Research map / Architectures and efficiency / Attention and transformers
Transformers meet Stochastic Block Models: Attention with Data-Adaptive Sparsity and Cost
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use data-adaptive SBM attention instead of full attention | SBM-Transformer · LRA (five tasks) | accuracy: outperforms full attention in all five tasks | Helped |
| Add density regularization with λ=10⁻⁴ (ablation) | SBM-Transformer · LRA average | accuracy: 59.93 -> 60.00 | Helped |
| Add density regularization with λ=10⁻³ (ablation) | SBM-Transformer · LRA average | accuracy: 59.93 -> 59.81 | No clear effect |
| Add density regularization with λ=10⁻² (ablation) | SBM-Transformer · LRA average | accuracy: 59.93 -> 59.84 | No clear effect |
| Add density regularization with λ=10⁻¹ (ablation) | SBM-Transformer · LRA average | accuracy: 59.93 -> 59.37 | Hurt |
| Add density regularization with λ=10⁻¹ (ablation) | SBM-Transformer · LRA average | attention density: 20.09% (ListOps), 26.10% (Text), 29.46% (Retrieval), 20.49% (Image), 18.56% (Pathfinder) -> 1.15%, 10.62%, 2.49%, 1.33%, | Helped |
| Manually add self-loops to attention masks (ablation) | SBM-Transformer · LRA average | accuracy: 59.93 -> 59.40 | Hurt |
| Add local attention with a sliding window of size 64 | SBM-Transformer · GLUE | F1 | Helped |
| Use SBM attention instead of the best reported baseline | SBM-Transformer · GLUE QQP | F1: 84.7 -> 85.2 | Helped |
| Use SBM attention instead of full attention | SBM-Transformer · GLUE QNLI | F1: 85.0 -> 83.6 | Hurt |
Closest papers by meaning
- Transformers Learn Faster with Semantic Focus (NeurIPS 2025 · similarity 0.89)
- XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025 · similarity 0.88)
- On the Role of Attention Masks and LayerNorm in Transformers (NeurIPS 2024 · similarity 0.88)
- kNN Attention Demystified: A Theoretical Exploration for Scalable Transformers (ICLR 2025 · similarity 0.88)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.88)
- Even Sparser Graph Transformers (NeurIPS 2024 · similarity 0.88)
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.