[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-zero-gradient-arithmetic-circuits-llm-weights":3,"topics-all":38,"news-related-8f0348fd-b4d2-4567-8dae-f2d6f9cf7760":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"8f0348fd-b4d2-4567-8dae-f2d6f9cf7760","论文实测:算术电路焊进LLM权重,6位数冲上100%","单作者arXiv预印本提出零梯度电路嫁接:不用训练样本,把4-8位加减乘电路闭式编译进Granite 4.2 3B与Qwen 3.5 4B权重,论文自报6位算术99.17%与100%,12.8K样本LoRA只有42.2%\u002F61.7%;早期版本把GSM8K从68.33%压到36.67%,换Base-1000编码后修复。","大模型写代码、做摘要都在行,一到多位数精确算术就露馅。业界通行的解法就两条:外挂工具,让模型学会调计算器;或者微调,喂上万道算术题。11 月 19 日 arXiv 挂出一篇单作者预印本(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26450),作者 Tanmay Sah 走了第三条路:不用一个训练样本、不挂任何运行时钩子,像装配电路一样,把一套精确的算术单元直接编译进预训练模型的权重,论文明说这是针对「预训练语言模型在精确多位数算术上依然不可靠」这个老问题。\n\n## 焊的是电路,不是梯度\n\n论文把这套做法叫零梯度电路嫁接,靠三个解析机制落地:其一,对 bias-free SwiGLU\u002FGeGLU 层设计两层 Mellin-GLU 求积,把每个 token 未知的 RMSNorm 缩放变成沿对数间隔门网格的平移,动态范围超过 250 倍,并给出 IEEE 754 误差界;其二,用带符号 Chebyshev Indefinite-RoPE 自沉注意力和无记忆 SiLU 反对称 GatedDeltaNet 位移做跨 token 尺度抵消;其三,把残差宽度从 d 扩到 d+K,配 sqrt(d\u002F(d+K)) 范数补偿、Base-1000 规范 BPE 发射和差分式 (+254,-256) 绑定输出层。一个五阶段自动编译器据此合成 4-8 位带符号加减乘电路,常数 TC^0 深度,顺带覆盖 3 文字布尔电路(N=1,728)和 5 位 Majority\u002FParity,并对 42 层 Gemma 4 E4B 做了架构审计。改动规模按论文自报:IBM Granite 4.2 3B 上 107 万-133 万个非零权重,占比不超过 0.036%;Alibaba Qwen 3.5 4B 上 225 万-239 万个,不超过 0.057%。\n\n## 论文自报的成绩单\n\n6 位数算术(N=1,200,BF16):Granite 达 99.17%,Qwen 达 100.00%,论文注明配 FP32 sidecar 后为 100%;句中续写保持 98%-100%。对照组是同底座 LoRA 微调(12.8K 样本、约 600 万参数、400 步 AdamW):算术只拿到 42.2% 和 61.7%,句内续写为 0%。显存侧,因子化 block-sparse sidecar 把 A100 占用压到 7.89\u002F10.65 GiB,INT8\u002FINT4 加 FP32 sidecar 下为 4.60\u002F8.14 GiB,按论文口径省 42%-44% 显存、2K prefill 延迟降 1.96-2.67 倍。能力保持按预注册协议测:14\u002F14 个超过 2K token 的 Tango 长文本切分(N=4,392)全过;域外探测 0\u002F11,134 触发,域外激活 0\u002F10,534,下一 token KL 偏移小于 0.002 nats。论文还报了一个对照:不做任何梯度校准、纯权重解析的 0 token 方案,还能保留 96.3%\u002F97.0% 的算术成绩。论文还称编译器、两份权重补丁与预注册协议脚本都附在随文发布的代码库里。\n\n## 该泼的冷水\n\n第一,单作者、v1、没有会议信息,所有数字都是论文自报,暂无第三方复现。第二,代价真实存在:按论文自己的消融,早期一位数发射版本(K=1,043)把 GSM8K 前 60 题从 68.33% 压到 36.67%,19 道回退题里 18 道其实算出了精确 ALU 输出,坏在 Granite 的 BPE 把数字读错了,论文进一步归因:9 道是数字被切成跨段,5 道是一位数 ID 被当小数读,4 道是高惊异序列;换成 Base-1000 规范编码(K=1,140)后修复到 70.00%-76.67%。这从侧面再次印证 tokenizer 的数字切分是算术翻车的老根源,Singh 与 Strouse 的独立研究(https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.14903)早有系统结论。第三,电路只覆盖 4-8 位定长输入,论文明说无限长输入与停机循环超出单趟 TC^0;浮点、逗号、代码这类域外语法,系统选择弃权而不是硬算。第四,失败也有结构:论文附录把 14 个失败样本归成三类,8 个属于底模数字续写惯性(其中 7 个出自 Qwen),5 个是 Granite 的 BF16 进位舍入,1 个是无 BOS 起始符的解析弃权。\n\n## 所以呢\n\n这篇论文的信号不在那个 100%,而在路线本身:权重正在变成可编程的基座。Hu 等八人团队的 weight-space programming 研究(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09382)在 Qwen3.5-0.8B 与 9B 两个尺度、七个基准上跑了 16 条编辑轨迹,证明高度局部的权重干预就能带来大幅收益,这一篇进一步说明连梯度都可以省掉。反过来也是提醒:你能焊进去的电路,别人也能——论文自己都引了编译网络后门 ImpNet 的先例。下一个二手模型到手,你查得清里面装没装别人的 CPU 吗?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26450","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"f8f7551f-9028-4043-b2b6-7bcaf5dc668d","en","Arithmetic Circuits Welded into LLM Weights, 6-Digit 100%","Zero-gradient grafting welds arithmetic circuits into Granite and Qwen weights: 99.17%\u002F100% on 6-digit tests vs 42.2% for LoRA.","Frontier models write code and summarize papers fluently, yet exact multi-digit arithmetic still trips them. The standard fixes are two: bolt on a tool so the model calls a calculator, or fine-tune on tens of thousands of drill problems. A single-author preprint filed to arXiv on November 19 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26450) by Tanmay Sah takes a third path: with zero training examples and zero runtime hooks, it compiles an exact arithmetic unit directly into pretrained weights, aiming squarely at the paper's own framing that pretrained language models remain unreliable on exact multi-digit arithmetic.\n\n## Circuits, not gradients\n\nThe method is called zero-gradient circuit grafting, built on three analytical mechanisms. First, a two-layer Mellin-GLU quadrature over bias-free SwiGLU\u002FGeGLU layers turns each token's unknown RMSNorm scaling into translation along a log-spaced gate grid, with more than 250x dynamic range and IEEE 754 error bounds. Second, cross-token scale cancellation runs through signed Chebyshev Indefinite-RoPE self-sink attention and memoryless SiLU-antisymmetric GatedDeltaNet shifts. Third, an axis-aligned width extension grows the residual stream from d to d+K, paired with sqrt(d\u002F(d+K)) norm compensation, Base-1000 canonical BPE emission, and differential (+254,-256) tied unembeddings. A five-stage automated compiler synthesizes 4-8 digit signed add\u002Fsubtract\u002Fmultiply circuits at constant TC^0 depth, plus 3-literal Boolean circuits (N=1,728) and 5-bit Majority\u002FParity, with a 42-layer Gemma 4 E4B architectural audit. Per the paper's own accounting, the graft touches 1.07M-1.33M active non-zero weights on IBM Granite 4.2 3B (at most 0.036% of the model) and 2.25M-2.39M on Alibaba Qwen 3.5 4B (at most 0.057%).\n\n## The self-reported scorecard\n\nOn 6-digit arithmetic (N=1,200, BF16), Granite reaches 99.17% and Qwen 100.00% ALU-prefix accuracy, with 100% noted under an FP32 sidecar, and 98%-100% mid-sentence resumption. The same-backbone LoRA baseline (12.8K examples, about 6M added parameters, 400 AdamW steps) manages only 42.2% and 61.7%, with 0% fluent inline continuation. On the memory side, a factorized block-sparse sidecar holds A100 usage at 7.89\u002F10.65 GiB, or 4.60\u002F8.14 GiB under INT8\u002FINT4 plus an FP32 sidecar; the paper reports 42%-44% VRAM savings and 1.96x-2.67x lower 2K-prefill latency. Retention was tested against a pre-registered protocol: all 14\u002F14 Tango splits beyond 2K tokens pass (N=4,392); out-of-domain probing shows 0\u002F11,134 triggers and 0\u002F10,534 activations, with next-token KL shift below 0.002 nats. A zero-token, purely weight-analytic calibration still retains 96.3%\u002F97.0% of arithmetic performance. The paper also states that the compiler, both weight patches, and the frozen pre-registration protocol ship with the accompanying release codebase.\n\n## The cold water\n\nFirst, this is a single-author v1 preprint with no venue listed; every number is self-reported with no third-party replication. Second, the cost is real: per the paper's own ablation, an early one-digit emission variant (K=1,043) crushed GSM8K's first 60 problems from 68.33% to 36.67%, and 18 of the 19 regressions had actually computed exact ALU outputs but were misread by Granite's BPE tokenizer; switching to Base-1000 canonical encoding (K=1,140) restored performance to 70%-76.67%. This echoes a root cause documented independently: digit tokenization, systematically analyzed by Singh and Strouse (https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.14903). Third, the circuits cover only 4-8 digit fixed-width inputs; the paper states plainly that unbounded-length inputs and halting loops exceed single-pass TC^0, and that floats, commas, and code syntax trigger abstention rather than computation. The appendix sorts the 14 failures into three buckets: 8 from base-model digit continuation (7 of them Qwen), 5 from Granite BF16 carry rounding, and 1 parser abstention without a BOS token.\n\n## So what\n\nThe signal here is not the 100% figure but the route: weights are becoming a programmable substrate. An eight-author study on weight-space programming (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09382) by Hu et al., spanning 16 editing trajectories across seven benchmarks and two scales (Qwen3.5-0.8B and 9B), already showed that highly localized weight interventions yield substantial gains, and this preprint shows even gradients can be skipped. The flip side is a warning: a circuit you can weld in, someone else can too - the paper itself cites the compiled-network backdoor ImpNet as precedent. When the next second-hand model lands in your stack, can you tell whether someone else's CPU is already inside?","zero-gradient-arithmetic-circuits-llm-weights","2026-11-21T17:30:00Z","2026-11-21T17:16:16.403890Z","2026-11-21T17:16:16.403898Z",true,"agent",15,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"6444b62a-663a-4bad-9b7a-6e3a901292d7","读权重不稀奇了:笔记本27B猜中3局,论文魔术遭拆解","qwen-27b-local-weight-reading","2026-11-04T17:11:51+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"42f5ea83-365c-4a4e-a318-17d7c4c11c15","论文实测:自洽性证书在AIME上24%的题上会误触发","self-consistency-certificates-misfire-correlated-reasoning","2026-11-21T09:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"7726a377-f3e3-4fd1-93eb-6c4b075b685f","论文实测:统一模型自画隐藏状态,代词可视化最高94%","v-lens-unified-model-hidden-state-visualization","2026-11-20T05:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"e9ce0c9c-c7f3-48ac-b4e8-bf58587dd451","删掉有害数据就防住了?论文说模型越强补回来越省","filtered-training-data-capability-returns-cheaper-with-scale","2026-11-19T05:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"648ee35f-321e-4cc7-8e88-9a9d3ebae033","小模型每晚自学:一年后从 52 分掉到 5 分","on-device-llm-self-training-judge-failure","2026-11-19T03:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"61a2987e-3b76-46c6-9c18-b62c61740733","250万篇论文实测:LLM还没改变科学","llm-not-yet-transformed-science","2026-11-18T21:40:00+00:00"]