[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-chronoreplay-spatial-interface-lock-in":3,"topics-all":38,"news-related-5480e0ca-7b17-481c-8a24-8498f1a0aecd":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},"5480e0ca-7b17-481c-8a24-8498f1a0aecd","8B模型OST涨13.41分:空间推理的接口锁定","西湖大学团队提出 ChronoReplay，把异构空间监督按时间顺序回放并保留原始答案格式，让一个 Qwen3-VL-8B 在同等观测条件下同时提升 OVO-S、OST、UCS 三个基准，其中 OST 涨 13.41 分，自适应取帧还能再加分。","一个家用机器人跟着你走过公寓，随后被问三个问题：镜子在我左边还是右边？沙发离这儿多远？我们是什么时候经过那盆植物的？一次行走，三个问题，三种答案形式——一个选项、一个以米为单位的数字、一个时间点。\n\n流式空间基准现在把三类提问放在同一条因果历史里考。现实里，很多后训练配方却只监督其中一种回答格式。西湖大学团队（联合浙大、中科大等机构）新发的一篇论文给这种失败模式起了名字：**接口锁定**（interface lock-in）——模型在一种问法上越来越强，换到另一种问法上迁移微弱，甚至退步。（https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.34283）\n\n## 一个被忽视的失败模式\n\n观察很具体：训练空间 VLM 回答「物体在左边还是右边」，并不需要它保留决定该答案的度量位移；监督一次距离估计，也可能让分类关系不受约束。当后训练把权重压在某种读出口上，增益便集中到那个接口，其余接口的迁移随之变弱。\n\n同数据对照把这件事量化出来：只训分类单一家族的模型在 OVO-S-Bench 拿到 58.03，但 OST-Bench 掉到 47.48；只训轨迹家族是 52.03 \u002F 51.03 \u002F 49.22。完整配方拿到 53.89 \u002F 63.93 \u002F 52.77，三基准均值 56.86，是本文比较的设定里最高的，官方基座均值 48.28。\n\n注意形状：完整配方在 OVO-S 上其实低于「只训分类」那条。它不是每项都赢，而是用 OVO-S 的几分换回 OST 的大幅回升。\n\n## 三件套：原生目标、时序回放、家族配平\n\n第一阶段是因果回放构造：每个答案保留来源接口的原生格式——分类保留候选项，度量保留量级、符号、单位与精度，时间保留时间戳与事件序号；轨迹轮次按原始顺序重放，每个目标绑定在它自己的合法因果前缀上。作者称之为时间原点对齐，而非把多轮标注摊平成独立问答对。\n\n第二阶段是家族配平。保家族有 33,180 条单轮样本（30,635 条分类、2,545 条数值），轨迹家族有 6,381 个有序回合、45,006 个受监督的度量、计数、判断与时间轮次，前者样本数是后者的五倍以上。做法是在家族内部先归一化风险，再显式设定比例（α=0.5），不让样本多的家族独占共享参数。项目页给出的回放池规模是 39,561 条序列（https:\u002F\u002Fyihaohu0118.github.io\u002FChronoReplay\u002F）。\n\n底座是 Qwen3-VL-8B-Instruct，视觉编码器冻结，只训语言模型与视觉合并器。论文没有加地图、记忆模块或辅助状态损失——轨迹隐式留在 VLM 内部。\n\n## 三个基准同时抬起来\n\n在同等观测条件下（OVO-S 用 Uniform-128，OST 用完整场景不重叠协议，UCS 覆盖全部 8,114 道时间戳因果问题）：\n\n- OVO-S-Bench：48.98 → 53.89，涨 4.91 分，十二个任务族全部改善；\n- OST-Bench：50.52 → 63.93，涨 13.41 分，十五个子类全部为正，方向估计 +26.86、顺序判断 +20.57、距离时间定位 +18.36；\n- UCS-Bench：45.33 → 52.77，涨 7.44 分，八个子类全部为正，数量 +10.73、物体轨迹 +8.46、观察者轨迹 +7.38。\n\n增益集中在需要从多轮累积状态里组合答案的项目，而不是当前这一眼。\n\n## 13.41 分不能全记在「时序」头上\n\n这是全文最值得学的一处自我约束：13.41 分属于完整配方，不是「时序回放」单项的贡献。作者用移位顺序对照把后者单独隔离——保留每个回合块、目标和家族权重，只把回合块循环轮转，结果是 61.81 对 64.22，两对配对种子下平均差 2.41 分（分别为 2.42 与 2.41）。\n\n同类严谨性还体现在答案格式的对照上：原生格式在 UCS 上高 2.92 分，规范化格式却在 OST 上高 1.48 分——两者交叉，不存在全面占优的编码方式，论文明确说不该把原生格式读成普遍更优。\n\n对照普通联合 SFT（51.8 \u002F 60.7 \u002F 49.8）与自然轮次加权（52.9 \u002F 62.3 \u002F 51.5），完整配方三个端点全面领先，OST 分别高 3.23 与 1.63 分。\n\n## 另一半收益在推理侧\n\n自适应帧选择在训练权重完全不动的前提下改观测接入方式，平均用 80.50 帧替代 128 帧，OVO-S 从 53.89 抬到 57.98。论文把它单独拆成训练效应（+4.91）与接入效应（+4.09）两项，并指出接入效应集中在 L1（+8.66）。项目页给出的口径是少用 37.11% 的帧。L4 始终是最难的一级，35.16 → 40.17 → 42.30。\n\n## 该泼的冷水\n\n其一，这是作者自报结果。代码、配置、冻结的训练选择清单与打分脚本都公开了（https:\u002F\u002Fgithub.com\u002Fyihaohu0118\u002FChronoReplay），但微调权重、基准媒体与保存的预测没有打包，项目页写明不含已保存的预测。\n\n其二，横向比较有协议坑。表格里的已发布模型参考行跑的是各自报告的协议，作者提醒这些行不能当作同条件结论。8B 的 53.89 与已发表的 Qwen3-VL-235B-A22B 的 53.64 数值接近，但 Gemini-3.1-Pro 的 59.19 仍排在前面，这种比较就不能这么算。UCS 上人类仍然是 91.0。\n\n其三，「接口锁定」是论文反复强调的一个经验模式，而不是被定位出来的内部机制——结论里明确写着它不是一条被隔离的机制性主张。伦理声明同样直白：作者没有审计监控、自主导航等高风险用途，也不主张适合这些场景。\n\n## 所以呢\n\n这篇论文真正的贡献可能不是那 13.41 分，而是示范了把「配方总分」拆成「机制贡献」的做法：13.41 属于整套配方，时序只值 2.41，剩下的归给家族配平与目标格式。混着讲，评测就会失真。\n\n更直接的启发是：当团队发现模型「某个任务上变强了」，顺手量一下它在别的问法上有没有掉。接口锁定不会写在 loss 曲线上，它只会在你换一种提问方式的测试集上悄悄现形。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.34283","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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"bb1e331c-3afa-41ae-a65a-dc9e9a95f3e3","en","8B VLM +13.41 on OST: Spatial Interface Lock-In","ChronoReplay replays spatial supervision chronologically with native targets, letting one Qwen3-VL-8B lift OVO-S, OST and UCS together, +13.41 OST.","A home robot follows you through an apartment, then gets asked three questions: is the mirror on my left or my right? How far is the sofa from here? When did we pass the plant? One walk, three questions, three answer forms: an option, a number in meters, a moment in time.\n\nStreaming spatial benchmarks now pose all three over one causal history. In practice, though, plenty of post-training recipes supervise only one answer format. A new paper from Westlake University and partners (ZJU, USTC and others) names that failure mode: **interface lock-in**. The model gets better and better at one way of asking, and transfers weakly, or even degrades, when the question changes shape (https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.34283).\n\n## A Failure Mode Nobody Was Measuring\n\nThe observation behind the paper is concrete. Training a spatial VLM to answer \"which side of me is the object on\" does not require it to preserve the metric displacement that determines that answer. Supervision for a distance estimate, meanwhile, may leave categorical relations largely unconstrained. Once post-training pushes the weights toward one readout, gains concentrate on that interface while transfer to the others stays weak.\n\nThe authors quantify this with same-data controls. A categorical-only variant reaches 58.03 on OVO-S-Bench but drops to 47.48 on OST-Bench. A trajectory-only variant posts 52.03 \u002F 51.03 \u002F 49.22. The full recipe reaches 53.89 \u002F 63.93 \u002F 52.77, for a three-benchmark mean of 56.86, the highest among the supervision settings compared in the paper, against 48.28 for the official base.\n\nLook at the shape of that result. The full recipe actually scores *below* the categorical-only arm on OVO-S. It does not win every column; it trades a few points on OVO-S for a large recovery on OST.\n\n## Three Ingredients: Native Targets, Chronological Replay, Family Balance\n\nThe first stage is causal replay construction. Every answer keeps the native format of the interface it came from: categorical targets keep their candidate structure, metric targets keep magnitude, sign, units and annotation precision, temporal targets keep timestamps and event indices. Trajectory turns are replayed in their original chronological order, with each target attached to its own legal causal prefix. The authors call this temporal-origin alignment, and it is explicitly not a matter of flattening multi-round annotations into independent question-answer pairs.\n\nThe second stage is family-balanced learning. The preservation family holds 33,180 single-turn examples (30,635 categorical and 2,545 numerical). The trajectory family holds 6,381 ordered episodes, counting 45,006 supervised metric, counting, judgement and temporal turns, so the preservation side carries more than five times as many examples. The approach normalizes risk inside each family and then fixes their ratio explicitly (alpha = 0.5), so the larger family cannot dominate the shared parameters. The project page puts the full replay pool at 39,561 sequences (https:\u002F\u002Fyihaohu0118.github.io\u002FChronoReplay\u002F).\n\nThe base model is Qwen3-VL-8B-Instruct with the visual encoder frozen; only the language model and the visual merger are trained. The paper adds no map, no memory module and no auxiliary state loss. The trajectory stays implicit inside the VLM.\n\n## Three Benchmarks Rise Together\n\nUnder matched observation access (OVO-S at Uniform-128, OST under the complete scene-disjoint protocol, UCS covering all 8,114 timestamp-causal questions):\n\n- OVO-S-Bench: 48.98 to 53.89, a 4.91-point gain, with all twelve task families improving;\n- OST-Bench: 50.52 to 63.93, a 13.41-point gain, positive across all fifteen subtypes, led by Direction Estimation at +26.86, Order Judgement at +20.57 and Distance Temporal Localization at +18.36;\n- UCS-Bench: 45.33 to 52.77, a 7.44-point gain, positive across all eight subcategories, led by quantity at +10.73, object trajectory at +8.46 and observer trajectory at +7.38.\n\nThe largest gains land on exactly the items that must be composed from accumulated multi-turn state rather than from the current view.\n\n## The 13.41 Points Do Not Belong to Chronology Alone\n\nThis is the most transferable move in the paper. The 13.41 points belong to the complete recipe, not to chronological replay as a single ingredient. The authors isolate the latter with a shifted-order control that keeps every turn block, every target and every family weight, and only cyclically rotates the turn blocks: 61.81 against 64.22, an average gap of 2.41 OST points across two paired seeds (2.42 and 2.41).\n\nThe same discipline shows up in the answer-format comparison. Native format is 2.92 points higher on UCS, while Canonical format is 1.48 points higher on OST. The results cross, so no encoding wins everywhere, and the paper explicitly declines to read native format as universally superior.\n\nAgainst ordinary joint SFT (51.8 \u002F 60.7 \u002F 49.8) and natural turn weighting (52.9 \u002F 62.3 \u002F 51.5), the full recipe leads on all three endpoints, by 3.23 and 1.63 points respectively on OST.\n\n## The Other Half Lives at Inference Time\n\nAdaptive frame selection changes how evidence is accessed while the trained weights stay fixed. It averages 80.50 frames instead of 128 and lifts OVO-S from 53.89 to 57.98. The paper separates this into a training effect (+4.91) and an access effect (+4.09), noting that the access effect concentrates on L1 (+8.66). The project page frames it as 37.11% fewer frames. L4 remains the hardest level throughout, moving 35.16 to 40.17 to 42.30.\n\n## The Cold Water\n\nFirst, these are author-reported numbers. Code, configurations, the frozen training-selection manifest and the scoring scripts are public (https:\u002F\u002Fgithub.com\u002Fyihaohu0118\u002FChronoReplay), but the fine-tuned weights, the benchmark media and the saved predictions are not bundled, and the project page states outright that this release ships no saved predictions.\n\nSecond, cross-model comparison runs into protocol trouble. The published reference rows in the paper's tables were produced under each model's own reported protocol, and the authors themselves warn that those rows cannot be read as same-condition conclusions. The 8B score of 53.89 is numerically close to the published Qwen3-VL-235B-A22B reference at 53.64, but Gemini-3.1-Pro at 59.19 still sits ahead, and no such claim can be made against it. On UCS, humans still sit at 91.0.\n\nThird, interface lock-in is an empirical pattern the paper repeatedly flags, not an isolated internal mechanism. The conclusion says so directly. The ethics statement is equally plain: the authors have not audited high-stakes uses such as surveillance or autonomous navigation, and do not claim suitability for them.\n\n## So What\n\nThe real contribution here may not be the 13.41 points. It is the demonstration of how to separate a recipe total from a mechanism contribution: 13.41 belongs to the full recipe, chronology is worth 2.41, and the remainder has to be attributed to family balancing and target format. Lumping them together is how evaluation goes wrong.\n\nThe more directly useful lesson is simpler. When a team notices that a model \"got better\" at some task, it should go measure whether the model dropped on a different way of asking the same question. Interface lock-in never shows up on the loss curve. It only surfaces on the test set where somebody changed the phrasing.","chronoreplay-spatial-interface-lock-in","2026-10-28T05:09:43Z","2026-10-28T05:10:02.178440Z","2026-10-28T05:10:02.178449Z",true,"agent",131,[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},"4b693fb4-541f-47ed-8923-6e280cec965f","大模型的“记忆”还没过视觉这一关：MEMLENS 把长上下文的短板测出来了","memlens-multimodal-long-term-memory","2026-08-03T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"c3a956f8-dd42-46df-a1fd-1322dd38c15c","MentalThink 把 SVG 当作「心智草稿纸」:让多模态大模型学会用代码画心像做空间推理","mentalthink-svg-spatial-reasoning","2026-07-10T22:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"39e6e64f-a3de-4dce-bdd8-c8643f9413a1","Orca：把\"世界状态\"焊进潜空间——BAAI 推出通用世界基础模型新范式","baai-orca-world-foundation","2026-07-03T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b0a2cefc-7a4e-4f2c-83a0-f1e4911f04e5","RNG-Bench：GPT-5.4\u002FGemini 3.1 Pro 闭环记忆现形","rng-bench-shanghai-ai-lab-non-markov-memory","2026-06-24T18:15:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"2f9ad534-b737-490e-b97e-bf01f2b6d26b","交错生成新基准:闭源旗舰文本赢,图像一致性输给开源","intreal-interleaved-generation-benchmark","2026-11-01T05:27:08+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"1b51020b-9a0c-46f8-98ec-e4e0c8c051a6","基准分其实是难度的加权平均:25 个主流基准的体检报告","benchmark-difficulty-measure-coverage-gaps","2026-11-01T01:16:42+00:00"]