Get Started
Home
Topics
Search
Library
Research questionHow can input-adaptive sparse attention reduce long-context prefilling cost without losing retrieval-relevant context?Long-context self-attention prefilling grows quadratically, while fixed sparse patterns can miss input-dependent structure. Dynamic routing can add overhead and misallocate attention mass as context length increases.
AI
Evaluation & Benchmarks
Inference Optimization
Natural Language Processing
Latest papersRecent research connected to this question, newest first.CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated RoutingThe source presents CRISP for dynamic sparse prefilling, using a structural routing proxy and a sink-aware threshold for attention-mass selection. Evidence covers InfiniteBench, RULER, and LongBench across two model families, including retrieval performance and attention speed at contexts up to 512k tokens.research paper · Sep 1, 2026
Related questions
How can long-context LLM inference reduce KV-cache memory without losing attention-head-specific information?How can we measure and control an LLM’s reliance on token-frequency priors when context is sparse?How can long-context language models reduce per-token reads from million-token KV caches without losing accuracy?How can language-model attention remain reliable beyond its training context?