We report a method to enhance the capabilities of a frozen small language model while dramatically reducing costs, without changing any weights. Verified knowledge is deposited as a byte-exact key-value (KV) state artifact, which is later restored into a fresh inference context via grafting.
The restoration is bit-exact: under a pinned deterministic configuration, the grafted logits are byte-for-byte identical to a fresh computation (SHA-256 equality), with zero KL divergence and 100% argmax agreement over fifty samples.
We demonstrate that own-position graft is the unique numerically exact operating point on a model with floating-point rotary encoding, verifying byte-exactness on two model scales (12B, 31B) and two GPU targets, one through a pre-registered replay.
In AIME 2025, a frozen Gemma-4-12B model moves from 80.0% to 93.3% after grafting a verified solution library, surpassing its own 77.5% and its 31B sibling's 89.2% published anchors.
For recurring cases, eight problems that the base model never solves within a 401,026-token budget are answered from cached verified solutions in just 61 total decode tokens, reducing the token count by a factor of 6,574 and energy usage by about 8,700x; the capability claim rests on held-out transfer (7 of 7 at 31B).
The same byte-exact store widens usable context from 32,768 to 2,854,766 tokens at zero extra accelerator memory, and remains byte-identical between machines of the same architecture.
We describe the system at the behavior level; the engine is proprietary, and every reported number is backed by committed input and output hashes for re-checking the scoring without the engine.
Blogger's Review: This byte-exact KV-cache grafting technology significantly enhances model performance and reduces energy consumption without altering model weights, showcasing the potential for efficiently utilizing verified knowledge. Further exploration in practical applications is warranted.