Research map / Architectures and efficiency / Attention and transformers
vAttention: Verified Sparse Attention via Sampling
Aditya Desai, Kumar Krishna Agrawal, Shuo Yang, Alejandro Cuadron, Luis Gaspar Schroeder, Matei Zaharia, Joseph E. Gonzalez, Ion Stoica
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 vAttention sampling to HashAttention | HashAttention · RULER32K-HARD (Llama-3.1-8B) | average accuracy: 81.94 -> 86.56 | Improved |
| add vAttention sampling to HashAttention | HashAttention · RULER32K-HARD (DeepSeek-R1-Distill-Llama-8B) | average accuracy: 60.70 -> 65.06 | Improved |
| add vAttention sampling to HashAttention | HashAttention · RULER32K-HARD (Mistral-7B) | average accuracy: 54.66 -> 56.90 | Improved |
| add vAttention sampling to oracle-top-k | oracle-top-k · RULER32K-HARD (Llama-3.1-8B) | average accuracy: 87.18 -> 88.61 | Improved |
| add vAttention sampling to oracle-top-k | oracle-top-k · RULER32K-HARD (DeepSeek-R1-Distill-Llama-8B) | average accuracy: 64.87 -> 65.15 | No clear effect |
| add vAttention sampling to oracle-top-k | oracle-top-k · RULER32K-HARD (Mistral-7B) | average accuracy: 64.37 -> 64.12 | No clear effect |
| add vAttention sampling to HashAttention | HashAttention · RULER32K easy datasets (Llama-3.1-8B) | average accuracy: 85.80 -> 84.57 | Hurt |
| add vAttention sampling to HashAttention | HashAttention · RULER32K full benchmark (Llama-3.1-8B) | average accuracy: 83.87 -> 85.57 | Improved |
| add vAttention sampling to oracle-top-k | oracle-top-k · LongBench (7 datasets, 10% sparsity) | average score: 39.35 -> 41.33 | Improved |
| add vAttention sampling to HashAttention | HashAttention · LongBench (7 datasets, 10% sparsity) | average score: 38.64 -> 40.04 | Improved |
| add vAttention to oracle-top-k for long generation | oracle-top-k · AIME 2024 (Avg@4) | accuracy: 36.66 -> 36.67 | No clear effect |
| use Hoeffding bound instead of CLT for sample budget (ablation) | vAttention · RULER 16K, eps=0.1 delta=0.2 | samples per head: 2.8× more samples (average 874) | Hurt |
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.
- HashAttention — HashAttention: Semantic Sparsity for Faster Inference (ICML 2025) · better on 5 of 6 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 0.89)
- 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)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · 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)
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.
