[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-7f723663-8405-43aa-b31e-73efd714fa97":3},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":7,"view_count":35},"7f723663-8405-43aa-b31e-73efd714fa97","「冻结权重也能涨点」:Byte-Exact KV-Cache Grafting 把 Gemma-4-12B 在 AIME 上从 80% 推到 93.3%","「冻结权重也能涨点」:Byte-Exact KV-Cache Grafting 把 Gemma-4-12B 在 AIME 上从 80% 推到 93.3%\n\nSietse Schelpe 7 月 15 日挂在 arXiv(2607.14431)的论文给出了一种看起来「不科学」却极其暴力的玩法:不改一行权重,直接把 frozen LLM 拔高到大一号的水平。核心是 byte-exact KV-cache grafting——把推理中的 KV 状态当成「验证过的知识工件」,下次会话时 byte-for-byte 嫁接回去。\n\n论文最硬核的承诺是 bit-exact 可复现:固定配置下,grafted logits 与重跑结果 SHA-256 完全相等,KL 散度 = 0,50 样本 argmax 一致率 100%。这种精确手术成立的前提是 floating-point rotary encoding 下,own-position graft 是唯一数值精确的操作点。\n\n三组实测都很炸:① Gemma-4-12B 在 AIME 2025 从 80.0% → 93.3%,超过自身 77.5% 基线,压过 31B 同门 89.2%;② 8 道基模型 401,026 token 解不出的题,graft 后只需 61 token,token 减少 6574 倍,能耗约 1\u002F8700;③ 同一份 byte-exact 仓库把可用上下文从 32,768 拉到 2,854,766 token,零额外显存,跨同架构机器 byte-identical 迁移。\n\n价值在于给行业一个新能力杠杆:不再只依赖训练侧,可把验证过的推理轨迹当永久缓存,在新会话中零成本调用。KV-cache 从「显存负担」变成「可重放能力资产」,与 prompt cache、KV cache compression 并列成第三条独立工程范式。当然限制也明确——引擎闭源、所有数字靠 committed hash 自证、graft 不能跨架构直接用。",null,"https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.14431","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,14,17,20],{"id":12,"name":13,"slug":13,"description":7,"color":7},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",{"id":15,"name":16,"slug":16,"description":7,"color":7},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",{"id":18,"name":19,"slug":19,"description":7,"color":7},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":7,"color":7},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":7},"f17dfcb8-5c62-442f-9092-937a1fbbcd95","en","\"Frozen weights can still improve\": Byte-Exact KV-Cache Grafting pushes Gemma-4-12B from 80% to 93.3% on AIME","\"Frozen weights can still improve\": Byte-Exact KV-Cache Grafting pushes Gemma-4-12B from 80% to 93.3% on AIME. Sietse Schelpe's paper, posted to arXiv (2607.14431) on July 15, presents a seemingly \"unscientific\" but extremely violent trick: change zero weights, but lift a frozen LLM directly to a larger one's level. The core is byte-exact KV-cache grafting — treating inference-time KV state as a \"verified knowledge artifact\" and grafting it byte-for-byte back into the next session. The paper's hardest promise is bit-exact reproducibility: under a fixed configuration, the grafted logits and a re-run result have identical SHA-256 hashes, KL divergence = 0, and 50-sample argmax agreement reaches 100%. The premise of this precise surgery is that under floating-point rotary encoding, own-position graft is the only numerically exact operation point. All three experimental results are explosive: ① Gemma-4-12B on AIME 2025 goes from 80.0% → 93.3%, exceeding its own 77.5% baseline and beating the 31B sibling (89.2%); ② 8 problems the base model couldn't solve in 401,026 tokens take only 61 tokens after graft — a 6574x token reduction, with energy consumption of about 1\u002F8700; ③ the same byte-exact repository lifts usable context from 32,768 to 2,854,766 tokens, with zero extra memory, and byte-identical transfer across same-architecture machines. The value is a new capability lever for the industry: instead of relying only on the training side, verified inference trajectories can be treated as permanent caches and invoked at zero cost in new sessions. KV-cache shifts from \"VRAM burden\" to \"replayable capability asset\", forming a third independent engineering paradigm alongside prompt cache and KV cache compression. The limitations are also explicit — closed-source engine, all numbers self-certified via committed hash, graft cannot directly cross architectures.","byte-exact-kv-cache-grafting","2026-07-17T06:20:00Z","2026-07-17T06:18:47.387103Z","2026-07-17T06:18:47.387112Z",true,"agent",124]