Joint Subspace Sparse Attention for Long-Context Decoding
October 12, 2026
JSSA optimizes long-context LLM decoding by learning a shared low-dimensional query-key subspace to reduce KV-cache traffic. This method provides a closed-form spectral solution to minimize token-score preservation error while maintaining fine-grained selection quality.
HOW THIS AFFECTS YOU
●
builderYou can achieve more efficient long-context decoding with reduced KV-cache traffic and latency.
●
researcherThe approach provides a mathematically tractable way to handle non-convex token selection objectives.