Training-free Q-axis sparse attention for long-context LLM decode — exact ranking, a measured entropy-concentration scaling law, long-range retention to 128K, and a per-query accuracy certificate, validated under matched-control discipline (Llama-3.2-1B).
reproducible-research transformers sparse-attention attention-long-context llm-inference-kv-cache q-axis-dcr-attention
-
Updated
May 31, 2026 - TeX