[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-latentpress-soft-token-context-compression":3,"topics-all":38,"news-related-bce9fc16-d31a-49be-b17b-f144619a58e2":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},"bce9fc16-d31a-49be-b17b-f144619a58e2","LatentPress:上下文压成软令牌直读，7.7 倍压缩反超原文，训练仅动 0.1% 参数","长上下文的压缩通常走文本摘要或渲染成图两条路，康奈尔与爱荷华州立团队提出第三条：把对话与文档写成连续记忆令牌，冻结的解码器直接读。LongMemEval 上 7.7 倍压缩拿到 0.504，反超未压缩的 0.490，写入仅 43 毫秒。","给语言模型看的上下文，长期只有两种压缩方式：写成人类可读的摘要，或渲染成图像再 OCR 回来。两条路都要把信息还原成「文本」，可读它的根本不是人，是模型。康奈尔大学与爱荷华州立大学的 Zhengze Zhou 和 Hejian Sang 提出的 LatentPress 走第三条路：把对话历史和长文档直接写成连续记忆令牌，冻结的解码器通过输入嵌入接口直读，推理时零文本重建（arXiv:2609.01507）。\n\n## 只训 0.1% 的 adapter\n\nLatentPress 把上下文拆成 WRITE 和 READ 两步：前者把文本映射成紧凑的连续状态，后者直接喂给冻结解码器的嵌入层。真正训练的只有一个与读取器匹配的小写入器，参数 4.2M 到 26.2M，约为解码器的 0.1%，压缩率覆盖 4 到 16 倍。对比以往路线：Gist 要微调整个解码器，ICAE 训 LLM 级编码器还得靠自编码器重建——LatentPress 的卖点是「只训 adapter、不重建、soft token 直读」。\n\n## 摘要路线近乎灾难\n\nLongMemEval 的 500 道 oracle-evidence 问答上（Llama-3.1-70B-Instruct 判分，Qwen2.5-7B 读取器）：未压缩证据 0.490；LatentPress 在 7.70 倍压缩下拿到 0.504，4.62 倍和 6.27 倍分别为 0.476 与 0.478——压得更狠反而更准。对照组惨烈：文本摘要在 12.06 倍压缩只剩 0.184；DeepSeek-OCR 路线从 2.33 倍的 0.426 掉到 9.34 倍的 0.312；ICAE 在 17.28 倍时崩到 0.174。\n\nLongBench-QA 上，领域内训练的写入器在 4 倍压缩全面超过读原文：Qwen2.5-14B 从 47.93 升到 57.99，Qwen2.5-7B 从 43.80 到 49.06，Qwen3-8B 从 30.80 到 39.62。但 16 倍压缩时三个读取器全部跌破原文——压缩红利有边界。\n\n## 写入 43 毫秒，读取快 5 到 9 倍\n\n写入一段对话仅 43 毫秒（Qwen3-8B，H100 80GB），批量 DeepSeek-OCR 重建约 934 毫秒、文本摘要 407-645 毫秒，都慢一个数量级。读取侧压缩前缀比原文快 5.0 到 9.2 倍：Qwen2.5-14B 读原文 4.14 秒，LatentPress 只要 0.49 秒。\n\n## 泼冷水的地方\n\n跨骨干泛化是 reader 相关的：UltraChat 训练的写入器 zero-shot 迁到 LongMemEval，Qwen2.5-7B 打平原文，更弱的 Qwen3-1.7B 超过原文，但更强的 Qwen3-8B 上仍低于未压缩基线。仓库目前 2 个 star，未放出模型权重（HF 上引用该论文的模型为 0），复现靠 Docker 脚本在 B200 验证。判分只有单个 judge，偏好值得留意。\n\n## 所以呢\n\n这篇论文戳的是接口问题：业界默认压缩产物该是「可读文本」，可消费者明明是模型。摘要 0.184 对软令牌 0.504 的差距说明，中间那道「重建给人看」的工序既慢又丢信息；16 倍全线跌破原文也提醒，连续记忆的甜点区在 4 到 8 倍。下一个值得盯的是谁先把 soft token 接口做进生产级 Agent 的记忆层——代码和复现脚本都给了，球在工程侧。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01507","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"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},"96e43c42-8b9d-495f-9311-c2f73683c910","en","LatentPress: 7.7x soft-token compression beats raw, 43ms writes","Soft-token context: LatentPress hits 0.504 at 7.7x compression vs 0.490 raw, writes in 43ms, reads 5-9x faster. Only a 0.1% adapter is trained.","Context for language models has long had only two compression routes: have another model write a human-readable summary, or render documents as images and OCR them back. Both restore information to \"text\" — yet the consumer of that compressed result is not a human, it is a model. LatentPress, from Zhengze Zhou (Cornell University) and Hejian Sang (Iowa State University), takes a third path: conversational histories and long documents are written directly into continuous memory tokens that a frozen decoder reads through its input-embedding interface, with no text reconstruction at inference (arXiv:2609.01507).\n\n## Training only 0.1% of the parameters\n\nLatentPress splits context use into WRITE and READ: WRITE maps text or a document into a compact continuous state; READ feeds that state straight into the frozen decoder's input-embedding layer. The only component trained is a small reader-matched writer adapter, 4.2M to 26.2M parameters — about 0.1% of the decoder. Compression rates span 4x to 16x. Compared with prior routes: Gist fine-tunes the whole decoder, ICAE trains an LLM-scale encoder and still reconstructs through an autoencoder — LatentPress's pitch is \"train only an adapter, never reconstruct, read soft tokens directly.\"\n\n## Summarization is close to a disaster\n\nOn the 500 oracle-evidence questions of LongMemEval (judged by Llama-3.1-70B-Instruct, Qwen2.5-7B reader): uncompressed evidence scores 0.490; LatentPress reaches 0.504 at 7.70x compression, with 0.476 at 4.62x and 0.478 at 6.27x — compress harder, score higher. The baselines fare badly: text summarization collapses to 0.184 at 12.06x; the DeepSeek-OCR route slides from 0.426 at 2.33x to 0.312 at 9.34x; ICAE crashes to 0.174 at 17.28x.\n\nOn LongBench-QA, in-domain-trained writers beat reading raw text across the board at 4x compression: Qwen2.5-14B rises from 47.93 to 57.99, Qwen2.5-7B from 43.80 to 49.06, Qwen3-8B from 30.80 to 39.62. At 16x, however, all three readers fall below raw — the compression dividend has a boundary.\n\n## 43ms writes, 5-9x faster reads\n\nWriting one conversation takes 43 ms (Qwen3-8B, H100 80GB), versus roughly 934 ms for batched DeepSeek-OCR reconstruction and 407-645 ms for text summarization — an order of magnitude slower. On the read side, a compressed prefix is 5.0-9.2x faster than raw context: Qwen2.5-14B takes 4.14 s on raw, LatentPress only 0.49 s.\n\n## Where the cold water goes\n\nCross-backbone generalization is reader-dependent: a writer trained on UltraChat transfers zero-shot to LongMemEval, matching raw on Qwen2.5-7B, exceeding raw on the weaker Qwen3-1.7B, but staying below the uncompressed baseline on the stronger Qwen3-8B. The repository currently has 2 stars, ships no model weights (zero models on Hugging Face cite the paper), and reproduction runs through Docker scripts validated on B200. Judging relies on a single judge — its preferences are worth keeping in mind.\n\n## So what\n\nThe paper's real jab is at the interface: the industry defaults to compressed context being \"readable text,\" when the thing consuming it is a model. The 0.184-vs-0.504 gap between summaries and soft tokens shows that the \"rebuild it for humans\" step is both slow and lossy; the 16x collapse reminds that continuous memory's sweet spot sits at 4-8x. The next question worth watching is who first wires a soft-token interface into a production agent's memory layer — the paper ships code and reproduction scripts, so the ball is in engineering's court.","latentpress-soft-token-context-compression","2026-09-05T19:06:09Z","2026-09-05T19:06:31.048733Z","2026-09-05T19:06:31.048744Z",true,"agent",183,[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},"c83af54b-79ed-445c-9482-07d98c26c36b","BeaconKV:长推理会回头看,只压最近窗口的 KV 缓存注定丢东西","beaconkv-beacon-query-kv-cache-compression","2026-09-09T11:25:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"d31bc388-b6c7-41a5-a6e9-6f00657c7616","加GPU还是压KV缓存？arXiv论文：压缩省钱1.2到2倍，但36B是道坎","tensor-parallelism-vs-kv-compression-cost","2026-08-30T17:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d2c56430-f9a7-4844-bfab-a8651279a70c","ResKV 不再把 KV 缓存压缩等同于删词：给被淘汰的信息留一份残差账本","reskv-residual-kv-cache-compression","2026-08-03T10:43:23+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"c32d3160-4e07-4128-890f-4e135aac2cce","CompactifAI 把 Llama 3.3 70B 砍到一半:Multiverse 在 Intel Xeon 6 上跑出 1.9 倍吞吐","compactifai-llama-3-3-70b-intel-xeon","2026-07-26T04:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5f745fe5-ea5d-453a-8b08-7dac524d1ac2","ACL 2026 综述 sKis：KV 缓存优化重塑为 LLM serving 系统学","acl-2026-skis-kv-cache","2026-07-12T18:15:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"dd9d199c-5cd9-4a15-8ae6-7e4fb40f4129","MosaicKV:把 KV 缓存压成「马赛克」,长上下文推理跑出 16× 注意力加速","mosaickv-mosaic-compression","2026-07-03T18:01:00+00:00"]