AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
Use dual-band spectral block scoring with energy calibration instead of mean-pooled coarse attention heuristicsPrism · PG19 128K prefillspeedup vs FlashAttention-2: XAttention 3.0x -> Prism 5.1xImproved
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 densityperplexityNo 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 densityperplexityImproved
Restrict the high-frequency branch to the dead zone (d_high=32) (ablation)Prism · PG19 32K, perplexity vs densityperplexityHurt
Use aggressive semantic slicing (d_low=64) that excludes the transition zone (ablation)Prism · PG19 32K, perplexity vs densityperplexityMixed
Add energy-based temperature calibration instead of fixed temperature 1.0 (ablation)Prism · PG19, perplexity vs densityperplexity at given densityImproved
Use Prism instead of FlexPrefill on LongBenchLlama-3.1-8B-Instruct · LongBench averageaverage score: 33.90 -> 41.08Improved
Use Prism sparse attention instead of full attention on Qwen3-8B Few-shotQwen3-8B · LongBench Few-shot Learningscore: 56.69 -> 58.36Improved
Use Prism sparse attention instead of full attention on RULERLlama-3.1-8B-Instruct · RULER averageaccuracy: 88.94 -> 87.54Hurt
Use Prism instead of full attention on long VideoMMEQwen3-VL-8B · VideoMME Longaccuracy: 63.11 -> 64.00Improved
Use Prism instead of full attention on LongVideoBenchQwen3-VL-8B · LongVideoBench Overallaccuracy: 65.00 -> 64.25Hurt
Use Prism instead of XAttention for dense video generation at matched qualityHunyuanVideo · HunyuanVideo generationspeedup: 1.54x -> 1.60xImproved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy