top-k 1024 · Qwen3-8B, 30 LongBench 8K samples | head logits MSE: 0.3368 -> 0.3126 | Improved |
| decay ratio mu = 0.7 instead of uniform budget (mu = 1.0) (ablation) | Stem on Qwen3-8B · LongBench | AVG accuracy: 31.64 vs 31.67 | No clear effect |
| lower decay ratio mu below 0.6 (ablation) | Stem · LongBench | AVG accuracy | Hurt |
| raise OAM coefficient beta above 0.3 (ablation) | Stem · LongBench | AVG accuracy | Hurt |
| add value magnitude term (beta 0 to 0.2) (ablation) | Stem · LongBench | AVG accuracy | Improved |
| apply Stem on top of DeepSeek-V3.2 DSA | DeepSeek-V3.2 · LongBench | AVG accuracy: 42.84 -> 43.16 | No clear effect |
| apply Stem decay schedule on top of MiniCPM-4.1 InfLLMv2 | MiniCPM-4.1 · LongBench | AVG accuracy: 42.75 -> 42.76 | No clear effect |
| use Stem instead of dense attention for prefill | Llama-3.1-8B-Instruct · 128K prefill latency | total latency: 1540ms -> 420ms | Improved |
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.
- XAttention — XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025) · better on 4 of 4 shared results
- FlexPrefill — FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025) · better on 2 of 2 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
- Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning (NeurIPS 2025 · similarity 0.88)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.88)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.87)
- Learning to Focus: Causal Attention Distillation via Gradient‐Guided Token Pruning (NeurIPS 2025 · similarity 0.87)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.87)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.87)
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.
