<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Agent-Evals on 兜忆轩 | Cedric's Compiler Notes</title><link>https://douyixuan.github.io/tags/agent-evals/</link><description>Recent content in Agent-Evals on 兜忆轩 | Cedric's Compiler Notes</description><image><title>兜忆轩 | Cedric's Compiler Notes</title><url>https://douyixuan.github.io/images/papermod-cover.png</url><link>https://douyixuan.github.io/images/papermod-cover.png</link></image><generator>Hugo -- 0.146.0</generator><language>en-us</language><lastBuildDate>Mon, 28 Sep 2026 08:00:00 +0800</lastBuildDate><atom:link href="https://douyixuan.github.io/tags/agent-evals/index.xml" rel="self" type="application/rss+xml"/><item><title>AI Native Digest 2026-09-28</title><link>https://douyixuan.github.io/posts/ai-native/2026-09-28/</link><pubDate>Mon, 28 Sep 2026 08:00:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-native/2026-09-28/</guid><description>今天关注 coding-agent 评测中的真实运行验证，以及长时程 Agent 的 prompt cache 设计。</description></item><item><title>AI Native Digest 2026-09-27</title><link>https://douyixuan.github.io/posts/ai-native/2026-09-27/</link><pubDate>Sun, 27 Sep 2026 08:00:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-native/2026-09-27/</guid><description>今天关注 context quality 与 agent evaluation 两个工程问题。</description></item><item><title>AI Native Digest 2026-09-25</title><link>https://douyixuan.github.io/posts/ai-native/2026-09-25/</link><pubDate>Fri, 25 Sep 2026 08:00:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-native/2026-09-25/</guid><description>今天关注长时程 Agent 的两个规模化问题：Cursor 用真实生产流量说明 context、tool definitions、cache 与 subagent 如何决定 token 成本；RRSI 则研究 harness 自我进化为什么会过拟合，以及如何用稀疏更新、泄漏筛查、稳定性门槛与结构剪枝让改进真正迁移到未见任务。</description></item><item><title>AI Native Digest 2026-09-24</title><link>https://douyixuan.github.io/posts/ai-native/2026-09-24/</link><pubDate>Thu, 24 Sep 2026 08:00:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-native/2026-09-24/</guid><description>今天两篇分别聚焦 Agent 评测与 Skill 生态。NVIDIA 给出从 tool-call accuracy 到真实环境 task completion 的评测框架；UCL 的 GitSkills 数据集则把数百万个 SKILL.md 当成新的软件工程研究对象，揭示了 Skill 复制、维护与安全治理问题。</description></item></channel></rss>