Research map / Architectures and efficiency / Attention and transformers
Prism: Spectral-Aware Block-Sparse Attention
Xinghao Wang, Pengyu Wang, Xiaoran Liu, Fangxu Liu, Jason Chu, Kai Song, Xipeng Qiu
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 |
|---|---|---|---|
| Use dual-band spectral block scoring with energy calibration instead of mean-pooled coarse attention heuristics | Prism · PG19 128K prefill | speedup vs FlashAttention-2: XAttention 3.0x -> Prism 5.1x | Improved |
| Use only the low-frequency band (d_low=96, d_high=0) instead of full dimension for pooled importance (ablation) | Prism block selection · PG19 32K, perplexity vs density | perplexity | No clear effect |
| Add a high-frequency branch covering the transition zone (d_high=64) instead of only the dead zone (32) (ablation) | Prism · PG19 32K, perplexity vs density | perplexity | Improved |
| Restrict the high-frequency branch to the dead zone (d_high=32) (ablation) | Prism · PG19 32K, perplexity vs density | perplexity | Hurt |
| Use aggressive semantic slicing (d_low=64) that excludes the transition zone (ablation) | Prism · PG19 32K, perplexity vs density | perplexity | Mixed |
| Add energy-based temperature calibration instead of fixed temperature 1.0 (ablation) | Prism · PG19, perplexity vs density | perplexity at given density | Improved |
| Use Prism instead of FlexPrefill on LongBench | Llama-3.1-8B-Instruct · LongBench average | average score: 33.90 -> 41.08 | Improved |
| Use Prism sparse attention instead of full attention on Qwen3-8B Few-shot | Qwen3-8B · LongBench Few-shot Learning | score: 56.69 -> 58.36 | Improved |
| Use Prism sparse attention instead of full attention on RULER | Llama-3.1-8B-Instruct · RULER average | accuracy: 88.94 -> 87.54 | Hurt |
| Use Prism instead of full attention on long VideoMME | Qwen3-VL-8B · VideoMME Long | accuracy: 63.11 -> 64.00 | Improved |
| Use Prism instead of full attention on LongVideoBench | Qwen3-VL-8B · LongVideoBench Overall | accuracy: 65.00 -> 64.25 | Hurt |
| Use Prism instead of XAttention for dense video generation at matched quality | HunyuanVideo · HunyuanVideo generation | speedup: 1.54x -> 1.60x | 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
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.86)
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · similarity 0.86)
- CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection (NeurIPS 2026 · similarity 0.85)
- Block Sparse Flash Attention (NeurIPS 2026 · similarity 0.85)
- XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025 · similarity 0.84)
- Fourier Position Embedding: Enhancing Attention’s Periodic Extension for Length Generalization (ICML 2025 · similarity 0.84)
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.
