O(NlogN) tree attention claims 550x KV compression with 79% accuracy at 16k context
Alarming-Emotion-894 · reddit · 2026-10-13
A Reddit update on ALHR (Adaptive Learnable Hierarchical Routing), a static binary-tree attention system using learnable functions to reduce keys read per token. On MQAR, it reads only 30-32 keys/token: 35x-550x KV compression vs dense baselines, with accuracy 92-93% at 4k context dropping to 79.2% at 16k. Author reports promising tiny-story tests but lacks resources for fuller results.
More from Research
- EXL details GNN fraud detection system with 27-33% detection gains — AI Engineer · 2026-10-13
- RTNav: open-source modular asynchronous agent for real-time zero-shot object navigation — rsasaki0109 · 2026-10-13
- AI-CFD benchmark culture: a mesh-free solver takes a cow to Mach 2 — burhop · 2026-10-13
- The Implicit Curriculum Hypothesis wins COLM outstanding paper award, powered by NSF public compute — IanArawjo · 2026-10-13
- Computational biology shifts as AI builds 'better predictors' in under a day — jmschreiber91 · 2026-10-13
- Model weights are only 1 of 4 VRAM allocations: why loaded LLMs still OOM — blaizedsouza · 2026-10-13