CRISP (arXiv 2609.01925, Thu 3 Sep digest; Adobe Research) is an input-adaptive sparse attention method for long-context prefilling. It replaces Jensen-Shannon-divergence-based head routing with a structured surrogate, C_struct, that measures the quality of Vertical-Slash-compatible positions, removing the pooled matmul and KL-divergence overhead; and replaces cumulative coverage thresholds with sink-aware thresholds that theoretically accumulate only O(n) background noise. Across InfiniteBench, RULER, and LongBench on two model families it is the strongest overall sparse method, matching or exceeding dense attention on retrieval-heavy benchmarks with recovery up to +28.0 points over baselines, and reaches up to 5.30x attention speedup at 512k tokens. Accepted to EMNLP 2026 main.
Prefill's quadratic cost is the wall for long-context serving, and RAG-style workloads are exactly the retrieval-heavy profile where CRISP claims no quality loss versus dense attention. For teams self-hosting long-context models, this is a drop-in accelerator candidate for the most expensive phase of the request.
| Arxiv | 2609.01925, announced in the Thu 3 Sep 2026 digest |
|---|---|
| Org | Adobe Research (first author Huu Huy Nguyen, with Franck Dernoncourt, Ryan Rossi) |
| Mechanism | C_struct structured surrogate replaces JSD-based head routing (removes pooled matmul + KL overhead); sink-aware thresholds replace cumulative coverage thresholds (O(n) background noise accumulation) |
| Results | strongest overall sparse method on InfiniteBench/RULER/LongBench across two model families; matches/exceeds dense on retrieval-heavy benchmarks; recovery up to +28.0 points; up to 5.30x attention speedup at 512k tokens |
| Venue | EMNLP 2026 main |