[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-atom-2-7m-ucr-arithmetic":3,"topics-all":36,"news-related-4545706f-48c2-43d4-a3c3-e60aba316fd1":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"4545706f-48c2-43d4-a3c3-e60aba316fd1","Atom2.7m 撕掉「参数越大越会算数」的迷信:UC Riverside 用 2.74M 反超 1.56B 的 GPT-2 XL","UC Riverside 团队在 Hugging Face 博客发布的 Atom2.7m 是一个只有 2.74M 参数的因果语言模型,却在 ArithMark2.0 基准上拿到 69.24% 的准确率,直接把参数规模是其 568 倍的 GPT-2 XL(1.56B)压在 29.92% 的水平线上。它的核心思想是把算术失败重新归因到表征层面,而不是参数不够。BPE 类的自然语言分词器在面对数字时会把 12345 切成 123+45、12+345、1+2+3+4+5 等不规则片段,破坏了位值与操作数角色;通用位置编码描述的是 token 在序列里的位置,却不告诉模型这个 7 是十位还是个位。Atom2.7m 把数字跨度、位值、操作数身份显式暴露给模型,辅以 Abacus 风格的位置嵌入,这正是 2024 年 Position Coupling 等论文验证过能让加法从训练长度 30 位泛化到 200 位的同一类思路。文章同时指出,140M 的 MobileLLM-R1-base 也在 ArithMark2.0 上大幅优于 GPT-2 XL,进一步佐证小模型靠结构就能赢大模型靠记忆。它给当下的启示很直接:在评估 LLM 的算术与逻辑结构化能力时,参数量早已不是首要变量,tokenizer 设计、数值表征和位置编码才是新的胜负手。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fucr-max\u002Fatom2-7m-arithmetic-representation","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"c46cc39a-516c-4226-96d4-292e77ed0db0","en","Atom2.7m: a 2.74M model out-maths GPT-2 XL's 1.56B","Atom2.7m, published on the Hugging Face blog by the UC Riverside team, is a causal language model with only 2.74M parameters, yet it scores 69.24% on the ArithMark2.0 benchmark, directly crushing GPT-2 XL (1.56B) — 568× its size — at 29.92%. Its core idea is to re-attribute arithmetic failure to the representation level, rather than to insufficient parameters. BPE-style natural language tokenizers, when facing numbers, will split 12345 into irregular chunks like 123+45, 12+345, 1+2+3+4+5, breaking place value and operand role; general positional encoding describes the position of a token in the sequence, but doesn't tell the model whether this 7 is the tens place or the ones place. Atom2.7m explicitly exposes the digit span, place value, and operand identity to the model, supplemented by Abacus-style positional embeddings — exactly the same line of thinking verified in 2024 papers like Position Coupling, which enabled addition to generalize from a training length of 30 digits to 200 digits. The article also points out that MobileLLM-R1-base at 140M also significantly outperforms GPT-2 XL on ArithMark2.0, further corroborating that small models can win big models by structure rather than memory. Its lesson for the present is direct: when evaluating LLM's arithmetic and logical-structural ability, parameter count is no longer the primary variable — tokenizer design, numeric representation, and positional encoding are the new battlegrounds.","atom-2-7m-ucr-arithmetic","2026-07-10T08:30:00Z","2026-07-10T08:25:29.744565Z","2026-08-19T02:08:40.142862Z",true,"agent",255,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"f57dc66f-e478-4a8c-9d0b-2b7543ebaf4f","微博 3B 模型 VibeThinker 登顶推理榜：小模型也能正面硬刚旗舰","vibethinker-3b-weibo-94-aime26-reasoning","2026-06-19T07:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"c99751d5-418e-49d5-99d3-e43b84c80ec7","IBM与NASA开源月球基础模型:Lunar Foundation Model","nasa-ibm-lunar-foundation-model-sombench","2026-09-19T09:30:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"8def771a-d936-4859-930d-02c3011dc55c","LimiX-2 开源：一个模型吃下分类回归插补，表格三榜登顶","limix-2-tabular-foundation-model","2026-09-17T21:09:27+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"176b4807-da61-479f-a514-9381cd13319e","SP3O:3 个锚点修复 PPO critic 的平坦化","sp3o-sparse-critic-supervision","2026-09-17T17:10:01+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"d056f67b-7e0d-4e44-8d39-e31ea50deeae","Bonsai 2 27B 三元压缩:Qwen3.8 压到 5.9 GB,benchmark 留存 98.2%","bonsai-2-27b-ternary-qwen3-8-compression","2026-09-17T15:47:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"2e27016d-b90e-45c7-825a-41fd1e435c80","JHU 新研究:组合持续学习机制,百任务记忆留存从 1.2% 提到 34.9%","compose-cl-long-horizon-memorization","2026-09-16T15:10:00+00:00"]