Research map / Architectures and efficiency / Attention and transformers
AdaSplash-2: Faster Differentiable Sparse Attention
Nuno Gonçalves, Hugo Pitorro, Vlad Niculae, Edoardo Ponti, Lei Li, Andre Martins, Marcos Vinicius Treviso
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 |
|---|---|---|---|
| replace softmax with alpha-entmax (AdaSplash-2) under NAPE | LLaMA-3 1B · RULER avg, 32K | accuracy: 36.8 -> 39.4 | Improved |
| replace softmax with alpha-entmax under RoPE | LLaMA-3 1B · RULER avg, 4K to 32K | accuracy: 51.4 -> 49.3 (4K); 27.3 -> 23.4 (32K) | Hurt |
| replace softmax with alpha-entmax under NAPE | LLaMA-3 1B · HELMET ICL avg, 32K | accuracy: 46.1 -> 62.2 | Improved |
| replace softmax with alpha-entmax under RoPE | LLaMA-3 1B · HELMET ICL avg, 32K | accuracy: 38.2 -> 39.3 | Improved |
| use NAPE instead of RoPE with softmax | LLaMA-3 1B · RULER avg, 32K | accuracy: 27.3 -> 36.8 | Improved |
| replace softmax with alpha-entmax under NAPE | LLaMA-3 350M · LAMBADA | perplexity: 19.23 -> 18.62 | Improved |
| replace softmax with alpha-entmax under RoPE | LLaMA-3 1B · LAMBADA | perplexity: 15.01 -> 15.76 | Hurt |
| replace softmax with alpha-entmax under NAPE | LLaMA-3 1B · OLMES 9mcqa avg | accuracy: 53.0 -> 53.1 | No clear effect |
| replace softmax with alpha-entmax under RoPE | LLaMA-3 1B · OLMES 9mcqa avg | accuracy: 52.5 -> 51.2 | Hurt |
| initialize tau with histogram (B=4/8/16 bins) in hybrid solver (ablation) | AdaSplash-2 solver · Gaussian scores n=4096, alpha=1.5 | mean absolute error: 10000x closer to tau* after 1 iteration | Improved |
| increase number of histogram bins (ablation) | AdaSplash-2 solver · Gaussian scores n=4096, alpha=1.5 | mean absolute error | Improved |
| use AdaSplash-2 instead of AdaSplash | Triton attention kernel · causal attention, 128K context | step time: 2.1s -> 1.5s | Improved |
Closest papers by meaning
- AdaSplash: Adaptive Sparse Flash Attention (ICML 2025 · similarity 0.93)
- Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction (NeurIPS 2025 · similarity 0.89)
- SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference (ICML 2025 · similarity 0.88)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.88)
- Transformers Learn Faster with Semantic Focus (NeurIPS 2025 · similarity 0.87)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 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.
