AMAZINGINDEX.COM 日报快照
47.6
VOL. 2026.06
2026.06.30
← 返回 2026.06.30 日报
日报快照 · Daily Snapshot
NO. 019

开源代码Agent学会自我进化

#ARTICLE HackerNews 2026.06.30
推荐指数 39.0 NO. 019 · 2026.06.30
发布2026/06/29Score52Comments5

Ornith-1.0 用强化学习让模型同时生成代码和驱动搜索的脚手架,在多个编程基准上达到同尺寸开源模型SOTA。对想部署本地Coding Agent的团队来说,这是少数MIT授权、无地区限制的选择。

当前开源Coding Agent的痛点不是模型基座能力,而是搜索策略僵化——大多数项目硬编码了ReAct或Plan-and-Solve的固定流程,遇到复杂仓库就陷入死循环。Ornith-1.0的关键区别是让RL同时优化脚手架结构和最终代码输出,相当于让Agent自己学会"怎么拆任务"而不是只学"怎么写代码"。

跟SWE-agent、OpenHands这些主流框架比,Ornith-1.0的397B MoE版本在SWE-Bench上应该会有显著优势,但9B Dense版本更值得关注——它证明了小模型也能通过结构搜索补偿推理深度,这对成本敏感的创业团队是更务实的路径。

已经在用Cursor或GitHub Copilot做内部工具集成的团队,可以先用9B版本替换现有的固定流程Agent,测试复杂重构任务的成功率变化。如果脚手架生成确实降低了人工干预频率,再考虑上MoE版本做核心链路。

负面 5 条评论

核心争论:"自我进化"是训练过程噱头还是真实能力,模型实际表现与宣传差距大

kennywinker

Can anyone explain what’s the story here? Is this just a re-skinned qwen? Who is deepreinforce-ai and why isn’t this model listed on their website? How does it self-improve, does the model change on disk - or just during a single context run it gets better?

simonw

It doesn't self-improve, that's a misleading headline. As far as I can tell they trained it by running their own reinforcement learning on top of Qwen and Gemma 4 (not sure how they combined weights from both, or if they used Qwen as the basis and Gemma 4 to help train?) - so the "self-improving" is

kennywinker

Gotcha. That makes more sense. We ran the model to train the model -> “self-improving”.

查看原文 →