Research map / Architectures and efficiency / Attention and transformers
SSA: Sparse Sparse Attention by Aligning Full and Sparse Attention Outputs in Feature Space
Zhenyi Shen, Junru Lu, Lin Gui, Jiazheng Li, Yulan He, di yin, Xing Sun
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 alignment loss (bidirectional) to SSA training (ablation) | SSA 300M · Commonsense avg | average accuracy: 48.38 -> 49.69 | Improved |
| train with a mix of full and sparse streams (FullRatio=0.5) instead of sparse only (ablation) | SSA 300M · Commonsense avg | average accuracy: 49.00 -> 49.69 | Improved |
| mix in the sparse stream (FullRatio=0.5) instead of full stream only (ablation) | SSA 300M · Wikitext PPL | perplexity: 25.16 -> 24.20 | Improved |
| use only one direction of alignment loss (Full to Sparse or Sparse to Full) (ablation) | SSA 300M · training stability | PPL / Comm. Avg.: NaN, 0.00 | Hurt |
| sample the attention mode randomly per layer instead of per step (ablation) | SSA 300M · training stability | PPL / Comm. Avg.: NaN, 0.00 | Hurt |
| train with larger receptive field (16x64 instead of inference-only extrapolation from 8x16) (ablation) | SSA 300M · Commonsense avg | average accuracy: 49.96 -> 49.16 | Hurt |
| train with receptive field 8x16 instead of extrapolating to it (ablation) | SSA 300M · Commonsense avg | average accuracy: 49.69 -> 48.88 | Hurt |
| alignment coefficient alpha 10 instead of 5 (ablation) | SSA 300M · Commonsense avg | average accuracy: 48.71 -> 49.69 | Improved |
| use L2 alignment loss instead of SmoothL1 (ablation) | SSA 300M · Commonsense avg | average accuracy: 49.69 -> 49.27 | Hurt |
| add sliding window to NSA (ablation) | NSA RF=256 · Commonsense avg | average accuracy: 58.9 -> 59.47 | Improved |
| add sliding window to NSA (ablation) | NSA, PG-19 32K RoPE scaling · PG-19 PPL 32K | perplexity: 18.16 -> 15.00 | Improved |
| SSA training instead of MoBA at sparse inference RF=256 | 1B model · LongBench 32k extrapolation | LongBench score: 16.88 -> 17.89 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MoBA — MoBA: Mixture of Block Attention for Long-Context LLMs (NeurIPS 2025) · better on 9 of 9 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
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.89)
- Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction (NeurIPS 2025 · similarity 0.88)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.88)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.88)
- SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference (ICML 2025 · similarity 0.87)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · 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.
