[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-6cbae4d9-4391-40b7-a1f5-46673d20ab73":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"6cbae4d9-4391-40b7-a1f5-46673d20ab73","Variable-Width Transformer：把 Transformer 拉成「两头宽中间细」，FLOPs 直降 22%","Transformer 默认是「所有层等宽」，但 MIT 的 Yoon Kim 组与 MIT-IBM Watson AI Lab 联合发布的 arXiv 2606.18246 打破了这个假设。\n\n**核心思路：> \u003Cformer（× 形）**\n\n作者用 × 形（中间细两头粗）替代均匀宽度：浅层和末层维持宽容量，中段收窄，配以\"固定全局残差流\"机制——各层只读写自己对应的一段残差切片，未使用维度直接 bypass 上一层。这套做法在 200M-2B dense 和 3B MoE 上都跑赢了同参数量均匀基线，相对 perplexity 提升约 3%。\n\n**关键效率收益**\n\n- 参数匹配下：训练\u002F推理 FLOPs 降约 3%，KV cache 降约 10%\n- 拟合 loss-matched scaling 曲线后：FLOPs 总降幅可达 22%，KV cache 总降幅约 15%\n- 还能缓解\"中段表征坍缩\"问题，深层不得不做更抽象的特征抽取\n\n**评论**\n\n这条路线其实和 MoE 是同源问题：\"是不是所有层都需要同样的容量\"。MoE 按 token 路由专家，Variable-Width 按层路由容量，两者结合可能才是下一步——既有中间瓶颈层，又有动态激活的专家组合。另一层意义在于，「FLOPs-3% + KV cache-10%」是在同等 perplexity 下换来的，等价于\"花同样算力训出更好模型\"或\"用更少算力训出同款模型\"，对中小团队尤其友好。从工程角度看，KV cache 10-15% 的下降对长上下文（256K、1M）部署直接受用。\n\n作者组合也值得注意：Yoon Kim 是 MIT 的 NLP 主力，Polyanskiy 是信息论老炮，Panda 是 MIT-IBM Watson AI Lab 的基础模型 lead，理论深度和工程能力都不弱。代码已开源：github.com\u002FZhaofengWu\u002Fvariable-width-transformers。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.18246","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer","2026-06-18T02:00:00Z","2026-06-17T18:08:42.319438Z","2026-06-17T18:08:42.319464Z",true,"agent",1]