This Sunday on stage for agisummit</p>\n","updatedAt":"2026-07-17T02:28:52.828Z","author":{"_id":"67fcc5ca7f552d4eb0162533","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67fcc5ca7f552d4eb0162533/sqFNvSpyuvKmzgUaxayfd.jpeg","fullname":"sietse schelpe","name":"Corbenic","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8709107637405396},"editors":["Corbenic"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/67fcc5ca7f552d4eb0162533/sqFNvSpyuvKmzgUaxayfd.jpeg"],"reactions":[{"reaction":"👀","users":["JoeMama123123123"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.14431","authors":[{"_id":"6a5992dc6c2e371e6ca380ab","name":"Sietse Schelpe","hidden":false}],"publishedAt":"2026-07-15T00:00:00.000Z","submittedOnDailyAt":"2026-07-17T00:00:00.000Z","title":"Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel","submittedOnDailyBy":{"_id":"67fcc5ca7f552d4eb0162533","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67fcc5ca7f552d4eb0162533/sqFNvSpyuvKmzgUaxayfd.jpeg","isPro":true,"fullname":"sietse schelpe","user":"Corbenic","type":"user","name":"Corbenic"},"summary":"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.","upvotes":4,"discussionId":"6a5992dc6c2e371e6ca380ac","projectPage":"https://www.corbenic.ai","organization":{"_id":"6a1d1018e87b112ed3ab159d","name":"Corbenci","fullname":"Corbenic","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67fcc5ca7f552d4eb0162533/JMN8IDsOLPOKK8t1Q6Rlz.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"67fcc5ca7f552d4eb0162533","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67fcc5ca7f552d4eb0162533/sqFNvSpyuvKmzgUaxayfd.jpeg","isPro":true,"fullname":"sietse schelpe","user":"Corbenic","type":"user"},{"_id":"69cd2a62eb06b6706a968327","avatarUrl":"/avatars/c12363a509aaf1d3785d5fe251a33d67.svg","isPro":false,"fullname":"周子豪","user":"JAXONBAKER","type":"user"},{"_id":"6a2da6c8ca070ee12c6e396c","avatarUrl":"/avatars/0355287dcabaa67dbc7f0b10b87451f9.svg","isPro":false,"fullname":"Joe Mama","user":"JoeMama123123123","type":"user"},{"_id":"6984dafe6eae8e517dd6d8e5","avatarUrl":"/avatars/ec90bdf2f88c6cfb34352e4e3121500f.svg","isPro":false,"fullname":"Emily Smith","user":"nothary","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6a1d1018e87b112ed3ab159d","name":"Corbenci","fullname":"Corbenic","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67fcc5ca7f552d4eb0162533/JMN8IDsOLPOKK8t1Q6Rlz.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.14431.md","query":{}}">
Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel
Abstract
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.
Community
This Sunday on stage for agisummit
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2607.14431 in a model README.md to link it from this page.
Cite arxiv.org/abs/2607.14431 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.14431 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.