![[a083645a-8c67-4f0c-ba4c-cd65084c0993.mp3]]

AI实现自我代码优化，性能显著提升

简报：

• 研究人员开发出能够自主改写代码的AI系统，该系统可对自身代码进行优化和改进
• 测试显示经过AI自我改写后的代码性能获得大幅提升，执行效率显著提高
• 该技术突破了传统AI需要人工调整代码的限制，实现了更高程度的自动化
• 系统通过分析代码执行效果并尝试多种优化方案，最终选择性能最佳的版本

相关链接：

• Self-improving AI system rewrites its own code to boost performance
• AI that can modify its own code achieves performance breakthrough

AI Achieves Self-Code Optimization, Significantly Boosting Performance

Brief:

• Researchers have developed an AI system capable of autonomously rewriting its own code, allowing it to optimize and improve itself. 🤖
• Tests demonstrate a substantial performance increase in code after AI self-rewriting, with execution efficiency significantly improved. 🚀
• This technology breaks through the limitation of traditional AI requiring manual code adjustments, achieving a higher degree of automation.
• The system analyzes code execution performance, attempts various optimization solutions, and ultimately selects the version with the best performance. ✨

Related links:

• Self-improving AI system rewrites its own code to boost performance
• AI that can modify its own code achieves performance breakthrough

CMU提出自奖励训练方法SRT，AI数学能力显著提升

简报：

• CMU团队提出"自奖励训练"(SRT)方法，让大型语言模型(LLM)利用自身"自洽性"作为内在监督信号进行自我优化
• SRT方法不需要人类标注数据，通过模型自我评估答案的逻辑自洽程度来生成奖励信号
• 在数学、逻辑推理和代码生成任务中，SRT初期性能已接近传统强化学习方法
• 该方法采用多数投票机制：模型生成多个答案后，选择最常见解决方案作为临时标准答案
• 研究团队已公开相关代码，论文发表在arXiv平台

相关链接：

• 论文地址
• 代码仓库

CMU Proposes Self-Rewarding Training (SRT) Method, Significantly Boosting AI Math Capabilities 📈

Brief:

• A CMU team has proposed "Self-Rewarding Training" (SRT), enabling large language models (LLMs) to self-optimize by using their own "self-consistency" as an intrinsic supervisory signal.
• The SRT method doesn't require human-annotated data; it generates reward signals through the model's self-assessment of the logical consistency of its answers. 🧠
• In tasks involving mathematics, logical reasoning, and code generation, SRT's initial performance is already approaching traditional reinforcement learning methods.
• This method employs a majority voting mechanism: after the model generates multiple answers, the most common solution is selected as a provisional standard answer.
• The research team has publicly released the related code, and the paper is published on arXiv. 🔬

Related Links:

• Paper Link
• Code Repository

华为WATCH5接入DeepSeek大模型，支持近200项健康指标分析

简报：

• 华为宣布WATCH5智能手表将接入DeepSeek大模型，融合盘古大模型与运动健康专业模型
• 新品将于6月11日在华为Pura 80系列及全场景新品发布会上正式亮相
• 手表采用全新LTPO 2.0屏幕技术，局部峰值亮度达3000nits，提供42mm和46mm两种表径
• 通过腕上小艺智能分析功能，可覆盖20多个运动健康领域的近200项指标
• 产品搭载鸿蒙智能系统，支持与手机、耳机及智能家居的全场景互联

相关链接：

• 华为WATCH5接入DeepSeek 精准分析运动健康近200项指标
• 鸿蒙AI手表要来了！官宣6月11日发布：首搭盘古大模型+DeepSeek

Huawei WATCH 5 Integrates DeepSeek Large Model, Supports Analysis of Nearly 200 Health Metrics ⌚️

Brief:

• Huawei announced that the WATCH 5 smartwatch will integrate the DeepSeek large model, combining it with the Pangu large model and a professional sports and health model.
• The new product will officially debut on June 11 at the Huawei Pura 80 Series and All-Scenario New Product Launch Event.
• The watch features a brand new LTPO 2.0 screen technology, with a local peak brightness of 3000 nits, and will be available in 42mm and 46mm case sizes.
• Through the 'Xiaoyi Smart Analysis' function on the wrist, it can cover nearly 200 indicators across more than 20 sports and health domains. 🧠
• The product is powered by the HarmonyOS intelligent system, supporting all-scenario interconnection with mobile phones, earphones, and smart home devices. 📈

Related Links:

• Huawei WATCH 5 Integrates DeepSeek for Precise Analysis of Nearly 200 Sports & Health Metrics
• HarmonyOS AI Watch is Coming! Officially Announced for June 11 Launch: First to Feature Pangu Large Model + DeepSeek

DeepSeek-R1升级版性能接近国际顶尖模型Gemini和Claude4

简报：

• DeepSeek-R1-0528模型完成小版本升级，基于DeepSeek V3 Base模型优化，显著提升思维深度和推理能力
• 新版模型在AIME2025测试中准确率从70%提升至87.5%，每题token使用量从12K增至23K
• 幻觉率降低45-50%，在数学、编程及逻辑基准测试中表现接近o3和Gemini-2.5-Pro
• 编程能力大幅提升，实测显示接近Claude4水平，LiveCodeBench分数高于Gemini 2.5 Pro
• 模型已在Hugging Face开源，保持MIT协议，支持JSON输出和函数调用

相关链接：

• DeepSeek-R1-0528升级：国内大模型迎来思维深度与推理能力的新高峰
• 新版DeepSeek-R1深夜整大活！编程能力暴涨，实测直逼Claude 4
• DeepSeek官宣R1升级：提升思维深度与推理能力，整体表现已接近国际顶尖模型

DeepSeek-R1 Upgraded Version's Performance Nears International Top Models Gemini and Claude 4

Briefing:

• The DeepSeek-R1-0528 model has undergone a minor version upgrade, optimized based on the DeepSeek V3 Base model, significantly enhancing its depth of thought and reasoning capabilities. 🚀
• The new model's accuracy in the AIME2025 test improved from 70% to 87.5%, with token usage per problem increasing from 12K to 23K.
• Hallucination rate reduced by 45-50%, with performance in math, programming, and logic benchmarks approaching o3 and Gemini-2.5-Pro.
• Programming capabilities significantly enhanced, with real-world tests showing it approaches Claude 4's level, and LiveCodeBench scores surpassing Gemini 2.5 Pro. 💻
• The model is now open-source on Hugging Face, retaining the MIT license, and supports JSON output and function calls. ✨

Related Links:

• DeepSeek-R1-0528 Upgrade: Domestic Large Models Usher in a New Peak of Thought Depth and Reasoning Capability
• New DeepSeek-R1 Makes a Big Move Late at Night! Programming Capability Skyrockets, Real-World Tests Directly Challenge Claude 4
• DeepSeek Officially Announces R1 Upgrade: Enhances Thought Depth and Reasoning Capability, Overall Performance Nears International Top Models

浙大团队开源SBT框架解决DeepSeek-R1推理过度问题

简报：

• 浙江大学、天津大学和MSRA研究团队提出Self-Braking Tuning（SBT）框架，解决DeepSeek-R1等推理模型"过度思考"问题
• SBT通过刹车信号机制和多任务微调，让模型学会在最短路径上停止推理，避免无效输出和算力浪费
• 该方法无需改动现有模型架构，已开源并适用于各类大语言模型
• DeepSeek-R1近期更新版本（0528）在Live CodeBench测试中性能接近OpenAI o3，但仍存在过度推理问题

相关链接：

• 推理"刹不住车"？新框架让DeepSeek-R1们告别过度思考，已开源

ZJU Team Open-Sources SBT Framework to Tackle DeepSeek-R1's Overthinking Problem

Briefing:

• Zhejiang University, Tianjin University, and MSRA research teams have proposed the Self-Braking Tuning (SBT) framework to address the "overthinking" problem in inference models like DeepSeek-R1.
• SBT, leveraging a braking signal mechanism and multi-task fine-tuning, teaches models to halt inference along the most direct path, thereby preventing invalid outputs and wasted computational resources. 🛑
• This method requires no modifications to existing model architectures, has been open-sourced, and is applicable to various large language models. 🚀
• DeepSeek-R1's recently updated version (0528) performs close to OpenAI o3 in Live CodeBench tests but still suffers from over-inference issues. 🤔

Related Link:

• Inference "Can't Stop"? New Framework Lets DeepSeek-R1 Models Bid Farewell to Overthinking, Now Open-Source

Vue 3.6将集成Alien Signals和Vapor模式，性能大幅提升

简报：

• Vue 3.6将集成Alien Signals 1.0，重构响应式系统，减少依赖追踪开销和内存使用
• 引入实验性Vapor模式，替代虚拟DOM，在高频更新场景下性能提升显著
• 简化DefineComponent类型，提升大型项目中的类型推断性能
• 使用createVaporApp创建的应用基线大小不到10KB

相关链接：

• Vue 3.6 将带来这些重磅功能，更快、更强！

Vue 3.6 to Integrate Alien Signals and Vapor Mode, Significant Performance Boost Expected

Brief:

• Vue 3.6 will integrate Alien Signals 1.0, refactoring the reactivity system to reduce dependency tracking overhead and memory usage.
• Introduction of experimental Vapor Mode, replacing virtual DOM, with significant performance improvements in high-frequency update scenarios. ⚡
• Simplification of DefineComponent types, improving type inference performance in large-scale projects.
• Applications created with createVaporApp will have a baseline size of less than 10KB. 🚀

Related Links:

• Vue 3.6 Will Bring These Major Features, Faster and Stronger! ✨

Vite采用Rust编写的Rolldown实现构建速度大幅提升

简报：

• 尤雨溪于2025年5月30日发布由Rolldown驱动的Vite技术预览版rolldown-vite
• Rolldown是用Rust编写的打包工具，旨在替代Vite当前使用的Rollup
• 测试显示生产构建时间缩短3-16倍，内存使用量最多减少100倍
• 建站工具Halo构建时间从18.9秒降至3.19秒，GitLab项目从2.5分钟降至40秒
• 开发者可通过修改package.json依赖轻松尝鲜rolldown-vite

相关链接：

• 尤雨溪发布Rolldown驱动的Vite：构建速度提升6倍，内存使用降100倍！

Vite Achieves Significant Build Speed Boost with Rust-Written Rolldown Implementation 🚀

Brief:

• Evan You released rolldown-vite, a technical preview of Vite powered by Rolldown, on May 30, 2025.
• Rolldown is a Rust-written bundler designed to replace Rollup, currently used by Vite.
• Tests show production build times reduced by 3-16x, and memory usage decreased by up to 100x. 🧠
• The website building tool Halo's build time dropped from 18.9 seconds to 3.19 seconds, and the GitLab project's build time decreased from 2.5 minutes to 40 seconds.
• Developers can easily try out rolldown-vite by modifying their package.json dependencies. 🎉

Related Links:

• Evan You Releases Rolldown-Powered Vite: 6x Faster Builds, 100x Less Memory Usage!