[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kvtc-iclr-2026-transform-coding-kv-20x":3,"topics-all":36,"news-related-4a2b4984-cbdd-44f5-8d70-4ed453649586":55},{"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},"4a2b4984-cbdd-44f5-8d70-4ed453649586","KVTC革新LLM推理：ICLR 2026用Transform Coding将KV Cache压缩20倍","想象一个场景：你的模型正在处理一份128K tokens的超长文档，每次生成新token时都需要从显存中读取数十GB的KV缓存——这就是当前大模型推理的核心瓶颈。\n\nICLR 2026上，一篇名为《KV Cache Transform Coding for Compact Storage in LLM Inference》的论文提出了KVTC方案，用经典媒体压缩思路解决这个难题。核心思路分三步：首先通过PCA对Key\u002FValue特征进行去相关；然后使用自适应量化分配比特位；最后用熵编码完成最终压缩。在Llama 3、Mistral NeMo和R1-Qwen 2.5上的实验显示，KVTC最高可实现20倍压缩比，在特定场景下甚至超过40倍，全面超越H2O StreamingLLM等token驱逐法和SVD类方法。\n\n为什么这个方向值得关注？因为KV Cache的内存瓶颈本质上是「特征相关性强 + 量化粒度粗」——传统方法要么直接丢弃部分token，要么用固定精度压缩，而KVTC用数据驱动的方式找到了更好的平衡。更重要的是，它的压缩不会带来精度损失，这让实际部署的可行性大幅提升。\n\n当然，挑战依然存在：PCA计算本身有开销，跨会话的缓存复用也需要工程上的配套。但对于追求高吞吐量的推理服务商而言，这条路指向的是在同等硬件上服务更多用户的能力。\n\n值得思考的是，NLP领域的模型优化往往借鉴隔壁多媒体压缩的思路。JPEG用DCT去相关，VP9用变换编码——现在轮到LLM推理站在同一个肩膀上了。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2511.01815","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},"ab2dfa08-546c-4636-8387-7928ae74a955","en","KVTC: transform coding squeezes KV cache 20x (ICLR 2026)","Imagine a scenario: your model is processing a 128K-token super-long document, and each new token generation requires reading tens of GB of KV cache from VRAM — this is the core bottleneck of current LLM inference.\n\nAt ICLR 2026, a paper titled \"KV Cache Transform Coding for Compact Storage in LLM Inference\" proposed KVTC, using classical media compression thinking to solve this problem. The core idea has three steps: first, PCA de-correlates Key\u002FValue features; then adaptive quantization allocates bits; finally, entropy encoding completes the final compression. Experiments on Llama 3, Mistral NeMo, and R1-Qwen 2.5 show KVTC achieves up to 20× compression ratio, and in specific scenarios even over 40×, comprehensively outperforming H2O StreamingLLM and other token-eviction methods and SVD-class approaches.\n\nWhy is this direction worth attention? The memory bottleneck of KV Cache is essentially \"strong feature correlation + coarse quantization granularity\" — traditional methods either drop tokens directly, or use fixed-precision compression, while KVTC uses a data-driven approach to find a better balance. More importantly, its compression doesn't bring accuracy loss, which greatly improves the feasibility of actual deployment.\n\nOf course, challenges remain: PCA computation itself has overhead, and cross-session cache reuse also requires engineering support. But for inference service providers pursuing high throughput, this path points to the ability to serve more users on the same hardware.\n\nIt's worth thinking about: NLP-domain model optimization often borrows from neighboring multimedia compression. JPEG uses DCT to de-correlate, VP9 uses transform coding — now it's LLM inference's turn to stand on the same shoulders.","kvtc-iclr-2026-transform-coding-kv-20x","2026-05-14T04:00:00Z","2026-05-14T04:08:18.707924Z","2026-08-19T02:08:40.142862Z",true,"agent",342,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"a046b158-396f-4db9-a8d3-22d8c193a166","清华把KV缓存当神经网络训练:32倍压缩准确率96.7%","latent-compaction-test-time-kv","2026-10-15T17:08:40+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"dab2e487-cfba-4d75-a93f-eff133476348","剪枝校准对不上解码分布?西湖蚂蚁让稀疏解码提速48%","sparsedecoding-decoding-aware-pruning","2026-10-10T19:09:23+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"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":73,"title":74,"news_slug":75,"published_at":76},"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","latentpress-soft-token-context-compression","2026-09-05T19:06:09+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"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":83,"title":84,"news_slug":85,"published_at":86},"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"]