DataFunSummit

Jul 19, 2026 · Artificial Intelligence

Asymmetric KV Decoupling and Head-Level Competition: Three Strategies for Efficient Long-Text Reasoning

The article analyzes the fundamental trade‑off between global competition in attention and efficiency when processing million‑token contexts, and presents three complementary techniques—AsymKV, Compression‑Aware Tuning (CAT), and Softmax Linear Attention (SLA)—that preserve or rebuild global competition while dramatically reducing memory and compute costs, achieving state‑of‑the‑art results on LongBench and various retrieval benchmarks.