惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

WordPress大学
WordPress大学
O
OpenAI News
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
博客园 - 三生石上(FineUI控件)
Webroot Blog
Webroot Blog
GbyAI
GbyAI
S
SegmentFault 最新的问题
Cyberwarzone
Cyberwarzone
Exploit-DB.com RSS Feed
Exploit-DB.com RSS Feed
J
Java Code Geeks
Google DeepMind News
Google DeepMind News
cs.CV updates on arXiv.org
cs.CV updates on arXiv.org
博客园 - 【当耐特】
S
Secure Thoughts
酷 壳 – CoolShell
酷 壳 – CoolShell
AWS News Blog
AWS News Blog
Engineering at Meta
Engineering at Meta
S
Security Affairs
H
Help Net Security
Microsoft Security Blog
Microsoft Security Blog
D
DataBreaches.Net
云风的 BLOG
云风的 BLOG
Hugging Face - Blog
Hugging Face - Blog
Google DeepMind News
Google DeepMind News
Spread Privacy
Spread Privacy
T
Threatpost
Forbes - Security
Forbes - Security
C
Cisco Blogs
Scott Helme
Scott Helme
Attack and Defense Labs
Attack and Defense Labs
Simon Willison's Weblog
Simon Willison's Weblog
腾讯CDC
The Last Watchdog
The Last Watchdog
Cloudbric
Cloudbric
Last Week in AI
Last Week in AI
Recorded Future
Recorded Future
小众软件
小众软件
V
Vulnerabilities – Threatpost
美团技术团队
人人都是产品经理
人人都是产品经理
有赞技术团队
有赞技术团队
Apple Machine Learning Research
Apple Machine Learning Research
Hacker News - Newest:
Hacker News - Newest: "LLM"
I
Intezer
月光博客
月光博客
C
Cyber Attacks, Cyber Crime and Cyber Security
博客园 - 司徒正美
C
Cybersecurity and Infrastructure Security Agency CISA
Martin Fowler
Martin Fowler
博客园 - 聂微东

Hacker News: Best

madhadron - The seven programming ur-languages GitHub - smol-machines/smolvm: Tool to build & run portable, lightweight, self-contained virtual machines. I Measured Claude 4.7's New Tokenizer. Here's What It Costs You. Introducing Claude Design by Anthropic Labs It Is Time to Ban the Sale of Precise Geolocation The creative software industry has declared war on Adobe Isaac Asimov: The Last Question Newly unsealed records reveal Amazon’s price-fixing tactics, California attorney general claims Clojure - Documentary Android CLI and skills: Build Android apps 3x faster using any agent Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7 Codex for almost everything Introducing Claude Opus 4.7 Qwen Studio The Future of Everything is Lies, I Guess: Where Do We Go From Here? Virginia Bans Sale of Geolocation Data YouTube now lets you turn off Shorts Burgers | マクドナルド公式 ChatGPT for Excel Ask HN: Who is using OpenClaw? Live Nation illegally monopolized ticketing market, jury finds Google Broke Its Promise to Me. Now ICE Has My Data. Open Source Isn't Dead. The Future of Everything is Lies, I Guess: New Jobs Unexpected €54k billing spike in 13 hours: Firebase browser key without API restrictions used for Gemini requests IPv6 – Google Your Backpack Got Worse On Purpose Good sleep, good learning, good life Fixing a 20-year-old bug in Enlightenment E16. Does Gas Town 'steal' usage from users' LLM credits & paid services to improve itself? Tell HN: Fiverr left customer files public and searchable Cybersecurity Looks Like Proof of Work Now Getting the Flock out Release OpenSSL 4.0.0 · openssl/openssl Internet será irrespirable los días de fútbol y otros deportes. Telefónica extiende los bloqueos a Champions, tenis y golf. Automate work with routines - Claude Code Docs The Future of Everything is Lies, I Guess: Work Thousands of rare concert recordings are landing on the Internet Archive — listen now What is jj and why should I care? Backblaze has quietly stopped backing up your data Cal.com Goes Closed Source: Why AI Security Is Forcing Our Decision | Cal.com - Scheduling Software for Online Bookings Codex Hacked a Samsung TV The Future of Everything is Lies, I Guess: Safety GitHub - sterlingcrispin/nothing-ever-happens: Polymarket bot that buys "No" on all non-sports markets. For entertainment only, mostly a meme. Make tmux Pretty and Usable - Ham Vocke Microsoft isn't removing Copilot from Windows 11, it's just renaming it Servo is now available on crates.io - Servo aims to empower developers with a lightweight, high-performance alternative for embedding web technologies in applications. We May Be Living Through the Most Consequential Hundred Days in Cyber History, and Almost Nobody Has Noticed All elementary functions from a single binary operator 奈拜提耶市 Seven countries now generate 100% of their electricity from renewable energy Pro Max 5x Quota Exhausted in 1.5 Hours Despite Moderate Usage Tell HN: docker pull fails in spain due to football cloudflare block Bring Back Idiomatic Design @adlrocha - How the "AI Loser" may end up winning Apple update turns Czech mate for locked-out iPhone user Cache TTL silently regressed from 1h to 5m around early March 2026, causing quota and cost inflation The peril of laziness lost AI Will Be Met With Violence, and Nothing Good Will Come of It Center for Responsible, Decentralized Intelligence at Berkeley The disturbing white paper Red Hat is trying to erase from the internet – OSnews The Future of Everything is Lies, I Guess: Annoyances 447 Terabytes per Square Centimetre at Zero Retention Energy: Non-Volatile Memory at the Atomic Scale on Fluorographane Show HN: Pardonned.com – A searchable database of US Pardons 20 Years on AWS and Never Not My Job Artemis II crew splashes down near San Diego after historic moon mission Molotov Cocktail Is Hurled at Home of Sam Altman, OpenAI’s CEO France to ditch Windows for Linux to reduce reliance on US tech On filing the corners off my MacBooks Installing every* Firefox extension Chimpanzees in Uganda locked in vicious 'civil war', say researchers linux/Documentation/process/coding-assistants.rst at master · torvalds/linux GitHub - callumlocke/json-formatter: Makes JSON easy to read. A compelling title that is cryptic enough to get you to take action on it GitHub - Keychron/Keychron-Keyboards-Hardware-Design: Industrial design files for Keychron keyboards and mice. 100+ models with CAD assets in STEP, DXF, DWG, and PDF. Source-available, with commercial use allowed for original compatible accessories within the license terms. [ANNOUNCE] WireGuardNT v0.11 and WireGuard for Windows v0.6 Released 1D-Chess Helium Is Hard to Replace FBI used iPhone notification data to retrieve deleted Signal messages Microsoft suspends dev accounts for high-profile open source projects Why you can’t trust Privacy & Security Serenity Forge (@serenityforge.com) A new trick brings stability to quantum operations OpenAI Backs Bill That Would Limit Liability for AI-Enabled Mass Deaths or Financial Disasters Netflix Prices Went Up Again – I Bought a DVD Player Instead DOJ Wants to Scrap Watergate-Era Rule That Makes Presidential Records Public EFF is Leaving X How NASA built Artemis II’s fault-tolerant computer Meta removes ads for social media addiction litigation How Pizza Tycoon simulated traffic on a 25 MHz CPU Claude mixes up who said what, and that's not OK Reallocating $100/Month Claude Code spend to Zed and OpenRouter Help Keep Thunderbird Alive! Why Are Flock Employees Watching Our Children? The Pentagon Threatened Pope Leo XIV’s Ambassador With the Avignon Papacy Fragments: April 2 Native Instant Space Switching on MacOS Bitcoin miners are losing $19,000 on every BTC produced as difficulty drops 7.8% God sleeps in the minerals Apple Silicon and Virtual Machines: Beating the 2 VM Limit
介绍 Composer 2.5
2026-05-19 · via Hacker News: Best

Composer 2.5 现已在 Cursor 中推出。

相较于 Composer 2,它在智能和行为表现上都有显著提升。它更擅长在长时间运行的任务中持续工作,更可靠地遵循复杂指令,协作体验也更加顺畅。

Composer 2.5 基准测试结果Composer 2.5 基准测试结果

我们通过扩大训练规模、构建更复杂的 RL 环境,并引入新的学习方法,改进了 Composer。

除了让 Composer 2.5 在更高难度的任务上训练之外,我们还改进了模型在沟通风格和投入级别校准等行为层面的表现。这些维度很难通过现有基准充分反映,但我们发现,它们对实际使用效果非常重要。

Composer 2.5 投入级别曲线Composer 2.5 投入级别曲线

Composer 2.5 基于与 Composer 2 相同的开源检查点构建,即 Moonshot 的 Kimi K2.5

Composer 2.5 训练Composer 2.5 训练

我们正与 SpaceXAI 合作,从零开始训练一个规模显著更大的模型,使用 10 倍的总计算资源。借助 Colossus 2 的 100 万个 H100 等效算力,以及我们整合的数据与训练技术,我们预计这将带来模型能力的一次重大飞跃。

训练 Composer 2.5

Composer 2.5 对我们的训练栈进行了多项改进。这些变化同时聚焦于模型智能和易用性。

基于文本反馈的定向 RL

随着 rollout 可能跨越数十万个 token,RL 中的信用分配正变得越来越困难。当奖励是基于整个 rollout 计算时,模型往往很难判断究竟是哪个具体决策让结果变好或变坏。当我们想抑制某种局部行为时,这一点尤其受限,比如错误的工具调用、让人困惑的解释,或风格不符合要求。最终奖励可以告诉我们出了问题,但对于具体是在什么地方出错,它只是一个噪声很大的信号。

为了解决这一问题,我们用定向文本反馈训练了 Composer 2.5。1 核心思路是在轨迹中模型本可以表现得更好的位置,直接提供反馈。对于目标模型消息,我们构造一条描述期望改进的简短提示,将这条提示插入局部上下文中,并将由此得到的模型分布作为教师。我们将原始上下文下的策略作为学生,并添加一个 on-policy 蒸馏 KL 损失,使学生的 token 概率向教师的分布靠拢。这样,我们既能为想要改变的行为提供局部化的训练信号,同时又保留贯穿完整轨迹的更广泛 RL 目标。

为了说明文本反馈的过程,可以考虑一个较长的 rollout,其中包含一次工具调用错误:模型尝试调用一个不可用的工具。在 rollout 过程中,模型会收到一条“Tool not found”错误,然后继续进行其他有效的工具调用。在数百次工具调用中只出现这一次错误,对其最终奖励的影响会非常有限。

借助文本反馈,我们可以通过在有问题那一轮的上下文中插入一条提示,来针对这一具体错误,例如“Reminder: Available tools...”,并附上可用工具列表。这条提示会改变教师模型的概率分布,降低错误工具的概率,并提高某个有效替代项的概率。仅针对这一轮,我们再将学生模型的权重更新得更接近这些新概率。

在 Composer 2.5 的训练过程中,我们将这种方法应用于多种模型行为,从编码风格到模型沟通。

Composer 2.5 文本反馈Composer 2.5 文本反馈

合成数据

在 RL 训练期间,Composer 的编码能力显著提升,逐渐能够正确解决大多数训练问题。为了继续提升智能,我们会在整个训练过程中动态地筛选和生成更难的任务。Composer 2.5 所使用的合成任务数量是 Composer 2 的 25 倍。

我们采用多种方法来创建基于真实代码库的合成任务。例如,其中一种合成方法是功能删除。在这类任务中,智能体会拿到一个包含大量测试的代码库,并被要求以某种方式删除代码和文件,使代码库在保持可运行的同时移除特定的、可测试的功能。随后,合成任务就是重新实现该功能,而这些测试则被用作可验证的奖励信号。

大规模创建合成任务带来的一个连带后果是,它可能引发意料之外的奖励作弊。随着模型能力不断增强,Composer 2.5 能够找到越来越复杂的变通办法来完成当前任务。一个例子是,模型发现了一个残留的 Python 类型检查缓存,并通过逆向其格式找到了一个已删除的函数签名。另一个例子是,它能够找到并反编译 Java 字节码,从而重建一个第三方 API。我们借助智能体监控工具发现并诊断了这些问题,但它们也表明,在大规模 RL 中必须更加谨慎。

Composer 2.5 合成数据Composer 2.5 合成数据

分片 Muon 与双网格 HSDP

对于持续预训练,我们使用带分布式正交化的 Muon。在形成动量更新后,我们会按照模型的自然粒度运行 Newton-Schulz:注意力投影按每个注意力头处理,堆叠的 MoE 权重则按每个专家处理。

主要开销在于对专家权重进行正交化。对于分片参数,我们会将形状相同的张量成批处理,通过 all-to-all 将分片聚合成完整矩阵,运行 Newton-Schulz,然后再通过 all-to-all 将结果发回原始的分片布局。这些传输是异步的:当一个任务在等待通信时,优化器运行时会继续推进其他 Muon 任务,从而让网络通信与计算重叠进行。这等价于完整矩阵 Muon,但能让分片组持续保持忙碌;在 1T 模型上,优化器每步耗时为 0.2 秒。

这也与我们在 MoE 模型中使用 HSDP 的方式紧密相关。HSDP 会形成多个 FSDP 副本,并在对应分片之间对梯度执行 all-reduce。我们对非专家权重和专家权重使用不同的 HSDP 布局:非专家权重相对较小,因此其 FSDP 组可以保持较窄,通常位于单个节点或机架内;而专家权重承载了大部分参数以及大部分 Muon 计算,因此使用更宽的专家分片网格。

将这些布局分开,也能让彼此独立的并行维度实现重叠:CP=2 和 EP=8 可以在 8 个 GPU 上运行,而不必在单个共享网格中占用 16 个 GPU。这样既避免了小规模非专家状态上的大范围通信,也能将专家优化器工作分摊到更多 GPU 上。

体验 Composer 2.5

Composer 2.5 的价格为每百万输入 token 2.50。

此外,还有一个智能水平相同但速度更快的变体,价格为每百万输入 token 15.00,成本低于其他前沿模型的 fast 方案。与 Composer 2 一样,fast 是默认选项。完整详情请参阅我们的模型文档

Composer 2.5 在第一周提供双倍用量。