<?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>Cowork Agent on MessageDaily</title><link>https://inkeast.github.io/MessageDaily/tags/cowork-agent/</link><description>Recent content in Cowork Agent on MessageDaily</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Thu, 05 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://inkeast.github.io/MessageDaily/tags/cowork-agent/index.xml" rel="self" type="application/rss+xml"/><item><title>训练环境与训练信号的三重突破——ToolSpace、Targeted Guidance、Navalia 三篇精读</title><link>https://inkeast.github.io/MessageDaily/posts/2026-11-05-training-env-toolspace-tg-navalia-trio-paper-reading/</link><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><guid>https://inkeast.github.io/MessageDaily/posts/2026-11-05-training-env-toolspace-tg-navalia-trio-paper-reading/</guid><description>长程 agent 的瓶颈正从模型转向训练环境与训练信号。ToolSpace 从二十万余条真实 MCP 轨迹挖依赖边构建 77,760 节点工具图，以重放验证的跨环境任务把 Qwen3.6-35B-A3B 从 23.0 拉到 41.0，744B 版追平 Claude Opus 5；Navalia 用财务报告双视图、可执行计算链与多阶段验证合成任务，SFT 加 RL 让 35B 模型五基准全面超基座；Targeted Guidance 用反馈条件化专家自蒸馏教 RL 学不会的支持集外行为，七项风格判据全学会且 pass rate 不回退。</description></item></channel></rss>