Original Summary

ALHR - Adaptive Learnable Hierarchical Routing, uses static binary trees and learnable functions to minimize the amount of keys to be read. It does use a dense teacher while phase 1 of training however. MQAR TEST AT 1024 TOKENS - Average keys read per query by dense - 512 Keys Average keys read per query by ALHR - 30 keys Top - 1 accuracy of dense - 94.9% Top - 1 accuracy of ALHR - 92.1% KV Compression of dense - 1x(100% read) KV Compression of ALHR - 35.3x(2.83% read) Peak VRAM of dense - 57 MB (Scales quadratically) Peak VRAM of ALHR - 422 MB (scales linearly) Cache compression of ALHR - 100% The true log and Kaggle cell used to run it are in the logs folder in the repo limitations: Full scale tests are still not completed, The training of this model would still be quadratic but the inference would be NlogN (as indicated in the logs in the repo) Would love your opinions<p>ALHR Repository: repo link in comments

  • 情报分类：综合情报
  • 分类依据：内容未命中明确的垂直分类规则，归入综合情报
  • 信息来源：Hacker News 新项目
  • 发布时间：2026/10/9 21:27:33