[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lecun-platonic-representation-theorem-jepa":3,"topics-all":38,"news-related-8a550513-ef18-4d19-b609-d5fee5a853af":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},"8a550513-ef18-4d19-b609-d5fee5a853af","LeCun证明柏拉图表示定理:109M JEPA对齐胜72B","LeCun等人把「柏拉图表示假说」证明成定理:JEPA的正对若是同一场景的好视角,不同模态表示只差一个旋转;next-token prediction拿文档片段当正对,属坏视角,只能靠规模补齐。109M文本JEPA对齐超72B级受测模型,好视角省42倍样本,一个旋转拼出COCO 0.838的captioner。","10 月 29 日,arXiv 上挂出一篇署名 Randall Balestriero 和 Yann LeCun 的论文,标题干脆把流传两年多的「柏拉图表示假说」(PRH)改叫了定理:Proving and Improving the Platonic Representation Hypothesis with JEPAs。这条假说说的是,不同模态上训练的模型,随着规模变大,会收敛到同一个「世界表示」;它被反复引用,却一直没人证明,更没给出达成它的配方。这篇论文两件事一起做,顺手给 next-token prediction 贴了张「坏视角」标签。\n\n## 定理:好视角让表示只差一个旋转\n\n在论文的设定(单一世界状态的忠实渲染、总体最优)下,JEPA——联合嵌入预测架构——在任意多个模态上分开训练,只要正样本对是同一个场景的两次观测,作者称之为好视角(good views),学到的表示就只差一个旋转;一旦正对混入破坏单一观测的信息,失配有界,且编码器会把维度浪费在本模态的噪声上,任何旋转都对不齐。关键一步是把 next-token prediction 形式化成坏视角 JEPA:语言模型的正对是同一篇文档的不同片段,不是同一场景的两次观测。这顺带解释了一个老现象——语言模型为什么只能靠规模才与视觉对齐。\n\n## 42 倍样本,109M 胜 72B\n\n实验从 14.5 万条 Flickr 描述一路做到 9.4 亿张网页图片的文本。同语料下,好视角 JEPA 用 2160 万样本就达到 next-token prediction 要 9 亿样本才有的对齐度(0.109 对 0.111),少 42 倍;算文本 token 少 14 倍,算唯一图片少 435 倍。规模对比更狠:109M 参数的 BERT-B 文本 JEPA 一共只读过 1160 亿 text tokens,在 PRH 原论文基准上与 DINOv2 的对齐分数 0.184,高于作者评测的每一个语言模型——Qwen2.5-72B 是 0.154,Llama-3-70B、Mistral-7B 全在下方,连 CLIP 的文本编码器(0.124)都没守住。这些语言模型参数量是它的六百多倍,训练 token 多出一百多倍(15-18T)。作者还拟合出对齐的 scaling law:对齐度随好视角占比与样本量按对数线性增长,预测 85 个未见过的训练 run,误差与换个随机种子重跑同量级;等对齐线下,好视角占比与样本量近似 q∝M^-0.57——加训练可以亚线性地替代好视角,但永远不划算。\n\n## 一个旋转,拼出一个 captioner\n\n应用展示部分:用一个正交旋转,把只在文本上训练的语言模型与冻结的 DINOv2 接起来,全流程唯一用到图文配对的环节就是拟合这一个旋转。COCO 测试集 CIDEr 0.838;随机旋转只有 0.038;靠 CLIP 4 亿图文对训练的 ZeroCap 是 0.345。在无配对基准上,这个 109M JEPA 的 ImageNet-100 top-5 准确率 64.4%,高于基准论文报告的全部 21 对模型组合(最高 56%);同条件重跑其中最强的 Qwen3-8B 只有 47.6%,被拉开 16.8 个百分点;next-token 模型在同一批文本上一半 run 直接失败。\n\n## 但先泼三盆冷水\n\n把它读成「LLM 路线被证伪」就过了。其一,对齐不等于能力:论文证据止于表示几何与零样本分类,0.838 的 captioner 用了 3 万对 Flickr 数据拟合旋转,完全无配对的版本 CIDEr 只有 0.158。其二,配对训练没有被淘汰:CLIP 文本编码器在 FOSCTTM 和 ImageNet-100 上仍然更好。其三,scaling law 自己承认在四倍拟合范围外会高估好视角的收益。更要紧的是,语言模型的价值不止跨模态对齐——推理、指令跟随、代码,都不在定理射程内。42 倍是对齐效率,不是智能效率。\n\nLeCun 批评 autoregressive 路线多年,这次把批评写成了数学。定理真正改变的是问题本身:不再问「要多少数据」,而是问「正样本是什么视角」。规模能缓解坏视角的病,但那是症状缓解,不是药。下次有人拿 72B 参数压你,不妨先问一句:你的正样本,是什么视角?(论文:[arXiv:2610.39302](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39302))","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39302","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",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},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"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},"55b1816f-477f-47be-81a1-7d03e05d0378","en","LeCun proves Platonic theorem; 109M JEPA tops 72B","LeCun proves the Platonic theorem: JEPAs with good views align across modalities; a 109M JEPA beats every tested LLM up to 72B with 42x fewer samples.","On October 29, an arXiv paper by Randall Balestriero and Yann LeCun quietly renamed the two-year-old \"Platonic Representation Hypothesis\" (PRH) into a theorem: Proving and Improving the Platonic Representation Hypothesis with JEPAs. The hypothesis says models trained on different modalities converge to the same representation of the world as they scale; it has been cited endlessly, but never proven, and it offered no recipe. This paper does both — and slaps a \"bad views\" label on next-token prediction along the way.\n\n## The theorem: good views leave one rotation\n\nUnder the paper's setting (faithful renderings of one world, population optima), Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn representations that differ by at most one rotation, provided their positive pairs are two observations of the same scene — good views. If the positives corrupt a single observation, misalignment is bounded, and the encoder spends dimensions on its own modality's nuisance, which no rotation can align. The pivotal step formalizes next-token prediction as a bad-view JEPA: a language model's positives are parts of the same document, not two views of one scene. Which explains an old observation — language models only align with vision through scale.\n\n## 42x fewer samples, 109M beats 72B\n\nExperiments span from 145k Flickr captions to the texts of 940M web images. On the same corpus, a good-view JEPA reaches with 21.6M samples the alignment that next-token prediction needs 900M samples for (.109 vs .111) — 42 times fewer; that is also 14 times fewer text tokens and 435 times fewer unique images. The scale comparison is harsher: a 109M-parameter BERT-B text JEPA that read only 116B text tokens scores .184 against DINOv2 on the PRH paper's benchmark, above every language model the authors evaluated — Qwen2.5-72B sits at .154, Llama-3-70B and Mistral-7B below that, and even CLIP's text encoder (.124) does not hold. Those language models are over 600 times larger and trained on over 100 times more tokens (15–18T). The authors also fit a scaling law for alignment — log-linear in the fraction of good views and in samples seen — that predicts 85 unseen runs within seed noise; along iso-alignment curves, the good-view fraction trades against samples as q∝M^-0.57: more training substitutes for good views, sub-linearly, and never at a fair rate.\n\n## One rotation makes a captioner\n\nAs a showcase: one orthonormal rotation joins a text-only language model to frozen DINOv2, and the only part fitted on image-caption pairs is that rotation. It reaches CIDEr .838 on COCO; a random rotation gets .038; ZeroCap, built on CLIP's 400M pairs, reports .345. On the latest pair-free benchmark, the 109M JEPA reaches 64.4% top-5 accuracy on ImageNet-100, above all 21 vision-language pairs that benchmark reported (at most 56%); rerun under identical conditions, the benchmark's strongest model Qwen3-8B reaches 47.6% — 16.8 points behind; next-token models fail in half of their runs on the same texts.\n\n## Three buckets of cold water\n\nReading this as \"the LLM route is falsified\" overshoots. Bucket one: alignment is not capability — the evidence stops at representation geometry and zero-shot classification; the .838 captioner fits its rotation on 30k Flickr pairs, and the fully pair-free version manages CIDEr .158. Bucket two: paired training is not dead — CLIP's text encoder still wins on FOSCTTM and ImageNet-100. Bucket three: the scaling law itself admits it overestimates good views at four times its fitted range. And language models are worth more than cross-modal alignment — reasoning, instruction following, and code sit outside the theorem's range. 42x is alignment efficiency, not intelligence efficiency.\n\nLeCun has criticized the autoregressive route for years; this time the criticism is mathematics. What the theorem really changes is the question: not \"how much data\" but \"what counts as a positive\". Scale relieves the bad-view disease — symptom relief, not a cure. Next time someone flexes a 72B model at you, ask first: what kind of views are your positives? (Paper: [arXiv:2610.39302](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39302))","lecun-platonic-representation-theorem-jepa","2026-10-30T23:11:14Z","2026-10-30T23:11:17.799299Z","2026-10-30T23:11:17.799306Z",true,"agent",103,[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},"58aafc2c-b297-4f8c-a817-cad6ed7f1eda","跨模态检索头:18个模型实测,砍掉后音频检索归零","mllm-retrieval-heads-cross-modal-optimization","2026-11-01T17:09:14+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"e4ac848d-d24a-465f-bece-37e090da86de","全双工语音模型长对话必崩:KAIST 揪出初始化锚点被挤走","full-duplex-speech-collapse-fd-lhb-kv-cache","2026-10-27T03:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"a491d4db-2d5a-4723-b093-28d1512fd647","VLA-Δ:物理状态变化当监督信号,LIBERO 98.4","vla-delta-physical-state-transition","2026-10-20T19:07:27+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"fc533af5-9e8a-43c2-8408-2f931a6aa398","EngramEdit:把事实塞进 LLM 的记忆抽屉","engramedit-decoupled-knowledge-llm","2026-10-09T04:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d78ea8a8-bebb-4b94-b340-127eb2874a73","Ovis 全模态嵌入 3B:综合分领先 5.19","ovis-omni-embedding-3b","2026-09-23T21:08:45+00:00"]