[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-paw-program-as-weights":3,"news-related-c53b958f-dde3-450f-86d4-9024621391e3":31},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":24,"published_at":25,"created_at":26,"modified_at":27,"is_published":28,"publish_type":29,"image_url":13,"view_count":30},"c53b958f-dde3-450f-86d4-9024621391e3","PAW 把 LLM 变成「编译器」：0.6B 小模型在 MacBook 上跑出 32B 性能","arXiv 2607.02512 昨天发布了一项值得所有关注本地 LLM 部署的工程师注意的工作：Program-as-Weights（PAW），把「用 LLM 解决模糊函数」从「按次调用 API」变成「一次性编译、本地复用」的新范式。\n\n团队提出的 fuzzy-function programming 针对一类常见痛点：日志告警判断、JSON 修复、搜索结果重排——这些规则无法写死、又不值得为每次调用付出 GPT-4\u002FClaude 级别的成本和延迟。PAW 的思路是反向操作：把自然语言描述的「模糊函数」编译成参数高效 adapter，挂在一个 0.6B 的 Qwen3 解释器上。\n\n最关键的一组数字：0.6B 解释器 + 编译出的 adapter，匹配了直接 prompt Qwen3-32B 的性能，但推理内存只有 1\u002F50，并且在 MacBook M3 上跑出 30 tok\u002Fs。架构是 4B 编译 + 0.6B 解释的「分而治之」——4B 编译器训练在 1000 万样本的 FuzzyBench 上，一次性把函数定义烧成 adapter，0.6B 解释器冻结权重只负责执行。LLM 的角色由此从「逐题求解器」变成「工具构造器」：每个函数只需付一次编译开销，之后任意调用都是本地、离线、可复现的。\n\nPAW 的真正意义不在「又一个小模型」，而在于把 LLM 的「一次性智慧」沉淀为「可重用资产」。对边缘部署、隐私敏感场景（医疗、法律日志）、成本敏感业务尤其有价值。当然限制也明确：每个函数仍需先编译一次，且 adapter 质量受 4B 编译器和 FuzzyBench 覆盖范围制约。代码与数据集已开源。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.02512","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[],"paw-program-as-weights","2026-07-03T04:00:00Z","2026-07-03T04:10:33.371163Z","2026-08-19T02:08:40.142862Z",true,"agent",128,{"items":32},[33,38,43,48,53,58],{"id":34,"title":35,"news_slug":36,"published_at":37},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":39,"title":40,"news_slug":41,"published_at":42},"fdbe1ee2-131c-4632-baf6-03109d7c1814","Qwen3.8-Next 架构论文:125B 参数 6B 激活,1\u002F9 训练 FLOPs 对标 397B 前辈","qwen3-8-flash-next-architecture","2026-09-01T23:15:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"86c380ed-bdb5-47d0-bf9a-3c55f8573d61","on-policy 蒸馏真的在蒸馏吗?普渡论文:固定负优势就能追平教师","on-policy-distillation-teacher-free-opsa","2026-09-01T15:05:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"a9832c84-4903-4d3a-93a4-e0768bcf69b2","SkewAdam 把 MoE LLM 的 optimizer state 砍到 2.6%：让 6.78B 模型塞进 40 GB 显卡","skewadam-moe-optimizer-state","2026-07-23T00:10:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"b362eb89-32ef-46ed-b65a-dd65f6f305f2","Direct-OPD 把「RL 经验」跨模型规模可复用：字节×清华让弱模型的策略差当强模型的隐式奖励","direct-opd-rl-experience-transfer","2026-07-14T14:10:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"09adc53e-c559-4899-bc68-117b19b717a3","单层 Transformer 就能打平全参数 RL 后训练:Qwen 上的「中段层集中增益」现象","single-layer-transformer-rl","2026-07-03T08:00:00+00:00"]