Search papers, labs, and topics across Lattice.
This paper introduces a method for enhancing the capabilities of a frozen small language model by employing byte-exact key-value (KV) cache grafting, allowing for the integration of verified knowledge without altering model weights. The approach achieves bit-exact restoration of logits, demonstrating zero KL divergence and 100% argmax agreement across multiple scales and GPU targets. Notably, the method significantly improves performance on specific tasks while drastically reducing energy consumption and increasing usable context length, showcasing a transformative leap in efficiency and capability for smaller models.
A frozen 12B model can leap from 80% to 93.3% accuracy by leveraging byte-exact KV-cache grafting, achieving unprecedented efficiency with 8,700x less energy consumption.
We report a way to make a frozen small language model both more capable and dramatically cheaper at once, without changing any weights. Verified knowledge is deposited once as a byte-exact key-value (KV) state artifact and later restored, by graft, into a fresh inference context. The restore 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 show that own-position graft is the unique numerically exact operating point on a model with floating-point rotary encoding, and we verify byte-exactness on two model scales (12B, 31B) and two GPU targets, one through a pre-registered replay. On AIME 2025, a frozen Gemma-4-12B moves from 80.0% to 93.3% once a verified solution library is grafted, above its own 77.5% and its 31B sibling's 89.2% published anchors. On the recurring case, eight problems the base model never solves within a 401,026-token budget are answered from cached verified solutions in 61 total decode tokens, a factor of 6,574 fewer tokens and about 8,700x less energy; the capability claim proper 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 moves 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 so the scoring can be re-checked without it.
