[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deletion-sensitivity-reasoning-trace-routing":3,"topics-all":38,"news-related-343f6cae-ae14-4617-a317-7316748e1d6e":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},"343f6cae-ae14-4617-a317-7316748e1d6e","删掉推理链再转发:同档32B模型差51分","推理模型每轮都写思维链,请求跨供应商转发时这段历史会被剥掉。arXiv:2611.13223 用 BFCL v4 的 794 道多轮工具任务、九个开放权重模型、三个种子实测:同一次删除,成功率落差从 -3.9 分到 +51.3 分,同属 32B 档的三款模型内部就差 51.5 分。","做 Agent 路由的都知道:请求跨供应商转发时,模型的多轮工具调用历史就不完整了。推理模型每一轮都先写一段思维链再动手,下一轮靠读它继续规划。上游厂商把思维链绑在签发账号上,跨签发方转发的请求到达时,思维链整段被清掉,动作和工具返回结果还留着。\n\narXiv:2611.13223(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13223) 把「删掉思维链」当成受控变量,量任务成功率掉了多少分。实测跨度是 −3.9 分到 +51.3 分,两端都不含零;同一个 32B 参数档内部,三个模型就差出 51.5 分。\n\n## 三个条件,把删除做成可复现的实验\n\n任务池取 BFCL v4 的四个多轮子集,794 道题,每题 2 到 8 轮,覆盖基础场景、长上下文、缺函数、缺参数四类。评分用 BFCL 官方 checker,对实验条件盲评,每个模型每个条件跑三个随机种子。被测模型九款:Qwen3 的 8B \u002F 14B \u002F 32B 三档,GLM-Z1-32B、GLM-Z1-9B-0414、Baichuan-M2-32B、Ling-mini-2.0、R1-Distill-14B、Olmo-3-7B-Think。\n\n三个渲染条件:保留全部历史;跨轮删除(历史轮思维链全清,当前轮不动);会话中删除(第 K=⌊0.4T⌋ 轮切换,前段清后段留)。两个指标:Δdel(S) 是保留减跨轮删除的成功率差,Δdel(R) 是删除带来的重复调用率上升。\n\n## 同一档参数内部差 51.5 分\n\n实测跨度是 −3.9 分到 +51.3 分,跨九模型共 55.2 分。\n\nBaichuan-M2-32B 掉得最狠:保留时成功率 51.8%(309\u002F596),跨轮删除后掉到 0.5%(3\u002F573),Δ 是 +51.3 分,95% 区间 [+47.2, +55.3]。GLM-Z1-9B-0414 形态相同:19.2%(109\u002F567)掉到 2.7%(16\u002F591),Δ +16.5 分,区间 [+13.1, +20.1]。\n\n反过来 R1-Distill-14B 删掉历史后成功率反而从 0.9%(17\u002F1852)升到 4.9%(92\u002F1892),Δ 是 −3.9 分,区间 [−5.1, −2.9]。论文的说法是:它原本背着扛不住的思维链,删掉反而卸了负担。\n\n扎眼的是参数档内部。Qwen3-32B 读 −0.2、GLM-Z1-32B 读 +1.3、Baichuan-M2-32B 读 +51.3,同为 32B 就差出 51.5 分。同家对比更说明问题:GLM 的 32B 是 +1.3,9B 是 +16.5。论文因此把结论落在「只有模型身份能预测结果」上——尺寸、删除比例、任务类型都解释不了。\n\n## 失败有两种样子\n\n论文没只看成功率,它从原始调用日志里另开一条通道:统计「同一个调用原样发了两次」的会话比例得到重复率,与成功率合成两个轴,分成三种失效签名。\n\n补偿型两个指标都基本不动,Qwen3 三档、GLM-Z1-32B、Ling-mini-2.0 属这一类。循环型重复率飙升:R1-Distill-14B 在会话中删除下从 16.8% 冲到 66.4%,跨轮删除是 22.7%;Olmo-3-7B-Think 跨轮删除从 30.2% 冲到 66.5%。沉默型不给重复,直接放弃:Baichuan-M2-32B 成功率 51.8% 掉到 0.5%,重复率只升到 26.5%。\n\n调用流更能说明问题:Qwen3-8B 保留时 6 次调用就完成,跨轮删除后调用数涨到 410 次,结果照样对。R1-Distill-14B 在缺参数题上把 ls(a=True) 发出 418 次。\n\n## 删一半和删光,差别不到 2 分\n\n论文给唯一被做剂量实验的模型(Qwen3-32B)跑了六条臂:删 0%、25%、50%、75%、100%,外加一次随机删一半。成功率依次是 56.5、53.0、54.5、56.0、55.0、55.0。零臂与全删臂相差 1.5 分,每臂相对零臂的差值区间都跨零;随机删一半(54.5)和顺序删一半(55.0)也基本一致。分量在于:模型在「全删」下能扛住,那在任何剂量下都扛得住。\n\n子集分解提示伤害落在哪。Baichuan 在基础场景付出 +64.7 分,长上下文 +59.3 分,缺参数 +45.6 分,缺函数只有 +7.8 分——工具本身缺失时,删历史也救不回来。R1-Distill-14B 在切换点重复率冲到 66.4%,比从一开始全删的 22.7% 更狠;Baichuan 删掉前一半历史损失 37.1 分,全删是 51.3 分,伤害集中在早期轮次。\n\n## 两天之内可以测出来\n\n落地版是一个两格筛查:同一个模型跑保留格和跨轮删除格,各 200 题、一个种子、两台设备一天跑完,读 Δdel(S) 区间。判定用闭式标准误,这个规模上 1.96 倍标准误约 9.6 分,要判「保留」得点估计落在 20 分以上。\n\n三条样例读数:0 分区间跨零,删;+15 读 [+5.4, +24.6],归到不敏感档,判断出错概率约 17%;+25 读 [+15.4, +34.6],保留。\n\n重复率通道可零成本搭在旁边:循环型模型在跨度大的条件对上重复率上升 36 到 50 分,补偿型在筛查那一对上不超过 3.8 分。论文提醒,有一个循环型模型在筛查对本身只读到 +5.9 分,只看筛查读数会漏掉它,监控得盯日志。\n\n## 机制上厂商文档自己承认了\n\n论文的前提是「上游厂商把思维链绑在签发方上」。这不必只信论文:OpenAI 官方开发者文档(https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Freasoning)写着,持久化推理只能在同一模型家族内复用,gpt-5.6-sol、gpt-5.6-terra、gpt-5.6-luna 可以互用,但 GPT-5.6 与 GPT-5.5 家族之间推理不携带;切换模型家族时,即使把 reasoning.context 设成 all_turns,API 也会把不兼容推理从上下文省略。\n\n被删掉的不只是「历史里的一段文本」,而是模型续写推理所依赖的隐状态。跨供应商转发在今天不是例外路径,而是聚合生态的常规操作。\n\n## 但论文自己留了口子\n\n几处论文主动承认还没定论:补偿型是重新推导丢掉的历史,还是把思维链当闲置货物带着走,论文说是开放的;这个读数能否跨任务族迁移,是结论的开放边界;重复率上升是否先于失败发生,论文只当佐证通道,不作早期预警的主张;敏感档的剂量响应,论文明确说是另一个实验。\n\n所以用法是「先测出来,再决定转发策略」。做模型路由或多 Agent 编排,换个供应商接着跑不是无损操作,损失量级取决于具体模型,参数规模完全不提示。拿线上流量跑一次两格筛查——200 题、一个种子、一天——把删除敏感度读出来,再决定哪些模型固定在原厂、哪些可以放心转发。\n\n删除敏感度是模型的属性,不是接口的属性,这是最值得带走的判断。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13223","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},"74119d0f-ca56-4d21-9b95-d5a490a5cdf5","en","Deleting the Reasoning Trace: 32B Models Span 51 Points","arxiv:2611.13223 measures what deleting a reasoning model's history costs: the same deletion spans -3.9 to +51.3 points, and 32B models differ by 51.5","Routing a request across providers quietly guts part of the model. Reasoning models write a trace at every turn before acting, and later turns read that history back to keep planning. Providers bind the trace to the issuing account, so a request forwarded across issuers arrives with its reasoning history already stripped, while the tool calls and their results survive.\n\narXiv:2611.13223 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.13223) does something plain that nobody had done before: it treats deleting the reasoning trace as the controlled variable and measures what that costs in task success. The measured span runs from -3.9 to +51.3 points, with both ends excluding zero, and three models that all sit in the same 32B parameter class differ by 51.5 points.\n\n## Three conditions, one controlled operation\n\nThe task pool is the four multi-turn subsets of BFCL v4: 794 tasks of 2 to 8 turns each, spanning base scenarios, long context, missing functions, and missing parameters. Scoring uses the official BFCL checker, blind to the experimental condition, with three random seeds per model per condition. Nine open-weight reasoning models run the measurement, covering Qwen3 at 8B, 14B and 32B, GLM-Z1-32B, GLM-Z1-9B-0414, Baichuan-M2-32B, Ling-mini-2.0, R1-Distill-14B and Olmo-3-7B-Think.\n\nThree rendering conditions define the operation. Retention keeps the full history. Cross-turn deletion strips the trace from every historical turn while leaving the current turn intact. Mid-session deletion switches at turn K = floor(0.4T), clearing the early turns and keeping the late ones. Two indicators come out: Delta_del(S) is retention minus cross-turn deletion on success, and Delta_del(R) is the rise in repeated-call rate that deletion triggers. The harness is a modified DeepSeek Harness on a self-hosted vLLM engine, with tensor parallelism across two devices, temperature 0.6, and bootstrap intervals over 10,000 resamples.\n\n## 51.5 points of spread inside one parameter class\n\nThe measured span runs from -3.9 to +51.3 points, 55.2 points across nine models.\n\nBaichuan-M2-32B takes the heaviest hit. Retention holds 51.8 percent success (309\u002F596) and cross-turn deletion drops it to 0.5 percent (3\u002F573), a Delta of +51.3 points, 95 percent interval [+47.2, +55.3]; mid-session deletion lands at 14.7 percent (93\u002F632). GLM-Z1-9B-0414 has the same shape: 19.2 percent (109\u002F567) falls to 2.7 percent (16\u002F591), a Delta of +16.5 points, interval [+13.1, +20.1].\n\nR1-Distill-14B moves the other way. Deleting the history lifts success from 0.9 percent (17\u002F1852) to 4.9 percent (92\u002F1892), a Delta of -3.9 points, interval [-5.1, -2.9]. The paper's reading is that this model was carrying a trace it could not afford, so removing it removes a burden rather than a capability.\n\nThe sharp part is what happens inside a parameter class. Qwen3-32B reads -0.2, GLM-Z1-32B reads +1.3, and Baichuan-M2-32B reads +51.3: three 32B models spread across 51.5 points. Within a single house, GLM's 32B reads +1.3 while its 9B reads +16.5. The paper therefore lands on model identity as the only predictor among the factors it varied, with size, deletion ratio and task type leaving the reading unexplained.\n\n## Two shapes of failure\n\nSuccess rate alone misses half the damage, so the paper opens a second channel on the raw call log: the share of sessions that issue the same call twice. Combined with success, this splits the models into three behavioral signatures.\n\nCompensating models hold both indicators: the three Qwen3 sizes, GLM-Z1-32B and Ling-mini-2.0. Looping models spike on repetition: R1-Distill-14B goes from 16.8 percent to 66.4 percent under mid-session deletion and reads 22.7 percent under cross-turn deletion, while Olmo-3-7B-Think goes from 30.2 percent to 66.5 percent under cross-turn deletion, re-issuing the same call until the budget runs out. Silent models do not repeat, they give up: Baichuan-M2-32B falls from 51.8 percent to 0.5 percent success while repetition rises only to 26.5 percent, and GLM-Z1-9B-0414 repeats the shape at a smaller scale, 19.2 percent to 2.7 percent, with repetition dropping 7.2 points.\n\nThe call streams make the mechanism concrete. On one base task, Qwen3-8B finishes in 6 calls under retention; under cross-turn deletion the same plan is re-derived, the call count rises to 410, and the outcome still lands. On a missing-parameter task, R1-Distill-14B issues ls(a=True) 418 times. Baichuan reaches the correct state in 4 calls, and under deletion is left with a single cd(temp) and silence.\n\n## Half deleted and fully deleted land within two points\n\nThe one model put on a dose ladder, Qwen3-32B, runs six arms: 0, 25, 50, 75 and 100 percent of the history deleted, plus a second arm that deletes a random half. Success reads 56.5, 53.0, 54.5, 56.0, 55.0 and 55.0 percent. The zero and full arms differ by 1.5 points, every arm's difference from the zero arm has an interval crossing zero, and the random half sits at 54.5 against 55.0 for the block half. The practical weight: if a model holds up when everything is deleted, it holds up at every dose, and there is no reason to measure a partial deletion separately.\n\nThe per-subset decomposition says where the damage lands for sensitive models. Baichuan-M2-32B pays +64.7 points on base scenarios, +59.3 on long context, +45.6 on missing parameters and only +7.8 on missing functions, because when the tool itself is absent, deleting the history cannot rescue anything. Mid-session deletion behaves the same way: R1-Distill-14B spikes to 66.4 percent repetition at the switch point, harder than the 22.7 percent it shows when everything is deleted from the start, and Baichuan loses 37.1 points to deleting the first half of its history against 51.3 points for deleting all of it, so the damage concentrates in the early turns.\n\n## A screen that runs inside two days\n\nThe deployment version is a two-cell screen: one model, the retention cell and the cross-turn deletion cell, 200 tasks each, one seed, one day on two devices, reading the interval on Delta_del(S). The rule rests on a closed-form standard error whose half-width at this scale is about 9.6 points, so calling a model keep requires a point estimate above roughly 20 points.\n\nThree worked readings fix the arithmetic. A model at zero reads an interval straddling zero, so delete. A model at +15 reads [+5.4, +24.6], which excludes zero but still holds +10 inside, so the screen puts it in the insensitive tier, and under a normal approximation the chance of that call is about 17 percent. A model at +25 reads [+15.4, +34.6] and keeps.\n\nThe repetition channel rides alongside at zero scoring cost. Looping models rise 36 to 50 points in repetition at the widest-spread condition pair, while compensating models stay within 3.8 on the screening pair. The paper flags the case that makes the monitoring rule necessary: one looping model reads only +5.9 on the screening pair itself, so a deployment that trusts the screening reading alone would miss it, and the monitor has to watch the log.\n\n## The mechanism is in the vendor docs too\n\nThe paper's premise is that providers bind the trace to its issuer. That does not have to rest on the paper alone. OpenAI's own developer documentation (https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Freasoning) states that persisted reasoning can be reused only within the same model family, so gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna can reuse each other's reasoning while reasoning does not carry between the GPT-5.6 and GPT-5.5 families. It further states that when you switch model families, the API omits incompatible reasoning from the model's context even when reasoning.context is set to all_turns. The same document describes the current_turn mode as making reasoning from the active turn available while it does not render reasoning from earlier turns into the next sample.\n\nWhat gets stripped is therefore not a stretch of text in the history but implicit state the model relies on to keep reasoning. Forwarding across providers is not an edge case in today's aggregation ecosystem, it is the normal path.\n\n## Where the paper leaves it open\n\nSeveral limits are stated by the authors themselves. Whether compensating models re-derive the lost trace or carry it as idle cargo is open, and separating the two needs a finer probe. Whether the reading transfers across task families is the open boundary of the claim. Whether the repetition rise precedes the failure inside a session stays unclaimed as an early warning, since the paper uses it only as corroborating evidence. And the dose response for the deletion-sensitive regime is explicitly another experiment.\n\nSo the usable form of this finding is to measure it before routing, not to configure a router from the paper.\n\n## So what\n\nFor anyone doing model routing or multi-agent orchestration, handing the request to a different provider and keeping going is not a lossless operation. How much the reasoning chain costs depends on the specific model, and parameter size gives no hint. The cheap move is to run the two-cell screen on your own traffic, 200 tasks, one seed, one day, read out that model's deletion sensitivity, then decide which models stay pinned to their origin and which are safe to forward.\n\nDeletion sensitivity is a property of the model, not of the interface. That is the part of the paper worth carrying away.\n","deletion-sensitivity-reasoning-trace-routing","2026-11-13T07:05:00Z","2026-11-13T07:26:14.326606Z","2026-11-13T07:26:14.326634Z",true,"agent",108,[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},"6c8ae96a-53e0-4ef7-8609-40b345cdee7c","决策模型实测:自己演砸42.9%,教学生82.9%","decision-model-actor-judge-teacher","2026-11-15T21:45:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bd350599-7253-4f32-a874-9224bab89a49","多智能体搜索实测：32组对照里委派多数时候反而更差","multi-agent-search-delegation-accuracy","2026-11-15T09:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"a2038f05-8ad2-4b57-bb29-47ff522e09a0","失败轨迹偷走权重:ALFWorld成功率涨7到9.4点","outcome-mass-alignment-multi-turn-agent-rl","2026-11-14T11:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9531b4f7-3f4d-474c-b233-c715925a3e76","8.4B轨迹压过25.3B技能语料:形式化证明中训配方","davinci-prover-agent-native-mid-training","2026-11-14T07:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"365a0af2-d564-4eff-bf03-a138cdea0c5f","研究智能体大考:只查交付过69%,查全要求剩28%","are-bench-research-agent-context-requirements","2026-11-13T01:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f2723003-43ac-47fe-bb59-49030446919a","28个决策模型大考:多选最高才36.81%","mmdm-decision-model-benchmark","2026-11-12T13:30:00+00:00"]