标签
ALHR是一种基于静态二叉树和可学习预算预测器的注意力机制，将推理复杂度降至O(NlogN)，在MQAR长上下文任务上保持约97%的密集基线精度，同时显著提升KV压缩和显存扩展效率。
This paper presents a certificate-carrying sub-quadratic method for computing bisimulation metrics in Markov decision processes using approximate nearest neighbors, with coverage-augmented guarantees and two-sided bounds. Experiments show improved scaling and accurate metric recovery compared to baselines.