⌘K
Change language Switch ThemeSign In
Narrow Mode
估值 1200 亿后 Kimi 再扔王牌,新架构爆改 Transformer 老配件,比 DeepSeek 同款还省钱 ===========================================================
!Image 12: AI前线 AI前线 @AI前线
One Sentence Summary
Kimi published a new paper proposing 'Attention Residuals,' which introduces an attention mechanism to address the dilution problem in Transformer's residual connections, achieving a 1.25x computational efficiency improvement while maintaining performance, and earning praise from industry leaders like Andrej Karpathy.
Summary
This article provides an in-depth analysis of Kimi's latest paper, which introduces innovative improvements to the 'residual connection,' a fundamental component of the Transformer architecture. To address the 'dilution problem' caused by traditional residual connections, Kimi proposes the 'Attention Residuals' mechanism, which introduces a 'smart filter' for each layer, enabling the model to selectively focus on important content from preceding layers. The paper further introduces 'Block Attention Residuals' to reduce computational costs, complemented by optimization techniques such as 'cross-stage caching' and a 'two-stage computation strategy.' Experiments demonstrate that this method achieves a 3-7.5 point improvement on challenging tasks like graduate-level exams, mathematical reasoning, and code generation, delivering performance equivalent to traditional residuals requiring 1.25 times the computational power. The article also compares the distinct approaches of Kimi's solution and DeepSeek's mHC scheme.
Main Points
* 1. Traditional residual connections suffer from a 'dilution problem,' where information accumulates layer by layer, making it difficult for later layers to focus on key content.Analogous to a message relay team, where later workers receive a stack of content from everyone before them, key points get drowned out, making the AI less intelligent. * 2. Attention Residuals introduce a 'smart filter,' enabling each layer to selectively focus on important content from preceding layers.This is achieved by assigning scores to preceding layers using learnable attention weights, allowing the model to 'listen' more to important information and less to unimportant information, thereby solving the dilution problem. * 3. Block Attention Residuals significantly reduce computational costs while maintaining performance, achieving 1.25x computational efficiency.This involves dividing layers into groups, using standard residuals within groups and attention between groups, which preserves selectivity while saving costs. * 4. Kimi's and DeepSeek's improvements to residual connections represent two distinct approaches: selective listening versus multi-path parallelism.Kimi enables the model to selectively focus on key original information, while DeepSeek broadens multiple dedicated channels to transmit mixed information; both have different application scenarios.
Metadata
AI Score
79
Website mp.weixin.qq.com
Published At Today
Length 3629 words (about 15 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
原创 允毅 2026-03-17 15:51 北京
Kimi在残差连接发大招,让AI处理信息更聪明,比Deepseek的mHC还省钱
作者 | 允毅
连马斯克、Andrej Karpathy 都纷纷点赞,DeepSeek 和 Kimi 前后脚都盯上的“残差连接”,到底是什么?
最近,Kimi 放出一篇重磅新论文,瞄准一个过去十年几乎没人动过的 Transformer 底层根基:残差连接(Residual Connection)。残差连接由何恺明于 2015 年在 ResNet 论文中提出,此后便成为深度学习领域的标配。
简单来说,可以把大模型的 Transformer 架构,想象成一支几十人排成长队的“传话小组”,那么残差连接就像一条规定:每个工人听完前面所有人的话后,都往里面再补一句,然后原封不动往后传。
这套规则长这样:
但这会带来一个麻烦:队尾的工人收到的话,是前面几十个工人的内容全堆在一起的,越往后话越乱、越长,前面工人说的重点被埋住了,后面工人加的内容也没人听得清,AI 就变笨了。这叫“稀释问题”。
于是,Kimi 想到把“注意力机制”引进来解决这一问题,它提出一个新的规则:“注意力残差”(Attention Residuals)。如同给工人们配备了“智能筛选器”,不用再全盘收下前面堆出来的大杂烩,可以选择重点听前面自己关心的内容。
结果,AI 变得更聪明了。不仅实现了全场景的稳定提升,在研究生级专业考试、高等数学推理、代码生成、超长文本理解这类真正考验 AI 硬实力的高难度任务上,更是直接实现了 3-7.5 个点的暴涨。
而且更省钱了。在 Kimi 自家的大模型架构上验证显示可以节约1.25 倍算力,而训练端到端额外开销不到 4%,推理延迟增加不到 2%,基本没什么额外负担。
而且它完全适配大模型的缩放定律,模型规模越大,这套规则的优势越明显。
Kimi 论文里的精彩关注点
残差连接之所以能存在这么久,是因为它为 AI 信息的流动开了一条“主干道”,把教 AI 的纠错信号一路传下去,从而让深层网络更容易训练,不会把前面的好东西丢掉。
但这有三个致命问题改不了,分别是:
- 不能挑着听:不管是负责“抓重点”的工人,还是负责“整理内容”的工人,都只能听前面堆成一团的话,不能自己选听谁的
- 丢了的内容找不回来:前面工人说的重点,一旦被后面的内容盖住,就再也挖不出来了,AI 聊着聊着就忘了您最开始说的话
- 内容越堆越多:越往后的工人,必须扯着嗓子喊、加更多内容,才能让队尾听到,不然就被前面的内容盖住了,教 AI 的时候很容易教乱套
而 Kimi 的灵感,来自于他们发现,“序列维度上的时间问题和深度维度上的曾问题,本质是同一件事”。
当年 AI 读长句子,只能一个词一个词往后传,前面的内容记不住,后来发明了注意力机制,让 AI 能自己选句子里哪个词重要,一下子就解决了。而这和“工人队伍里传话的先后顺序”,本质上是一样的。
注意力残差链接的具体做法就是:
- 每个工人,都能直接听到前面所有工人的原声,还能打分,分数高就多听,分数低就少听,最后把所有话按分数合起来,再自己加工,传给下一个人
- 打分是不固定的:处理数学题的时候,就多听负责逻辑的工人的话;处理聊天的时候,就多听负责语气的工人的话,特别灵活。
为了让这一套机制真正落地,Kimi 推出了“省力好用版”的注意力残差,即“分块注意力残差”(Block AttnRes)。
它会把工人分成几个小组,每个小组内部用标准残差连接做累加,类似于“小组总结”,而组与组之间再用 Attention 来做选择性收听。这样一来,大大节省了成本开支。
为了进一步优化成本,Kimi 团队还做了两项针对性优化:
* 面向训练过程的“跨阶段缓存”:已经传过的旧内容直接存在本地不用反复来回发,每次只传新生成的内容,大幅节省算力、提升训练效率。
* 面向推理过程的“两阶段计算策略”:将 AI 生成回复前,先把要用到的历史内容一次性找齐,再一步步算结果,不用每算一个字都重新翻一遍历史,响应更快,用户几乎感知不到任何延迟。
最后,Kimi 实际测试了 5 种不同大小的 AI,从 194M 到 528M,分对比了传统残差连接、注意力残差连接和分块注意力残差的实际效果。
结果发现:
* 不管 AI 是大是小,注意力残差连接都比传统残差连接犯错更少、更聪明
* 分块注意力残差连接,花同样的钱,能达到传统残差连接 1.25 倍算力才能达到的效果
* AI 规模越大,分块版和完整无损版的注意力残差连接效果越接近,超大 AI 里,两者几乎一模一样
Kimi 还做了一系列消融实验,证明注意力残差连接中的每个设计,都必不可少:
* 必须用“随内容变的打分”,如果用固定的分数,效果就很差,甚至不如传统残差连接
* 必须用 softmax 打分,不然 AI 就不会果断选重点,效果变差
* 必须先把内容整理通顺再打分,不然信息容易盖住其他人,效果变差
* 分成 8 个左右的小组,效果最好
Kimi 的新设计给了 OpenAI 大神 Andrej Karpathy 不少启发,他忍不住深思,“注意力机制”是不是可以运用到更多的地方去?
有网友直接指出,或许 Transformer 中不少“长期沿用的配置”,都可以动一动了。
还有网友激动地将 Kimi 奉为“新的 DeepSeek”。
!Image 24 DeepSeek 与 Kimi 的 解法有什么不同?
虽然都是对残差连接进行创新,但 DeepSeek 推出的 mHC 却和 Kimi 的 AttnRes 走出了两条路。
如果把残差连接理解成 AI 内部的一套“传话机制”,那两家的分歧,首先来自于它们对“问题到底出在哪”的判断不同。
DeepSeek 的思路是,信息主干道太挤,信息才会乱,那就拓宽多条并行专线,用学习好的规则在专线之间混合、传递。
Kimi 的思路则是,信息主干道不能挑重点才会乱,那就给筛选权。
这也带来了两者在能力边界上的关键差别,DeepSeek 的做法只能听到 “混合后的二手内容”,听不到原版原话,而 Kimi 的做法能直接听到 “所有前面的原版原话”,想找谁的就找谁的。
两者都能提升基础效果,但擅长的领域完全不同,差距主要在复杂任务上。
* mHC:主打 “训练稳定性”,基础能力有提升,复杂任务乏力
它的核心价值是能把模型做的更深,训练的时候不会乱套,解决了深层模型训练崩溃的问题,在基础的语言理解、闲聊场景有稳定提升。但在需要多步推理、长上下文记忆、精准逻辑回溯的任务上,比如数学、代码、专业考试、长上下文问答,提升非常有限 。原因正在于它不擅长让模型“精准找回某个早期关键步骤”。
* AttnRes:兼顾稳定性 + 复杂推理,长上下文 、逻辑题提升碾压
它不仅解决了深层模型的训练稳定性问题,更擅长需要精准回溯、多步推理的高难度任务。
两者在落地适配性上天差地别。
DeepSeek 的 mHC 更像伤筋动骨的大改造,得把现有 AI 的单流架构完全改成多流并行架构,每层都要加新的流处理模块,还要重调全套训练参数,没法直接替换传统残差连接,老模型基本没法直接升级,跟拆房重盖没两样,适配成本极高。
而 Kimi 的 AttnRes 是即插即用的替换件,只需要把模型里原来的残差连接模块换成 Block AttnRes,其他模型结构、训练参数全都不用动,老模型直接升级就能拿到效果提升,就像给房子换个好门锁不用拆墙,适配成本极低。
从已有对比看,Kimi 似乎拿到了更好的“性价比”。
有网友将 Kimi 的 AttnRes 将与 DeepSeek 的 mHC 进行了性能对比,发现在 Kimi 的消融实验中,Kimi 完整无损版本的 Full AttnRes 性能始终优于 mHC,而 Block AttnRes 的性能与之相当,但 Kimi 的方案需要读写的数据量,只有 DeepSeek mHC 的 1/6。
这意味着,在大模型真正走向大规模工程部署的语境下,Kimi 的路线可能具备更强的现实吸引力。
过去十年,Transformer 的大部分创新都集中在注意力、FFN、MoE、位置编码和训练数据上。残差连接则像空气一样存在,重要,但过于基础,以至于几乎没人认真去动它。
而现在,DeepSeek 和 Kimi 先后出手,或许正释放一个信号:当数据红利逐渐到顶,过去的大模型比拼的是“谁能堆更多算力,那么下一代大模型,或许要开始学会“管信息”来拉开差距。
参考链接: https://github.com/MoonshotAI/Attention-Residuals/blob/master/Attention_Residuals.pdf
声明:本文为 AI 前线整理,不代表平台观点,未经许可禁止转载。
会议推荐
OpenClaw 出圈,“养虾”潮狂热,开年 Agentic AI 这把火烧得不可谓不旺。在这一热潮下,自托管 Agent 形态迅速普及:多入口对话、持久记忆、Skills 工具链带来强大生产力。但这背后也暴露了工程化落地的真实难题——权限边界与隔离运行、Skills 供应链安全、可观测与可追溯、记忆分层与跨场景污染、以及如何把 Agent 纳入团队研发 / 运维流程并形成稳定收益。
针对这一系列挑战,在 4 月 16-18 日即将举办的 QCon 北京站上,我们特别策划了「OpenClaw 生态实践」专题,将聚焦一线实践与踩坑复盘,分享企业如何构建私有 Skills、制定安全护栏、搭建审计与回放机制、建立质量 / 效率指标体系,最终把自托管 Agent 从可用的 Demo 升级为可靠的生产系统。
##### 今日荐文 Cursor 经历生死存亡 第一批能活到1000岁的人已出生?Neuralink联创抛出惊人预言 模型不再是关键?LangChain 创始人:真正决定Agent 上限的是运行框架 360 周鸿祎回应“龙虾”玩梗图:“龙虾”不是病毒,不会因为存在安全问题就否定一项技术 狂裁1600人,换掉CTO、晋升“下一代AI人才”!SaaS巨头的转型焦虑
!Image 28: AI前线 AI前线 @AI前线
One Sentence Summary
Kimi published a new paper proposing 'Attention Residuals,' which introduces an attention mechanism to address the dilution problem in Transformer's residual connections, achieving a 1.25x computational efficiency improvement while maintaining performance, and earning praise from industry leaders like Andrej Karpathy.
Summary
This article provides an in-depth analysis of Kimi's latest paper, which introduces innovative improvements to the 'residual connection,' a fundamental component of the Transformer architecture. To address the 'dilution problem' caused by traditional residual connections, Kimi proposes the 'Attention Residuals' mechanism, which introduces a 'smart filter' for each layer, enabling the model to selectively focus on important content from preceding layers. The paper further introduces 'Block Attention Residuals' to reduce computational costs, complemented by optimization techniques such as 'cross-stage caching' and a 'two-stage computation strategy.' Experiments demonstrate that this method achieves a 3-7.5 point improvement on challenging tasks like graduate-level exams, mathematical reasoning, and code generation, delivering performance equivalent to traditional residuals requiring 1.25 times the computational power. The article also compares the distinct approaches of Kimi's solution and DeepSeek's mHC scheme.
Main Points
* 1. Traditional residual connections suffer from a 'dilution problem,' where information accumulates layer by layer, making it difficult for later layers to focus on key content.
Analogous to a message relay team, where later workers receive a stack of content from everyone before them, key points get drowned out, making the AI less intelligent.
* 2. Attention Residuals introduce a 'smart filter,' enabling each layer to selectively focus on important content from preceding layers.
This is achieved by assigning scores to preceding layers using learnable attention weights, allowing the model to 'listen' more to important information and less to unimportant information, thereby solving the dilution problem.
* 3. Block Attention Residuals significantly reduce computational costs while maintaining performance, achieving 1.25x computational efficiency.
This involves dividing layers into groups, using standard residuals within groups and attention between groups, which preserves selectivity while saving costs.
* 4. Kimi's and DeepSeek's improvements to residual connections represent two distinct approaches: selective listening versus multi-path parallelism.
Kimi enables the model to selectively focus on key original information, while DeepSeek broadens multiple dedicated channels to transmit mixed information; both have different application scenarios.
Key Quotes
* One can imagine the Transformer architecture of large models as a message relay team of dozens of people lined up in a long queue. * It's like equipping the workers with smart filters, so they no longer have to accept the entire hodgepodge piled up from before. * Block Attention Residuals, for the same cost, can achieve the same effect that traditional residual connections would require 1.25 times the computational power to reach. * Perhaps many long-standing configurations in Transformer are ripe for re-evaluation.
AI Score
79
Website mp.weixin.qq.com
Published At Today
Length 3629 words (about 15 min)
Tags
Transformer
Residual Connections
Kimi
Attention Mechanism
Model Architecture
Related Articles
* After Reading 200 Papers: Analyzing DeepMind, Meta, DeepSeek - What AGI Narratives Are Chinese and American Giants Describing? | 2025 AI Annual Review * Software Engineering Outlook for the Next Two Years: From Writing Code to Managing AI, Programmers are Splitting into Two Careers * Google Co-founder Sergey Brin's Unusual Reflection: Underestimating Transformer and the Risks of AI Code Generation, 'The Higher Cost of Erroneous AI Code' * OpenAI Frontline Development Observations: Those Who Can Manage 10–20 Agents and Run Hour-Long Tasks Are Leaving Other Engineers Far Behind * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * Behind Meta's Email Deletion Incident: Why Does OpenClaw 'Lose Its Memory'? A Codebase Maintainer Deconstructs the Memory Architecture * This is Probably the Most Comprehensive, Readable, and Easy-to-Understand Article on AI Large Models I've Ever Read.") * Gemini Leadership: Pro's Main Role is to Distill Flash! The Greatest Potential for Breakthroughs Lies in Post-training; Noam, Jeff Dean: Continual Learning is a Key Direction for Improvement * 1,500 PRs, 0 Humans Coding: Codex-Driven Million-Line Internal Product Practice * Year-end AI Reflections by Salvatore Sanfilippo, the Creator of Redis HomeArticlesPodcastsVideosTweets
Kimi's New Ace After $120 Billion Valuation: Radically Re... ===============