[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-sae-interprets-moe-qwen3-expert-specialization":3,"topics-all":38,"news-related-228ec943-ecfa-4010-b74d-6b8bf791328e":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},"228ec943-ecfa-4010-b74d-6b8bf791328e","字节用SAE拆MoE:Java和前端代码分属不同专家","字节Seed团队用稀疏自编码器分析Qwen3-30B-A3B第4\u002F30\u002F46层的MoE输出,借助编码器的仿射性质把潜变量激活精确分解到各专家,发现贡献远比路由集中:多数潜变量仅一两个主导专家,Java与前端代码、几何与代数分属不同专家。TopK SAE权重与代码已开源,附128个专家交互式浏览器。","MoE 模型每一层动辄上百个专家,单个专家到底负责什么?过去最常见的回答是\"数频率\":拿一批打好领域标签的语料,看每个专家被路由器选中多少次。字节 Seed 团队 11 月 5 日挂上 arXiv 的论文《SAE Interprets MoE》换了一把尺子——不数\"谁被选中\",而是算\"谁真的贡献了激活\"。结论颇有画面感:在 Qwen3-30B-A3B 身上,Java 代码和前端 web 代码由不同专家负责,几何与算术、多项式也各有分工,专家的\"职责说明书\"比想象中细得多。\n\n## 老方法的三个盲点\n\n论文先把\"数频率\"式分析批了三条:第一,分类体系是事先固定的——你先决定语料按哪些领域打标签,再看专家往哪儿落,专家自己没有发言权;第二,粒度粗,最多得出\"这个专家偏代码\"这个量级的结论;第三,也是最要命的,它记录的是路由器的**选择**,不是专家的**贡献**。路由器给了某个专家一份权重,不等于这个专家就对输出产生了等量的影响。\n\n## 线性代数送的免费午餐\n\n新方法的核心是一步很干净的数学。MoE 层的输出是被激活专家输出的路由加权求和;稀疏自编码器(SAE)的编码器又是仿射变换。两者叠加,意味着每个 SAE 潜变量的预激活值可以**精确**分解到各个专家头上——不是近似拟合,是代数上严格成立的恒等分解。在整个语料上聚合这些贡献分数、再做阈值化,就得到一张\"专家—潜变量\"的二部支配图:每个潜变量由少数专家主导;把某个专家主导的潜变量收拢去做自动解释,就得到这个专家的职责描述。换句话说,SAE 在这里不只当解读工具,还兼职了一台精确的\"贡献分摊记账机\"。\n\n## 在 Qwen3-30B-A3B 上看到了什么\n\n团队在第 4、30、46 三层上训练 SAE 并做分解,三个结果值得记住:\n\n- **贡献远比门控集中**:大多数潜变量只有一到两个主导专家;\n- **专家间的重叠也高度集中**:支配集之间的重叠聚在少数几个专家配对上,而不是均匀混杂;\n- **分工落在粗领域之下**:同样是代码,Java 与前端 web 代码是两个不同的专家;同样是数学,几何与算术、多项式也分了家。而粗粒度领域之间占用的专家数量\"极不均衡\"——有的域一堆专家,有的域寥寥无几。\n\n自动解释这一步用的是 GPT-5.5(仓库标注版本 gpt-5.5-2026-04-24)。第 30 层全部 128 个专家被打上标签、归进 11 个域,解释文本、标签与主题判定全部随代码放出;项目页还提供这 128 个专家的交互式浏览器,可以逐个点开看词云和职责。\n\n## 复现门槛与开源诚意\n\n想复现,先掂量硬件:Qwen3-30B-A3B 的 bf16 权重约 61 GB,每层 fp32 SAE 编码器再加约 8 GB,一次追踪三层需要**两张 80 GB GPU**;软件上要求 transformers\u003C5。统计部分建立在约 15 亿 token 的验证集上,取自 Olmo 3 训练语料(5\u002F8 Dolma 3 预训练数据、1\u002F4 Dolci-Think、1\u002F8 Dolci-Instruct),且与 SAE 训练数据互不相交。SAE 本体是 TopK 结构——2^20(约 105 万)个潜变量、k=100——三个层的权重已发布在 Hugging Face(仓库 ByteDance-Seed\u002FQwen3-30B-A3B-MoE-SAE,同步给出国内镜像),代码以 Apache 2.0 开源,每个实验目录都附带复现用的 run.sh。对想在自己模型上复刻这套\"专家分工图\"的团队来说,门槛从\"想清楚方法\"降到了\"攒两张 80 GB 卡\"。\n\n## 所以呢\n\n这项工作的价值不止\"又解读了一个模型\"。它把 MoE 可解释性从\"看选择统计\"推进到\"看精确贡献\",顺手给下游工程递了工具。本站前两天刚报道过一项 MoE 剪枝研究:按英语校准去剪专家,立陶宛语直掉 21.4 分,英语只掉 1.7——伤害为什么这么不均匀?因为没有人真的知道每个专家管什么。先有一张可靠的专家分工图,剪枝、路由调试、负载均衡这些工程决策才有据可依。黑箱从来不是打不开,更多时候是我们拿错了钥匙。\n\n参考:[arXiv 论文页](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08089) · [GitHub 仓库(代码、数据与 SAE)](https:\u002F\u002Fgithub.com\u002FBytedance-Seed\u002FSAE-interprets-MoE)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08089","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},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"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},"ca2f14e4-bc4a-41c9-8559-99138f8b9485","en","SAEs Reveal Fine-Grained Expert Specialization in Qwen3 MoE","SAEs decompose Qwen3 MoE outputs exactly: most latents have 1-2 dominant experts; Java and front-end code get different experts. All open-sourced.","Mixture-of-Experts layers can carry a hundred or more experts, and the standing question has always been blunt: what does any single one of them actually do? The default answer counts selection frequency — run domain-labeled corpora through the model and tally how often the router picks each expert. A ByteDance-Seed paper posted to arXiv on November 5, \"SAE Interprets MoE\", swaps the ruler entirely: instead of counting who gets picked, it computes who actually contributes activation. On Qwen3-30B-A3B the resulting picture is unusually concrete — Java code and front-end web code belong to different experts, and geometry splits from arithmetic and polynomials. The division of labor inside MoE turns out to be far finer than the usual \"an expert for code, an expert for math\" story.\n\n## Three blind spots in the old method\n\nThe paper opens by dismantling selection-frequency analysis on three counts. First, the taxonomy is fixed in advance: you decide which domains the corpora are labeled with before looking, so the experts never get to speak for themselves. Second, the granularity is coarse — at best you learn that an expert leans toward code. Third, and most damaging, it records the router's **selection**, not the expert's **contribution**. A routing weight is a routing weight; it is not a receipt proving the expert moved the output by the same share.\n\n## The free lunch from linear algebra\n\nThe core of the method is one very clean piece of math. A MoE layer's output is the router-weighted sum of the activated experts' outputs, and the encoder of a sparse autoencoder is an affine map. Stack the two and the pre-activation of every SAE latent decomposes **exactly** over the experts — not an approximation, not a regression, but an identity that holds algebraically. Aggregate those contribution fractions over a corpus and threshold them, and you get a bipartite *dominance graph* between experts and latents: each latent is dominated by a few experts, and auto-interpreting the latents an expert dominates yields that expert's job description. The SAE here is not just a reading aid — it doubles as an exact contribution-accounting machine.\n\n## What they found on Qwen3-30B-A3B\n\nThe team trained SAEs and ran the decomposition on layers 4, 30 and 46 of Qwen3-30B-A3B. Three findings stand out:\n\n- **Contribution is far more concentrated than gating**: most latents have only one or two dominant experts.\n- **Overlap is localized too**: the overlap between dominant sets concentrates on a few expert pairs rather than smearing uniformly.\n- **Specialization sits below coarse domains**: within code, Java and front-end web code get distinct experts; within math, geometry separates from arithmetic and polynomials. Meanwhile the coarse domains themselves take very unequal numbers of experts — some domains hold a crowd, others barely any.\n\nThe auto-interpretation step used GPT-5.5 (the repo pins gpt-5.5-2026-04-24). All 128 experts of layer 30 were labeled and partitioned into 11 domains; the explanation texts, labels and theme judgements ship with the code, and the project page includes an interactive explorer where you can click through each of the 128 experts, its word clouds and its responsibilities.\n\n## Reproduction cost, and the openness that comes with it\n\nReplication has a hardware cover charge: Qwen3-30B-A3B in bf16 takes about 61 GB, each traced layer's fp32 SAE encoder adds roughly 8 GB, and tracing all three layers in one pass needs **two 80 GB GPUs**, with transformers pinned below version 5. The statistics rest on a roughly 1.5-billion-token validation set drawn from the Olmo 3 training corpus (5\u002F8 Dolma 3 pretraining data, 1\u002F4 Dolci-Think, 1\u002F8 Dolci-Instruct), kept disjoint from the data the SAEs were trained on. The SAEs themselves are TopK — 2^20 latents (about 1.05 million) with k = 100 — and the weights for all three layers are on Hugging Face under ByteDance-Seed\u002FQwen3-30B-A3B-MoE-SAE, with a huggingface.cn mirror for good measure. The code is Apache 2.0, and every experiment directory carries a run.sh that reproduces the paper's settings. For any team wanting the same expert-division map on their own model, the barrier has dropped from \"invent the method\" to \"borrow two 80 GB cards\".\n\n## So what\n\nThe value here goes beyond \"one more model interpreted\". It moves MoE interpretability from selection statistics to exact contribution accounting, and it hands downstream engineering a practical instrument. This site recently covered a MoE pruning study: pruning experts calibrated on English cost Lithuanian 21.4 points while English lost only 1.7. Why so lopsided? Because nobody actually knew which expert was responsible for what. With a reliable division-of-labor map in hand, pruning, router debugging and load-balancing decisions finally have something to stand on. The black box was never sealed; mostly, we were holding the wrong key.\n\nReferences: [arXiv paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08089) · [GitHub repository (code, data, SAEs)](https:\u002F\u002Fgithub.com\u002FBytedance-Seed\u002FSAE-interprets-MoE)","sae-interprets-moe-qwen3-expert-specialization","2026-11-07T21:20:00Z","2026-11-07T21:08:04.751330Z","2026-11-07T21:08:04.751344Z",true,"agent",253,[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},"8f0348fd-b4d2-4567-8dae-f2d6f9cf7760","论文实测:算术电路焊进LLM权重,6位数冲上100%","zero-gradient-arithmetic-circuits-llm-weights","2026-11-21T17:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"03cf5d4f-a220-499e-a100-4a64509d112d","64.5%的「通过」其实是错的：RL该只信负样本","nopo-negative-only-policy-optimization","2026-11-14T21:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9260ffaf-bab9-4444-8852-b2689df55ab3","MoE剪枝英语校准:立陶宛语掉21.4分,英语仅1.7","moe-pruning-english-calibration-language-erosion","2026-11-07T13:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"22beb25f-2981-4089-ae0a-e1305d467872","蒸馏打不过RL？换对教师省50倍算力","opd-vs-grpo-compute-efficiency","2026-11-06T17:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"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":95,"title":96,"news_slug":97,"published_at":98},"fe609cac-33ac-4ea5-a6b7-db1a5913ec71","GPT-6 徒手改权重:Qwen 数学7.9到48.1","gpt6-astra-weight-space-programming","2026-11-02T23:09:36+00:00"]