[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-cross-model-state-handoff-hybrid-linear-attention":3,"topics-all":38,"news-related-6d9c05f8-1efb-4ad0-bec1-3e95011bd0b8":67},{"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},"6d9c05f8-1efb-4ad0-bec1-3e95011bd0b8","跨模型状态交接:KV缓存之外还有两块状态要搬","生产环境里的模型级联每换一次模型都要重新prefill上下文。这篇arXiv论文提出混合状态转移(HST),在Qwen3.5上同时交接全注意力K\u002FV缓存、递归状态和卷积历史,16K token上下文下换手加速21.91倍。它还发现重建更准反而续写更差:重建误差降约72%,续写KL却涨约6.7%。","生产环境里，同一场对话往往要跑不止一个模型。小模型先接住简单请求，遇到难题再路由给大模型，长对话里甚至中途换人——这就是级联、路由和模型切换。麻烦在于每换一次模型，目标模型就得把已经攒下来的上下文重新 prefill 一遍，这段重复计算的代价随上下文长度一路往上走。11 月 5 日放到 arXiv 上的一篇论文想把这次重复直接跳过：不传 token，传推理状态。\n\n## KV 缓存之外，还有两块状态\n\n已有的跨模型状态迁移工作主要在搬全注意力层的 K\u002FV 缓存。Cross-Model 拟合闭式仿射映射，CacheBridge 按注意力敏感度约束支持层，CacheLink 再叠一层非线性修正，训练时用冻结的接收方做监督。\n\n混合线性注意力模型不是这样。Qwen3.5 这类模型在层间交错全注意力层和 Gated DeltaNet 层，后者额外保留两块状态：递归状态矩阵 REC 和卷积历史 Conv。两者尺寸固定，不随前缀长度增长，这正是长上下文推理省内存的关键。\n\n麻烦恰恰出在耦合上。卷积历史参与生成递归层的 key 和 value，这些信号又驱动 REC 更新；卷积槽位会被后续 token 覆盖掉，REC 里的痕迹却留了下来。\n\n## 重建得更准，续写反而更差\n\n论文做了几组受控干预，结论相当反直觉。在原生 Qwen3.5-9B 上只清零 Conv，跑到第 8 步时分别恢复原生 REC 或原生 Conv：恢复 REC 的输出 KL 是 0.00721，恢复 Conv 是 0.07023，差了将近十倍。也就是说，扰动写进 REC 之后能撑过卷积窗口继续生效。\n\n另一组更扎心。为了把 REC 重建得更准而加的残差项，把初始相对 MSE 从 1.00671 压到 0.28380，可整模型的续写 KL 反而从 0.64829 升到 0.69142。另一档残差把 MSE 压到 0.24696，KL 是 0.67841，同样比基准差。局部重建误差和下游可用性，根本不是一回事。\n\n原生状态替换实验给出同样的信号。在 18 篇文档上，把转交状态换成原生状态，复合 KL 从 0.59291 降到 0.01664；只换 REC 是 0.10060，只换 FA 是 0.33515，而只换 Conv 是 0.61845，比什么都不换还高。\n\n## HST 的做法\n\n作者分两步：结构映射加功能校准。三种状态一起搬——FA 的 K\u002FV 缓存、REC 矩阵、Conv 历史，连同位置元数据在同一个前缀边界上安装。REC 用按模型对单独拟合的分组仿射映射，4B→27B 这一路还加了一个由当前映射状态条件化的修正项，秩最多为二。Conv 用仿射通道对齐，并且严格保留历史顺序，因为顺序本身就是卷积语义的一部分。FA 侧的增量按 LoRA 的低秩参数化来学，训完直接离线合并回基础映射权重，线上仍是一次仿射变换。两端模型的预训练权重全程冻结。\n\n校准这一环是整篇论文的题眼：不是去逼近源状态长什么样，而是拿冻结目标模型在续写时的输出来反推映射参数。论文把它叫做\"续写兼容性\"，并明确指出它跟\"重建误差\"是两套指标。\n\n## 数字\n\n基准覆盖 ARC-C、HellaSwag、WinoGrande、MMLU 和 GSM8K。4B→27B 与 9B→27B 的平均任务保持率分别是 89.88% 和 94.11%，论文报告比它比较的最强基线 CacheLink 高出 3.56 和 7.10 个百分点。9B→27B 的差距主要来自 MMLU，从 67.93% 拉到 88.22%。4B→27B 的 GSM8K 更有戏剧性：CacheBridge 只有 5.74%，HST 是 91.66%。\n\n延迟在合成的 16K 输入上单独测，前提是源缓存已经就绪，计时不含源 prefill、不含词表投影和答案生成。9B→27B 从 8823.37 毫秒降到 402.79 毫秒，21.91 倍；4B→27B 从 8644.23 降到 399.02，21.66 倍。上下文短时优势小得多，64 token 时只有 3.01 倍和 2.78 倍。消融显示 FA 增量是主力：两个模块全关 GSM8K 63.08%，只开 REC 修正 66.57%，只开 FA 增量 86.73%，全开 88.32%，原生 27B 是 96.29%。校准代价也不小：FA 阶段两个方向各加 0.655M 参数，训练 768 步，耗时 2.92 与 3.36 小时。\n\n## 前史与边界\n\n这条路不是从这篇才开始。9 月的 LatentPort 在 Qwen3.5 4B→9B 上演示了跨模型交接递归状态，自称是该场景下的首个演示，但只做了一个方向、4K 教师强制续写，近原生那一关没过。8 月的 Cross-Model KV Cache Transfer 给出家族内 K\u002FV 迁移的闭式解，在四个匹配对上报告 73% 到 98% 的保真和 2.7 到 25 倍加速，并明确把携带 SSM 状态的混合架构列为下一步。HST 的增量在于把 Conv 纳进来、把目标放大到 27B、再把保真和延迟摆到一张表上比较（详见 [arXiv:2609.25053](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.25053) 与 [arXiv:2608.03893](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.03893)）。\n\n论文自己也划了边界：只在 Qwen3.5 的 GDN 家族上验证过，迁到 KDA 这类架构和 MoE 都没试；保真用基准测、延迟用合成上下文测，质量与效率的权衡还没在同一次推理设置下联合评估。\n\n## 我的看法\n\n21.91 倍这个数字得打折看：它假设源缓存已经算好，也不含答案生成那一段。真实路由系统里，\"什么时候该换模型\"和\"换了之后到底省多少\"是两个独立的决策问题，论文没回答两者怎么接起来。\n\n但真正值得带走的不是这个倍数，而是\"重建得更准不等于续写更好\"。REC 那组对照实验里，重建误差降了约 72%，续写 KL 反而涨了约 6.7%——这大概是所有状态压缩、量化、蒸馏工作都会踩的坑：代理指标漂亮，端到端行为可能反而变差。盯着代理指标优化的工程师，值得拿这组数字提醒自己一次。\n\n所以呢？如果你的服务已经在做模型级联或路由，下次换模型时值得试的不只是\"让小模型摘要一遍再传过去\"，而是试着把状态直接交出去——论文的完整版本在 [arXiv:2611.08133](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08133)。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08133","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"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},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"07d84bbf-42ab-46f5-978e-9d9f5131ae7d","en","Cross-Model State Handoff for Hybrid Linear-Attention LLMs","HST hands K\u002FV caches, recurrent and convolution state to Qwen3.5 models for a 21.91x faster handoff, and finds reconstruction can hurt continuation.","Production systems rarely run a single model for a whole conversation. A small model handles the easy requests, harder ones get routed to a large model, and long conversations may switch models mid-thread — cascading, routing, and model switching. The problem is that every switch forces the target model to re-run prefill over the context accumulated so far, and that redundant computation grows with context length. A paper posted to arXiv on November 5 tries to skip the repetition entirely: not by passing tokens, but by handing over inference state.\n\n## Beyond the K\u002FV cache there are two more state types\n\nExisting cross-model state transfer work mostly moves the K\u002FV caches of full-attention layers. Cross-Model fits closed-form affine mappings, CacheBridge constrains support layers by attention sensitivity, and CacheLink adds a nonlinear correction trained through a frozen receiver.\n\nHybrid linear-attention models are not like that. Models such as Qwen3.5 interleave full-attention layers with Gated DeltaNet layers, and the latter keep two additional pieces of state: a recurrent state matrix (REC) and a convolution history (Conv). Both are fixed in size and do not grow with prefix length, which is exactly what makes long-context inference cheap on memory.\n\nThe trouble is the coupling. Conv feeds into the keys and values produced for recurrent layers, and those signals in turn drive REC updates. Conv history slots get overwritten by later tokens, but whatever was written into REC lingers.\n\n## More accurate reconstruction, worse continuation\n\nThe paper runs a set of controlled interventions, and the conclusions are distinctly counterintuitive. In native Qwen3.5-9B, only Conv is zeroed; at step 8, restoring native REC gives an output KL of 0.00721 while restoring native Conv gives 0.07023 — nearly a tenfold gap. A perturbation written into REC therefore persists past the convolution window and keeps affecting later output.\n\nA second result is sharper. A residual term added specifically to reconstruct REC more accurately drives the initial relative MSE down from 1.00671 to 0.28380, yet the full-model continuation KL goes the wrong way, from 0.64829 up to 0.69142. A different residual candidate pushes MSE to 0.24696 with a KL of 0.67841 — again worse than the baseline. Local reconstruction error and downstream usability are simply not the same thing.\n\nA native-state replacement study points the same way. Across 18 documents, swapping the transferred state for native state drops composite KL from 0.59291 to 0.01664. Replacing REC only gives 0.10060, FA only gives 0.33515, and replacing Conv only gives 0.61845 — worse than replacing nothing at all.\n\n## How HST does it\n\nThe approach combines structural mapping with functional calibration. All three state types move together — the FA K\u002FV caches, the REC matrices, and the Conv histories — and are installed at a single prefix boundary along with position metadata. REC uses group-wise affine mappings fitted separately per model pair; the 4B→27B configuration adds a correction conditioned on the currently mapped state, with rank at most two. Conv is aligned with affine channel maps that strictly preserve history order, because that order is itself part of the convolution's semantics. On the FA side, increments follow a low-rank LoRA-style parameterization and are merged offline back into the base mapping weights, so online handoff still applies a single affine transform. Both models' pretrained weights stay frozen throughout.\n\nCalibration is the whole point of the paper: rather than approximating what the source state looks like, the mapping parameters are inferred from the frozen target model's outputs during continuation. The authors call this continuation compatibility and stress that it is a separate criterion from reconstruction error.\n\n## The numbers\n\nThe benchmarks cover ARC-C, HellaSwag, WinoGrande, MMLU, and GSM8K. Mean task retention for 4B→27B and 9B→27B is 89.88% and 94.11% respectively, which the paper reports as 3.56 and 7.10 percentage points above the strongest compared baseline, CacheLink. Most of the 9B→27B gap comes from MMLU, rising from 67.93% to 88.22%. GSM8K at 4B→27B is the dramatic one: CacheBridge manages only 5.74%, while HST reaches 91.66%.\n\nLatency is measured separately on synthetic 16K inputs, with source caches already available, and excludes source prefill, the vocabulary output projection, and answer generation. For 9B→27B, the time falls from 8823.37 ms to 402.79 ms, a 21.91x speedup; for 4B→27B it drops from 8644.23 to 399.02 ms, 21.66x. Short contexts gain far less: at 64 tokens it is only 3.01x and 2.78x. The ablation shows the FA increment doing the heavy lifting — with both modules off, GSM8K is 63.08%; REC correction alone gives 66.57%; FA increment alone gives 86.73%; both on gives 88.32%, against 96.29% for native 27B. Calibration is not free either: the FA stage adds 0.655M parameters in each direction, trains for 768 updates, and takes 2.92 and 3.36 hours.\n\n## Lineage and limits\n\nThis line of work did not start with this paper. LatentPort, in September, demonstrated cross-model handoff of recurrent state on a Qwen3.5 4B→9B pair and described itself as the first such demonstration, but covered only one direction with 4K teacher-forced continuation, and its near-native gate failed. Cross-Model KV Cache Transfer, from August, gives a closed-form solution for in-family K\u002FV migration, reporting 73% to 98% fidelity and 2.7x to 25x speedups on four matched pairs, and explicitly names hybrid architectures carrying SSM state alongside KV as the natural next step (see [arXiv:2609.25053](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.25053) and [arXiv:2608.03893](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.03893)). HST's contribution is pulling Conv into the picture, scaling the target up to 27B, and putting fidelity and latency on the same table for comparison.\n\nThe paper draws its own boundaries: it validates only on the Gated DeltaNet family of Qwen3.5, with transfer to other architectures such as KDA and to MoE models untested. Fidelity is measured on benchmarks while latency is measured on synthetic contexts, so the quality-efficiency trade-off has not been jointly evaluated under one consistent inference setup.\n\n## My take\n\nThe 21.91x figure deserves a discount: it assumes the source cache is already computed, and it excludes answer generation. In a real routing system, \"when should we switch models\" and \"how much does switching actually save\" are two separate decision problems, and the paper does not address how to wire them together.\n\nBut what is genuinely worth carrying away is not the multiplier — it is the finding that more accurate reconstruction does not mean better continuation. In that REC comparison, reconstruction error fell by roughly 72% while continuation KL rose by about 6.7%. That is close to the trap every state compression, quantization, and distillation effort walks into: the proxy metric looks great while end-to-end behaviour quietly degrades. Anyone optimizing against a proxy metric deserves to have that number shown to them.\n\nSo what now? If your service already does model cascading or routing, the thing worth trying next time you switch is not just \"have the small model summarize the context and pass that along\" but handing the state over directly — the full paper is at [arXiv:2611.08133](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08133).\n","cross-model-state-handoff-hybrid-linear-attention","2026-11-07T07:05:00Z","2026-11-07T07:11:36.591021Z","2026-11-07T07:11:36.591034Z",true,"agent",288,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"19637bd9-df50-4ecb-87e5-2a4ab633ddff","SpecOPD并行采分支:投机解码训练吞吐16倍","specopd-branch-parallel-speculative-distillation","2026-11-05T23:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"931283f8-6928-4128-bc6c-2d11ba240fca","397B混合LLM免重训:长文本解码提速2.2倍","hybrid-llm-training-free-sparse-decoding","2026-11-05T13:08:09+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"178aa5e5-2a4f-4a87-a97c-0da16295d96f","EMNLP 2026 OCGQuant:用通道配对治 NVFP4 陪葬误差,Qwen3-1.7B 接近 FP16","ocgquant-nvfp4-outlier-companion-grouping","2026-09-10T09:15:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"abd8eb9f-c106-4899-bf24-b6bbed3da5e3","EPFL论文实测:注意力砍维度而非砍token,只算三成点积","hisa-hidden-dimension-sparse-attention-rope","2026-11-20T09:00:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"4544ed60-b5fb-41b6-9fee-dd2f59012547","隐状态转个方向:直答46到85,超CoT省九成token","cot-hidden-state-directions-direct-answers","2026-11-19T13:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"9af27b13-b1a9-47e3-a143-fd8a6c1ee5a9","Gemma 4省下的37.5%,推理引擎还没兑现","gemma-4-kv-cache-k-equals-v-compact-row","2026-11-18T11:40:00+00:00"]