KV Cache Splicing Method for Frozen Gemma 4
MindPsychological140 · reddit · 2026-07-19
The author introduces a method to save verified knowledge as KV states and restore them when needed, achieving byte-level identical results compared to recomputation.
They report that on Gemma 4 12B, this cached knowledge approach boosted a routing system's performance on AIME 2025 from 76.7% to 90.0%. The post includes a link to the paper (arXiv:2607.14431) and notes that the author will present at the AGI Summit on July 19.
Related event: KV-Cache Grafting Boosts Frozen Small Models(2 posts)→
More from Research
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26
- Ortet evaluates 34 open-weight models on 50 health tasks, finds general leaderboards insufficient — elmanmansimov · 2026-10-26