[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-coca-cola-latent-tool-context-compression":3,"topics-all":38,"news-related-f28966de-08b7-442e-ba3a-1d644131df26":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},"f28966de-08b7-442e-ba3a-1d644131df26","工具输出不必是人话:清华潜token压缩最多省89%","智能体调搜索工具后,原始返回动辄吃掉九成上下文。清华 Coca-Cola 把工具输出压成 128 个连续潜 token 直接注入大模型嵌入空间,大模型全程冻结。四个问答基准上 token 消耗最多降 89.24%,耗时最多降 55.97%,EM\u002FF1 反而全面超过不压缩基线,代码已开源。","智能体调一次搜索工具,返回的网页内容动辄上千 token,而下一步决策真正用得上的,可能只是一个人名、一个年份。清华 AI Agent Group 本周挂上 [arXiv:2611.06500](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.06500) 的论文先算了一笔账:到第 4 步工具调用时,原始工具返回已经吃掉智能体上下文的 96.7%。这些又长又吵的\"观察\"不仅烧钱,还会稀释模型对关键证据的注意力。他们的方案叫 Coca-Cola(Context Compression via Continuous Latent Tokens),思路相当激进:工具输出根本不必是人能读懂的文字。\n\n## 现有压缩方法的隐形天花板\n\n主流上下文压缩分两派。token 剪枝一族(如 LLMLingua 系列)让模型判断哪些 token 信息量低,直接删掉;文本摘要一派(RECOMP、Search-o1 的 Reason-in-Documents 模块、ACON)用辅助 LLM 把工具返回缩写成短文。论文指出它们共享同一个软肋:压缩结果仍是离散文本序列,必须保持语义连贯,压缩比天然有上限。连续潜 token 不受这个约束——单个向量可以在多个数值维度上联合编码信息,理论容量可以超过同等输入位置上的词表符号,论文还专门用码本容量给这个直觉做了形式化论证。\n\n## 128 个潜 token 换掉整段返回\n\nCoca-Cola 的结构是\"小模型编码器 + 潜在适配器\":作为压缩器的 Qwen3.5-4B 或 GLM-Z1-9B 把任务指令、当前推理轨迹、工具返回三者拼在一起编码,latent adapter 再把隐状态映射成固定数量的潜 token,直接插进目标大模型的输入嵌入序列,两侧用特殊边界标记隔开。推理时每条工具返回固定压成 128 个潜 token;搜索场景每次最多取 5 条结果、每个问题最多 5 次工具调用。目标大模型可以是 Qwen3.5-35B-A3B 或 GLM-4.7-Flash。\n\n## 训练目标只有一个:别改变下一步动作\n\n最关键的设计叫 Action-Preserving Training。压缩器不需要重建原文,而是把完整返回和潜 token 版本分别喂给同一个冻结的大模型,用 KL 散度对齐两者的下一步动作分布(损失权重 0.3),再用轨迹里记录的真实下一步动作做交叉熵监督(权重 1.0)。潜 token 连续可微,梯度能穿过冻结的大模型直接训到压缩器,实际只更新编码器的 LoRA 参数(rank 16)和 adapter。训练数据 3,927 对(另留 436 对验证),4 张 A100-80G 里 3 张放冻结的 agent、1 张放压缩器,学习率 1.5e-4 线性衰减到 5e-5,KL 项在前 20% 步数里热身。\n\n## 省下最多 89% 的 token,分数反而涨\n\n在 NQ、TriviaQA、2WikiMultiHopQA、HotpotQA 四个问答基准、两个目标模型上,效率提升是碾压级的:输入 token 方面,Qwen 组 NQ 从 14,362 降到 1,766(降 87.7%),GLM 组 2Wiki 从 31,693 降到 3,003(降 90.5%);端到端耗时最多降 56.0%(GLM 组 2Wiki 从 106 秒到 47 秒)。压缩率整体落在 70.01% 到 89.24% 之间。更反直觉的是质量:压缩后 EM 和 F1 全面超过不压缩的 Vanilla 基线,Qwen 组 HotpotQA 的 EM 从 11.00 提到 25.70,F1 从 16.41 提到 33.75;GLM 组 TriviaQA 的 EM 从 41.50 提到 51.50。八个模型-数据组合里六个拿到全部方法中的最优 EM 和 F1。也有例外:Qwen 组 TriviaQA 上,以文本重建为目标的 ICAE 在 EM 上更高(61.80 对 59.20)——论文没有回避这一点。\n\n## 所以呢\n\n论文明确说压缩器没有任何搜索专属设计,换个工具不用改架构。这意味着\"工具返回用潜表示\"可以做成一层通用接口。我的看法:上下文是智能体最贵的资源,大多数工作在优化\"往上下文里塞什么\",这篇把问题改写成\"上下文里的东西必须是文字吗\",这一步跨得比多数压缩工作都远。代价也明摆着:潜 token 人读不了,智能体出错时排查链路会更痛苦;压缩器本身还要单独占一张 GPU。论文自称(to our knowledge)是第一个把工具输出压成潜 token 的工作,这个方向值得持续盯——如果哪天工具协议层开始原生输出潜表示,智能体的成本结构会整个改写。代码在 [GitHub: THU-Agent\u002FCoca-cola](https:\u002F\u002Fgithub.com\u002FTHU-Agent\u002FCoca-cola),中文 README 直接给出 Vanilla、LLMSummary、LLMLingua 三条基线的复现路径,网页检索默认走 Exa,训练与评测脚本齐全。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.06500","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},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"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},"0adbb7ef-76db-4d79-a87b-a0ba3fda9274","en","Coca-Cola: 128 Latent Tokens Cut Agent Context by 89%","Tsinghua's Coca-Cola compresses tool outputs into 128 latent tokens for frozen LLM agents, cutting tokens up to 89.24% and runtime up to 55.97%.","Every time an LLM agent calls a search tool, the returned page content can easily run to thousands of tokens, while the next decision may hinge on a single name or a single year. A paper posted this week on [arXiv:2611.06500](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.06500) by Tsinghua's AI Agent Group starts with an accounting exercise: by the fourth tool call, raw tool observations already occupy 96.7% of the agent's context. These long, noisy observations not only burn money but also dilute the model's attention to key evidence. Their proposal, Coca-Cola (Context Compression via Continuous Latent Tokens), takes an aggressive stance: tool outputs do not have to be human-readable text at all.\n\n## The hidden ceiling of existing compression\n\nMainstream context compression falls into two camps. Token-pruning methods (the LLMLingua family) have a model score token informativeness and drop the rest. Text-summarization methods (RECOMP, Search-o1's Reason-in-Documents module, ACON) use an auxiliary LLM to condense tool outputs into shorter prose. The paper's critique applies to both: the compressed result is still a discrete token sequence that must remain semantically coherent, which inherently caps the compression ratio. Continuous latent tokens escape this constraint — a single vector can jointly encode information across many numerical dimensions, and the paper formalizes this intuition with a codebook-capacity argument showing latent representations can carry more information per input position than a vocabulary symbol.\n\n## 128 latent tokens replace a whole observation\n\nCoca-Cola pairs a small encoder with a latent adapter. A compressor model — Qwen3.5-4B or GLM-Z1-9B — jointly encodes the task instruction, the agent's current reasoning trajectory, and the tool observation. A latent adapter then maps the hidden states into a fixed number of latent tokens placed directly into the target agent's input embedding sequence, delimited by special boundary markers. At inference, every tool observation is compressed into exactly 128 latent tokens; in the web-search setting the agent retrieves up to 5 results per call and makes at most 5 tool calls per question. Target agents are Qwen3.5-35B-A3B and GLM-4.7-Flash.\n\n## One training objective: do not change the next action\n\nThe core design is Action-Preserving Training. Instead of reconstructing the original text, the compressor feeds the full observation and its latent-token version separately through the same frozen agent, aligning the two next-action distributions with KL divergence (loss weight 0.3), plus cross-entropy supervision from the recorded next action in the trajectory (weight 1.0). Because latent tokens are continuous and differentiable, gradients propagate through the frozen agent back into the compressor; only the encoder's LoRA parameters (rank 16) and the adapter are updated. Training uses 3,927 pairs (436 held out for validation) on 4 A100-80G GPUs — 3 GPUs host the frozen agent, 1 hosts the compressor — with learning rate 1.5e-4 linearly decaying to 5e-5 and the KL term warming up over the first 20% of steps.\n\n## Up to 89% fewer tokens, and scores go up\n\nAcross NQ, TriviaQA, 2WikiMultiHopQA, and HotpotQA with two target models, the efficiency gains are decisive. Input tokens drop from 14,362 to 1,766 on NQ with the Qwen agent (down 87.7%), and from 31,693 to 3,003 on 2Wiki with the GLM agent (down 90.5%). End-to-end runtime falls by up to 56.0% (GLM on 2Wiki, 106 s to 47 s). Overall compression rates land between 70.01% and 89.24%. The counterintuitive part is quality: after compression, EM and F1 consistently beat the uncompressed Vanilla baseline. On HotpotQA with the Qwen agent, EM rises from 11.00 to 25.70 and F1 from 16.41 to 33.75; on TriviaQA with the GLM agent, EM rises from 41.50 to 51.50. Coca-Cola takes the best EM and F1 among all methods in six of eight model-dataset settings. There is an honest exception: on TriviaQA with the Qwen agent, ICAE — trained with a text-reconstruction objective — scores higher on EM (61.80 vs 59.20), and the paper does not hide this.\n\n## So what\n\nThe paper states the compressor has no search-specific design; swapping in another tool requires no architectural change. That means \"tool outputs as latent representations\" could become a general-purpose interface layer. My take: context is the most expensive resource an agent has, and most work optimizes what gets inserted into it — this paper reframes the question as \"must the things in context be text at all?\", a bigger conceptual step than most compression work. The costs are equally clear: latent tokens are unreadable to humans, making debugging a painful affair when the agent misbehaves, and the compressor needs its own GPU. The authors claim (to their knowledge) to be the first to compress tool outputs into latent tokens. The direction is worth watching: if tool-protocol layers ever start emitting latent representations natively, the entire cost structure of agents gets rewritten. Code is on [GitHub: THU-Agent\u002FCoca-cola](https:\u002F\u002Fgithub.com\u002FTHU-Agent\u002FCoca-cola), with a Chinese README providing reproduction paths for the Vanilla, LLMSummary, and LLMLingua baselines, Exa as the default web-search backend, and full training and evaluation scripts.","coca-cola-latent-tool-context-compression","2026-11-07T23:05:00Z","2026-11-07T23:07:07.083548Z","2026-11-07T23:07:07.083556Z",true,"agent",61,[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},"821cbcc4-a8e0-41fe-90a8-284abd96f641","Agent上下文压缩提速8.68倍:小模型起草大模型验收","speccompact-agent-context-compaction","2026-11-05T19:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"851c9f58-ab76-414d-86b4-4b10dc782422","Memorilla:16个记忆token反超全上下文","memorilla-latent-semantic-memory-tokens","2026-11-03T07:03:31+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"a820bb5a-ac48-4aab-8d8b-34da340900c1","REMORY 让 LLM 用 5% 上下文继续跑 Agent","remory-agent-memory","2026-10-11T09:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b8ca86ba-2401-441d-ba5c-06cb6ef09420","随机抽样几乎就够了:高误差风险再砍50.2%","bestrand-benchmark-compression-random-sampling","2026-11-06T15:40:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"53d00d00-e4fb-41b0-a36a-37161203b514","KV缓存拼接复用丢题可至100%,修补要盯答案块","kv-cache-repair-budget-target-chunk","2026-11-05T15:09:02+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"64088ca6-c244-474c-93ce-cdbfe55de6d6","投机解码没用满?agent循环把加速比挖到10.87倍","specpilot-agentic-speculative-decoding","2026-11-04T13:15:17+00:00"]