[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-smoothagent-ttft-12x":3,"news-related-6784d3bd-26c1-4fd5-a2e5-8c7b9e591dae":36},{"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},"6784d3bd-26c1-4fd5-a2e5-8c7b9e591dae","SmoothAgent 把上下文变换「提前做」：Agent 长链路 TTFT 砍到原来的 1\u002F12","长上下文 LLM Agent 跑多轮工具调用时，「上下文工程」听起来是个细节，实则是 TTFT（Time-To-First-Token）的隐形杀手。UCSD 与 UCLA 团队 6 月底在 arXiv 上公开的 SmoothAgent 论文把这个痛点一刀切开：他们提出 lookahead programming model，让 Agent 框架把上下文变换写成「异步操作」，运行时提前把变换后的 KV cache 准备好——实验数据是 TTFT 最多砍掉 11.9 倍。\n\n现代 Agent 框架靠 offloading、reduction、isolation 三类策略控制上下文长度，但每次变换都会让已有 KV cache 失效，触发一次完整的 re-prefill。这就是 TTFT 在多轮 Agent 中被反复推高的原因。\n\nSmoothAgent 的关键洞察是上下文变换是**段可分解**的——前缀的变换与未来 token 无关。抓住这一点，论文把变换操作改写成「异步任务」，运行时放到后台提前执行，等到真正需要时 KV cache 已经准备好，可以直接替换而不阻塞。配套的 lookahead-aware 调度器还能在延迟敏感的请求之间安排这些异步任务，控制相互干扰。\n\n论文在多种上下文工程策略上做了实验，并把方案接进既有 Agent 框架和 vLLM \u002F SGLang 这类 LLM serving 系统——不是停留在 demo，而是真的能在生产栈里跑。11.9 倍意味着过去十几秒的首 token 能压到一两秒，对长链 Agent 工作流的可用性是质变。\n\n和传统的 KV cache 优化思路比，SmoothAgent 不是把缓存「压得更小」或「留得更久」，而是从**调度时序**上把变换前置。Agent 框架几乎不用改业务代码，只要把同步的变换调用改成 async API 就能享受到加速——这种「零侵入」设计往往决定了一项优化能不能真正铺开。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.00151","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"248ff262-fe0c-469b-9772-69cb0c565f63","en","SmoothAgent pre-transforms context, cutting TTFT to 1\u002F12th","When long-context LLM Agents run multi-turn tool calls, \"context engineering\" sounds like a detail, but it's actually the invisible killer of TTFT (Time-To-First-Token). The SmoothAgent paper, posted at the end of June to arXiv by a UCSD + UCLA team, cuts into this pain point directly: they propose a lookahead programming model that lets the Agent framework write context transformations as \"asynchronous operations\", with the runtime pre-preparing the transformed KV cache — the experimental data is a TTFT cut of up to 11.9x. Modern Agent frameworks rely on three strategies — offloading, reduction, and isolation — to control context length, but each transformation invalidates the existing KV cache, triggering a full re-prefill. This is why TTFT keeps getting pushed up across multi-turn Agents. SmoothAgent's key insight is that context transformations are **segment-decomposable** — the prefix transformation is independent of future tokens. Catching this point, the paper rewrites the transformation operation as an \"asynchronous task\", executed in the background at runtime; by the time it's actually needed, the KV cache is already prepared and can be directly replaced without blocking. The accompanying lookahead-aware scheduler can also arrange these asynchronous tasks between latency-sensitive requests, controlling mutual interference. The paper ran experiments across multiple context-engineering strategies, and plugged the solution into existing Agent frameworks and LLM serving systems like vLLM and SGLang — it's not stuck at the demo level, it really runs in production stacks. 11.9x means the previous ten-plus seconds of first-token can be compressed to one or two seconds, which is a qualitative change for the usability of long-chain Agent workflows. Compared with the traditional KV cache optimization mindset, SmoothAgent doesn't \"compress the cache more\" or \"keep it longer\" — it **pre-positions** the transformation in the scheduling timeline. Agent frameworks barely need to change business code, just turn the synchronous transformation call into an async API to get the speedup — this kind of \"zero-intrusion\" design often decides whether an optimization can really take off.","smoothagent-ttft-12x","2026-07-23T03:00:00Z","2026-07-22T18:14:53.934396Z","2026-08-19T02:08:40.142862Z",true,"agent",127,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"83ef24e7-e2f5-4aec-8b4e-88bd10f8c28c","SelfCompact 把\"何时压缩\"从 LLM 智能体身上卸下:6 基准 18.1 分提点,token 成本砍 30-70%","selfcompact-jhu-when-to-compress-rubric","2026-06-23T08:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"5988b398-8533-4af5-9aca-ff6e51e9f429","Parallel-Synthesis：让 KV Cache 成为 Agent 多分支的「原生合成接口」，TTFT 提速 2.5–11×","parallel-synthesis-kv-cache-agent-2-5-11x","2026-06-15T14:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"c0fdcd77-4585-4b04-b042-ddf1789c2158","AI模型发布逻辑正在重写：2026年5月，四条主线改变竞争规则","may-2026-4-main-lines-competition-rewrite","2026-05-27T00:15:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"cac485ab-a429-4ebc-88c6-a1f924f978ff","AWS 开源 KeysAndValues:微调时就让模型学会“遗忘”,单张 A100 撑住 128K","aws-keysvalues-sparse-attention-finetuning","2026-08-26T05:20:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"c94bdf86-5de9-49fe-8c98-0f5c47611bfe","SGLang v0.5.18 发布:大模型冷启动提速 2.38 倍,710 个 PR 都改了什么","sglang-v0-5-18-cold-start-2-38x","2026-08-24T23:15:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f65e204c-0115-4b50-9113-2c3bb2ff6637","ReCache:给 Agent 的工具记忆装上独立缓存,KV 内存砍 92%、首 token 提速 3.655 倍","recache-agent-kv-cache-reuse","2026-08-24T15:30:00+00:00"]