[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lclm-1-16-compress-8-8x-75pct-accuracy":3,"news-related-c4380194-5bf9-43a0-8460-46436a4f2f97":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},"c4380194-5bf9-43a0-8460-46436a4f2f97","LCLM 把上下文压到 1\u002F16：8.8 倍提速的代价是 16 倍时准确率只剩 75%","Agent 跑得越久、检索文档越长，LLM 上下文窗口正在成为新的计算瓶颈。NYU、Columbia、Princeton、马里兰、Harvard 与 LLNL 联合发布的 LCLM（Latent Context Language Models）走了一条新路线：用 0.6B 编码器把输入 token 序列压成更短的「潜变量」再交给 4B 解码器生成，压缩在 prefill 之前完成——压缩比直接折算为解码端算力和显存节省。\n\n与 KV Cache 压缩的差别在于：主流方案要先 materialize 完整上下文再 evict，省的是存储而非解码端算力；LCLM 把整条压缩链路前置。论文在 RULER 长上下文基准上：4 倍压缩准确率 91.76%，相对 94.41% 的无压缩基线只掉 2.6 个百分点；16 倍压缩时为 75.06%，但生成速度比 KV Cache baseline 快 8.8 倍，所有被对比的 KV Cache 方法在同等压缩比下都更差。\n\n训练上 0.6B + 4B 在 350B token 上端到端训练，反直觉的发现是「解码器尺寸比编码器更重要，资源应优先放给解码端」。模型已开源 HuggingFace，可作为「前置压缩器」嵌进 RAG\u002FAgent 链路，研究者也演示了「选择性解压」——Agent 先扫读再聚焦关键段。\n\n「上下文即成本」在 2026 年已是常识。LCLM 的真正贡献不在 8.8 倍这个数字，而在于它证明「对输入本身做端到端压缩」是一条比 KV Cache 修剪更彻底、可工程化的路线。4 倍压缩（2.6 个百分点精度损失）是 RAG\u002FAgent 落地的甜点区间，16 倍更适合离线「先存后解压」类场景。对国内做长上下文优化的团队（DeepSeek V4 稀疏注意力、MIT CompreSSM、KVTC 等）来说，这篇论文给出的方向值得认真对照。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.09659","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",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":13},"697b8035-1f27-4daa-9634-99e22fb379c0","en","LCLM compresses context 16x; accuracy drops to 75% at 16x","arXiv 2606.09659 introduces LCLM (Lossy Context Language Model), an aggressive context compression method that achieves 8.8× inference speedup by compressing context to 1\u002F16 of its original size. The trade-off: 16× longer context results in only 75% accuracy (compared to full context).\n\nThe \"aggressive compression\" approach: LCLM uses a learned \"context compressor\" that reduces the context to 1\u002F16 of its size, with no specific quality preservation guarantees. The compressor is trained end-to-end with the language model, and the goal is to maximize inference speed, not context fidelity.\n\nThe trade-off: on a long-context QA benchmark, LCLM hits 75% accuracy with 8.8× speedup. The full-context baseline (no compression) hits 100% accuracy with 1× speedup. The \"accuracy vs speed\" trade-off is clear: 8.8× speedup for 25% accuracy loss. For some use cases (e.g., \"quick scan\" of long documents), this trade-off is acceptable; for others (e.g., legal review, medical records), it's not.\n\nThe \"informed choice\" highlight: LCLM is not claiming to be a universal solution — it's an aggressive compression method for specific use cases. The authors recommend it for \"scanning\" tasks (e.g., \"is this document relevant to my query?\") where some accuracy loss is acceptable. For \"precision\" tasks (e.g., \"extract specific information from this document\"), they recommend using less aggressive compression or no compression.\n\nThe bigger takeaway: \"lossy compression\" is becoming a real engineering trade-off. The \"lossless\" assumption is breaking, and the \"informed lossy compression\" approach gives users the choice of how to trade off accuracy and speed. For the industry, this means \"context compression\" tools will offer multiple \"compression levels\" (lossless, mild, moderate, aggressive), and users will pick the level that fits their use case.","lclm-1-16-compress-8-8x-75pct-accuracy","2026-06-15T04:00:00Z","2026-06-15T04:07:51.720780Z","2026-08-19T02:08:40.142862Z",true,"agent",110,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"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":44,"title":45,"news_slug":46,"published_at":47},"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":49,"title":50,"news_slug":51,"published_at":52},"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":54,"title":55,"news_slug":56,"published_at":57},"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",{"id":59,"title":60,"news_slug":61,"published_at":62},"c1ee21bd-4b59-418a-9b77-e46c790a8978","InfoKV 把 KV 缓存压缩推过「只看注意力」的临界点：用信息熵帮推理模型跑得更长","infokv-entropy-kv-cache-compression","2026-06-27T18:14:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"94e7739c-f218-4dfc-803d-3662a97321f3","DeepSeek V4 混合注意力架构解析：如何在1M上下文下将计算量降至原来的27%？","deepseek-v4-csa-hca-1m-27pct-flops","2026-05-27T07:20:00+00:00"]