Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel
A frozen Gemma-4-12B reportedly jumped to 93.3% on AIME 2025 by grafting verified KV-cache states instead of changing weights.
The paper says verified knowledge can be stored once as a byte-exact KV artifact and restored into later inference with identical logits under a pinned deterministic setup. It reports zero KL divergence, full argmax agreement across fifty samples, and SHA-256 equality for fresh versus grafted computation. The author says the method also cut recurring-case token use from a 401,026-token budget to 61 decode tokens, with an estimated 8,700x energy reduction. The engine is proprietary, but the paper says reported numbers are backed by committed input and output hashes for re-checking. HF Daily Papers' note
The paper says verified knowledge can be stored once as a byte-exact KV artifact and restored into later inference with identical logits under a pinned deterministic setup. It reports zero KL divergence, full argmax agreement across fifty samples, and SHA-256 equality for fresh versus grafted computation. The author says the method also cut recurring-case token use from a 401,026-token budget to 61 decode tokens, with an estimated 8,700x energy reduction. The engine is proprietary, but the paper says reported numbers are backed by committed input and output hashes for re-checking. HF Daily Papers' note
score 4