[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-userarena-shopping-agent-user-simulation-benchmark":3,"topics-all":38,"news-related-ba7f04d1-2fe5-4e1d-90a6-94ec3c4b70df":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},"ba7f04d1-2fe5-4e1d-90a6-94ec3c4b70df","阿里 UserArena:64,729 条购物轨迹当考场","阿里联合莱顿大学、中国人民大学等在 arXiv 发布 UserArena 交互式推荐基准,含 11,010 名真实用户、64,729 条会话轨迹与 566,102 件商品,并提出轨迹级意图蒸馏 OPTD。自测中 OPTD 的线上 A\u002FB 效果误差 0.287,优于 GPT-6-Sol 的 0.576。","用户模拟器一直是推荐系统的老大难：光判断\"下一步会不会点\"没用，得看模拟出来的用户能不能扛住长链路，最后还得上线 A\u002FB 才对得上账。10 月 22 日，阿里集团联合莱顿大学、中国人民大学、中科院计算所在 arXiv 放出一篇论文（[UserArena 原文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.30932)），把这件事做成了一张可以公开检验的考卷。\n\n## 家底：64,729 条真实购物轨迹\n\n论文称数据来自\"亚洲某大型电商平台\"（平台名未公开），作者单位里明确挂着阿里集团。规模是：11,010 名用户、64,729 条会话轨迹、577,670 个动作、566,102 个商品、6 个用户属性字段，点击覆盖 41 个品类；其中 73,717 次商品点击里服饰占 54.90%，剩下 45.10% 摊在另外 40 个品类上。长链路部分另切出 59,378 条按时间排序的会话，来自 5,768 名用户，人均 10.29 次会话、86.36 个动作。对照组 OPeRA 是 51 人 527 次（人均 1.00 次），USB 是 5,222 人 5,274 次（人均 1.00 次）——这才是\"跨会话\"该有的量级。\n\n数据本身也不靠爆款撑：67,267 个被点过的商品里 92.0% 只出现一次，最热的前 1% 商品也只占 3.23% 的点击。动作分布同样真实，feed 页下滑占 38.34%，商品点击 12.76%，会话结束 11.21%，加购 3.90%，下单 1.02%。许可写得很死：仅限非商业研究，禁止反识别用户、禁止与外部个人数据关联、禁止未授权再分发；界面截图是离线 React 环境重建的，不含真实用户 GUI 画面。代码与数据在 GitHub 的 RecGPT-Team\u002FUserArena。\n\n## 五级评测：从\"能解析\"到\"复现线上\"\n\n评测分五层：环境合规（L0）、局部行为保真（L1）、序列行为保真（L2）、个性化人群保真（L3）、结果保真（L4）。配套两种跑法：Replay 用录好的状态预测下一动作，看局部准不准；Rollout 让模拟器自己往下跑，看整段轨迹像不像。L0 两种都跑，L1 只在 Replay 下评，L2 到 L4 全在 Rollout 下评。\n\n## OPTD：在学生自己的轨迹上蒸馏\n\n论文点了两条老路的毛病：SFT 只学录好的回答，存在训练测试错配；GRPO 给整条回答一个标量奖励，信用分配太粗。OPTD 换了个位置——学生自己 rollout 出来的每一个 token，都拿冻结教师在\"轨迹级意图\"这个特权上下文下的全词表分布当监督目标；再用逐 token 的信任域（按 KL 预算挑插值系数）把教师拉回\"只看得到学生可见状态\"的参照附近，另加一项接地损失对齐录好的真实回答，GT 系数 0.1。消融实验里去掉任一组件都会掉分，去掉 SFT 预热掉得最狠。\n\n## 论文自己测的一组数字：2B 跑赢 GPT-6\n\n论文自己做的生产 A\u002FB 校验最狠：同一对排序策略，分别在模拟器和线上真实 A\u002FB 的对照\u002F实验组里跑，比较 ΔPV-CTR 的\"模拟值 ÷ 线上值\"。OPTD 是 0.713×（误差 0.287），明显好过第二名 STA 的 0.297×；其余任务型模拟器把线上效果压到 0.020×～0.297×，没微调的 Qwen3.5-2B 只有 0.004×。反方向的是 GPT-6-Sol 的 1.576× 和 GPT-6-Luna 的 2.154×，高估了线上反应。论文由此下的结论是：通用大模型的容量，不等于行为校准能力。\n\n跨会话那张表里，OPTD 在 L1 到 L4 的八个指标中拿了五项第一、两项第二：DM-JSD 0.0284、轨迹长度误差 0.72、人群保真 PF 0.8929、终态 T-JSD 0.0077；macro-F1 和 ΔCTR 是 STA 领先。作者还做了图灵测试式人工评测：200 对真实与模拟轨迹、每对 3 名标注者、来源盲化且打乱顺序，OPTD 的\"人类赢率\"最低、\"模型赢率\"最高。\n\n## 但这张榜单现在是空的\n\n有个细节值得单说：项目页的 leaderboard 写着\"暂无已验证的 SOTA 分数，只收录完整的 Replay 与 Rollout 报告，不预估任何分数\"。数据集本身也还标着\"计划中\"，GitHub 仓库只有 3 次提交、3 个 star。想复现，基本得自己动手。\n\n## 所以呢\n\n这份工作真正的价值，可能不在\"又一个 benchmark\"，而在于它把\"模拟器准不准\"这个长期靠感觉的问题，换成了\"能不能复现线上 A\u002FB 的效果比\"这个可证伪的指标。一旦这么问，2B 的专用小模型压过 GPT-6 就不奇怪了——模拟用户要的是行为分布对齐，不是知识量。而榜单留空、数据待发、非商用许可，也说明它眼下更像一份评测协议，而不是一份拿来就能用的资源。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.30932","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"a5f25581-b4f0-432e-b0d8-6a08f1c08732","en","UserArena: 64,729 Real Shopping Sessions as Test Bench","Alibaba and partners released UserArena: 64,729 real shopping sessions plus the OPTD method, whose A\u002FB effect error of 0.287 beats GPT-6-Sol's 0.576.","Simulating shoppers has always been the hard part of recommendation systems. Predicting whether someone will click next is not enough — what matters is whether a simulated user holds up across a long shopping journey, and whether the simulation still matches what happened in a live A\u002FB test. On 22 October, Alibaba Group, Leiden University, Renmin University of China and the Institute of Computing Technology at the Chinese Academy of Sciences published a paper on arXiv ([UserArena](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.30932)) that turns this into a public test bench.\n\n## The dataset: 64,729 real shopping trajectories\n\nThe paper says the data was recorded on one of Asia's largest e-commerce platforms, without naming it; Alibaba Group is listed among the author affiliations. The scale: 11,010 users, 64,729 session trajectories, 577,670 actions, 566,102 products, 6 user attributes, with clicks spanning 41 product categories. Of 73,717 item clicks, apparel accounts for 54.90%, and the remaining 45.10% is spread across 40 other categories. The long-horizon portion carves out 59,378 chronologically ordered sessions from 5,768 users, averaging 10.29 sessions and 86.36 actions per user. By comparison, OPeRA has 51 users and 527 sessions at 1.00 average sessions, and USB has 5,222 users and 5,274 sessions at 1.00 — that is what a cross-session set should look like.\n\nThe data is not propped up by blockbusters: 92.0% of the 67,267 clicked items appear only once, and the most-clicked 1% of items account for just 3.23% of clicks. The action mix is equally realistic — feed-page scrolling at 38.34%, item clicks at 12.76%, session termination at 11.21%, add-to-cart at 3.90%, purchase at 1.02%. The license is deliberately restrictive: non-commercial research only, prohibiting attempts to re-identify users, linkage with external personal data, and unauthorized redistribution. Interface images are reconstructed in an offline React environment and contain no real user GUI screenshots. Code and data live in the RecGPT-Team\u002FUserArena GitHub repository.\n\n## Five evaluation levels, from \"parses\" to \"reproduces production\"\n\nEvaluation runs across five levels: environment compliance (L0), local behavioral fidelity (L1), sequential behavioral fidelity (L2), personalized population fidelity (L3), and outcome fidelity (L4). Two modes support them. Replay predicts the logged next action from recorded states, isolating local accuracy; Rollout lets the simulator choose actions and drive states until the session ends, testing whether free-running sessions reproduce real trajectories. L0 runs under both modes, L1 under Replay, and L2 through L4 under Rollout.\n\n## OPTD: distilling on the student's own trajectory\n\nThe paper identifies the weakness of two standard approaches. SFT learns only from recorded responses, creating a train-test mismatch. GRPO assigns one scalar reward per generated response, which is coarse credit assignment. OPTD relocates the supervision: every token of the student's own rollout is supervised by the full-vocabulary distribution of a frozen teacher conditioned on trajectory-level intent as privileged context. A per-token trust region — picking an interpolation coefficient against a KL budget — pulls that teacher back toward a reference conditioned only on what the student can observe, and a separate grounding term aligns against the recorded real response, with a GT coefficient of 0.1. In ablations, removing any single component degrades performance on at least one metric, and removing the SFT warm-up degrades it most.\n\n## The paper's own numbers: a 2B model beating GPT-6\n\nThe paper's production A\u002FB validation is the sharpest part. The same pair of ranking policies is run in simulation and in the control and treatment arms of a live A\u002FB test, comparing the ratio of simulated to online ΔPV-CTR. OPTD reaches 0.713× (effect error 0.287), clearly ahead of second-place STA at 0.297×. The other task-specific simulators compress the online effect down to between 0.020× and 0.297×, and the untuned Qwen3.5-2B reaches only 0.004×. Going the other way, GPT-6-Sol at 1.576× and GPT-6-Luna at 2.154× overestimate the online response. The paper's conclusion is that general-purpose model capacity does not automatically translate into behavioral calibration.\n\nIn the cross-session table, OPTD takes first place on five of the eight L1-to-L4 metrics and second on two more: DM-JSD 0.0284, trajectory error 0.72, population fidelity PF 0.8929, terminal-state T-JSD 0.0077. Macro-F1 and ΔCTR go to STA. The authors also run a source-blinded Turing-style human evaluation: 200 real-versus-simulated trajectory pairs, three annotators each, with sources anonymized and order randomized, where OPTD posts the lowest human win rate and the highest model win rate.\n\n## And yet the leaderboard is empty\n\nOne detail deserves its own note: the project page leaderboard states that no verified SOTA scores exist yet, that only complete Replay and Rollout reports are ranked, and that no scores are estimated. The dataset itself is still marked as planned, and the GitHub repository has 3 commits and 3 stars. Anyone wanting to reproduce this will be doing the work themselves.\n\n## So what\n\nThe real value here may not be \"one more benchmark\" but the substitution of a falsifiable question for a long-standing matter of taste: instead of asking whether a simulator is accurate, ask whether it reproduces the effect ratio of a live A\u002FB test. Once you frame it that way, a dedicated 2B model beating GPT-6 stops being surprising — simulating a user requires matching behavioral distributions, not absorbing more knowledge. The empty leaderboard, dataset still pending, and non-commercial license all signal that what exists today is closer to an evaluation protocol than a ready-to-use resource.","userarena-shopping-agent-user-simulation-benchmark","2026-10-26T09:10:00Z","2026-10-26T09:19:12.425831Z","2026-10-26T09:19:12.425842Z",true,"agent",202,[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},"a50b0336-dabe-405e-b8c4-d27156ab6d24","60% 权重转私有测试集:开放权重落后榜首 12 分","artificial-analysis-intelligence-index-v5","2026-10-28T07:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"aa294551-fbef-4cf4-b466-1928ab1925bb","朴素 RAG 拿下 6\u002F7 组评测:复杂记忆系统被比下去了","naive-rag-beats-complex-agentic-memory","2026-10-25T07:17:45+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"da6cbb98-3b89-4631-8dbd-59411136fb30","哈佛用 700 家公司拆解 AI 编程智能体:30% 代码增量,49% 合并时延","harvard-coding-agent-review-bottleneck","2026-10-14T11:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"c368ad9f-9308-4a0b-8f5c-3ae4601b48b9","D2K-Bench: 专家设计把 LLM 写 GPU 核提速 33.9%","d2k-bench-llm-gpu-kernel-design-guidance","2026-10-07T03:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"85f7f1c2-d896-436b-a915-37faed8776eb","你改主意了,模型没改:被拒需求也会带偏大模型","intent-eval-rejected-change-confusion","2026-10-06T17:15:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00"]