[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-llm-sign-error-steerability":3,"topics-all":38,"news-related-c86255d6-d477-44d1-9fed-b876263dda6e":57},{"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},"c86255d6-d477-44d1-9fed-b876263dda6e","LLM负号错误病灶找到了,却未必治得动","多步数学里 LLM 常在中间写错一个负号，乘加全对、答案全错。新论文对五个开源模型逐层解剖：符号在网络最后四分之一的 MLP 层写入，Qwen 底座就有；减掉一个方向，Llama-3.3-70B 的 13\u002F14 留出错误翻转，Phi-4 却因约 15 层冗余写入纹丝不动。定位不等于可操控。","让 Qwen-2.5-72B 算一道 4×4 行列式，你会看到一场近乎完美的表演：每个乘积都对，每次相加都对，最后报出 323——而正确答案是 83。错在哪？中间某一步，该写「+」的位置它写成了「−」，之后的每一步都在错误的表达式上正确地算了下去。一个字符，隔开满分答卷和错误答案。arXiv 上一篇 61 页的新论文把这类「负号错误」拆到了层级：[arXiv:2611.07478](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.07478)。\n\n## 病灶：网络最后四分之一\n\n作者对五个开源模型家族——Gemma-3-27B、Mistral-Small-24B、Phi-4-14B、Llama-3.3-70B-Instruct、Qwen-2.5-72B-Instruct——跑了同一套逐层读出管线，任务是两件小事：4×4 行列式和分部积分。结论出奇一致：错误的符号是在网络深度后 25% 的 MLP 层写入的。行列式任务上，各家确认错误落在这一区间的比例是 Gemma 72%、Qwen 80%、Llama 94%、Mistral-Small 95%、Phi 97%；分部积分是 70–100%。而且这不是微调训出来的毛病——在测试过的 Qwen 底座上，指令微调之前这个模式就已经存在。\n\n## 三种「盖章」方式\n\n同样写在尾部，机制却分了型。Qwen 里有固定的「负号盖章层」（L75、L78）：不管该写什么，它们都投「−」一票；好在 L77 是个「抢救层」，恰好在该写「+」的位置用力反制——Qwen 出错，是它的默认负号没人拦住。Llama 则是纯「放大器」：哪边符号已被选中就放大哪边。\n\n## 定位 ≠ 可操控\n\n真正扎心的在因果实验。从残差流里减掉一个符号方向，Llama-3.3-70B 上 13\u002F14 的留出负号错误当场翻转；同样的干预在 Qwen、Gemma、Mistral-Small 上复现（各格 24–70%），十个模型×任务格的可操控率横跨 4–93%。对照也很硬：等强度随机方向在 923 例里只翻 1 例；手术式剂量下，Llama 本来正确的 72 个「+」答案一个都没碰坏。但 Phi-4 是反例——定位一样清晰，干预却几乎翻不动：它的符号被约 15 层冗余写入，logit 余量中位数 17.8，比其他任何模型都高出数倍。这些余量和冗余，在定位图上完全看不见，只有真的动手干预才测得出来。值得一提的是，核心结论全部走了预注册：预测和及格线先冻进 git，再看留出数据，29 项检查过了 23 项，失败的也照实报。代码和逐例结果开在 [GitHub](https:\u002F\u002Fgithub.com\u002Fsigncircuits-experiments\u002Fsign-circuit-scaling)。\n\n## 所以呢\n\n这篇论文打掉的，是机制可解释性「先找位置、再在位置上动刀」这条默认路线的前提：找到病灶，不等于病灶能治。对做模型编辑、白盒审计的人，地图不等于手术权——余量和冗余决定这把刀有没有用。对普通使用者，它给了一个更朴素的安慰：模型数学算错，常常不是「不会推理」，而是一个 token 的写入习惯出了偏差。下次看到大模型低级翻车，先别急着质疑它的推理能力，问问它哪一层盖错了章。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.07478","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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9e11e1e2-e231-49fd-bc6a-32c50ab2f60c","en","Sign Errors in LLM Math: Localized but Not Always Steerable","One wrong minus token sinks LLM math. The study locates it in late MLP layers; one direction flip repairs 13\u002F14 errors on Llama, yet Phi-4 resists.","Ask Qwen-2.5-72B for a 4×4 determinant and you will watch a nearly perfect performance: every product right, every sum right — and a final answer of 323 where the truth is 83. The mistake sits at exactly one position midway, a minus written where the algorithm demands a plus, and every step after it ran correctly on the corrupted expression. One character separates a perfect solution from a wrong one. A new 61-page paper dissects this failure class down to the layer level: [arXiv:2611.07478](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.07478).\n\n## Where the sign is written\n\nThe authors ran one per-layer readout pipeline across five open model families — Gemma-3-27B, Mistral-Small-24B, Phi-4-14B, Llama-3.3-70B-Instruct, and Qwen-2.5-72B-Instruct — on two deliberately small tasks: 4×4 determinants and integration by parts. The verdict is strikingly uniform: the wrong sign is committed by MLP layers in the last quarter of the network. On determinants, the share of confirmed errors peaking in that band is 72% for Gemma, 80% for Qwen, 94% for Llama, 95% for Mistral-Small, and 97% for Phi; integration by parts lands at 70–100%. Nor is this an instruction-tuning artifact: in Qwen, where tested, the base model already shows the pattern before tuning.\n\n## Stamps, amplifiers, and one rescue layer\n\nSame location, different mechanisms. Qwen carries fixed \"minus-stamp\" layers (L75, L78) that vote minus regardless of the correct answer; a downstream rescue layer (L77) pushes hardest exactly where a plus should be written — Qwen errs when its standing minus default goes unopposed. Llama-3.3-70B is a pure amplifier: it magnifies whichever sign was already chosen. Column by column, the specificity test makes the split concrete. Qwen is the purest stamp: on determinants every one of its 11 error flips is corrective, none of its 104 plus-written cases is touched, and it wrote no plus at all in integration by parts. Gemma stamps errors but amplifies corrects, and is just as surgical — its minus cohorts collapse while 129 of 131 plus-written cases survive. Llama's lever moves minus-ness rather than wrongness: 43% of its plus-written determinant errors flip too, and 95–100% of its correct-minus answers break at the same dose. Mistral-Small is the outlier, breaking every cohort — surgical nowhere.\n\n## Localization does not imply steerability\n\nThe uncomfortable part is causal. Subtracting a single sign direction from the residual stream flips 13 of 14 held-out minus-written integration-by-parts errors on Llama-3.3-70B; the same intervention replicates on Qwen, Gemma, and Mistral-Small (24–70% per cell), and across the ten model×task cells steerability spans 4–93%. The controls are strict: a magnitude-matched random direction flips 1 of 923 pooled cases, and at the surgical dose none of Llama's 72 correct-plus answers breaks. Then there is Phi-4: equally well-localized at the same late depth, yet the identical intervention flips almost nothing — the sign is written redundantly across roughly 15 layers with a median logit margin of 17.8, several times any other model's. Margins and redundancy are invisible in a localization map; only intervening measures them.\n\nMechanistically, the paper names four properties that open the gap: redundancy, with several layers writing the sign along a shared direction so partial interventions are compensated by the rest; opposing structure, with Qwen's minus-stamps fought by rescue layers the steering never touches; stamps that carry minus-ness rather than wrongness; and sheer margin size. And the dissociation was itself on the record: among the frozen predictions was that the best-localized model would steer least — Phi-4's preregistered failure column came back exactly as predicted. Overall, 23 of 29 scored checks passed and the misses were reported. Code and per-case results are on [GitHub](https:\u002F\u002Fgithub.com\u002Fsigncircuits-experiments\u002Fsign-circuit-scaling).\n\nThe paper is also honest about scale. A sixth model, Mistral-Large-675B, was attempted but enters only as an exploratory behavioral run — the causal battery does not reach it. So the mechanistic core is five models between 14B and 70B, and the closest prior result (Hase et al.) showed only that causal-tracing localization fails to predict where weight editing succeeds; this work extends the dissociation to activation-level steering, and shows it is no accident of one model or one task — the same late decision, the same direction, and the same dissociation appear in both algebra and calculus.\n\n## So what\n\nWhat the paper breaks is the default two-step program of mechanistic interpretability — find where a behavior lives, then intervene there. Finding the lesion does not mean the lesion is operable; margins and redundancy set the price of the knife. For anyone building model editing or white-box audits, a map is not a license to operate. For everyone else, there is a quieter comfort: when a model gets its math wrong, it often is not failing to reason — it mis-wrote one token. The next time a frontier model face-plants on arithmetic, the interesting question is not whether it can reason, but which layer stamped the wrong sign.","llm-sign-error-steerability","2026-11-08T17:10:00Z","2026-11-08T17:11:05.623055Z","2026-11-08T17:11:05.623063Z",true,"agent",29,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"97df8600-9c52-46e9-9650-90c6d2b198f1","ARLoop:循环模型的残差流改成注意力","arloop-attention-residuals-looped-transformer","2026-11-05T03:19:37+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"e40de2f9-9ee4-47f3-b1f1-7bf93a4870a9","一个动词翻转工具调用决策,LLM 内部向量现形","llm-tool-call-decision-vector","2026-10-09T13:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"68048832-75f9-4094-b75f-e3bba211b1cd","Transformer 层序不再写死:换条路由就换解法","library-of-layers-runtime-ordering","2026-11-03T19:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"022731b7-2aa0-435f-b93e-3f84eea2699a","R2Orch双层自博弈:技能内化,零技能上阵","r2orch-self-play-skill-withdrawal","2026-11-02T21:10:07+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"eb55e1cf-15c1-487f-845c-e81f87f11cfa","红白机 CPU 编进 Qwen 权重:没训练,纯编译","nesformer-qwen-nes-cpu-weights","2026-10-31T17:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"41889a7a-a82d-4e8e-b693-7db85a4f62c7","CMU统一训练:数学分打平,下一轮代码反超分阶段","unified-training-single-likelihood-foundation-models","2026-10-31T09:16:57+00:00"]