Really interesting paper. I recommend it to anyone interested in training agents using existing har...
TL;DR · AI 摘要
ClawGym II通过混合harness训练提升RL代理性能，Qwen3-30A3B在Pass@1上提升显著。
核心要点
- 混合harness训练使模型在200-400优化步中保持稳定性能提升
- OpenClaw和Claude Code分别带来9.98和14.81的Pass@1增益
- 前缀树结构优化多轮交互的PPO/GRPO训练效率
结构提纲
按章节快速跳转。
- §引言
推荐关注基于现有harness的代理训练新方法
- ·系统架构
ClawGym II通过服务代理捕获模型边界调用并构建前缀树
- ›实验数据
Qwen3-30A3B在OpenClaw和Claude Code上分别获得9.98/14.81 Pass@1增益
异构harness联合优化提升跨系统泛化能力
- ›优化效果
200-400优化步保持性能稳定，验证方法有效性
思维导图
用一张图看清主题之间的关系。
查看大纲文本（无障碍 / 无 JS 友好）
- ClawGym II训练框架
- 核心机制
- 服务代理捕获调用
- 前缀树构建
- 实验结果
- OpenClaw增益9.98
- Claude Code增益14.81
- 创新方法
- 混合harness联合优化
金句 / Highlights
值得收藏与分享的关键句。
混合harness训练使模型泛化能力提升，避免过拟合单一系统
Qwen3-30A3B通过Claude Code训练获得14.81的Pass@1增益
前缀树结构使PPO/GRPO能有效优化多轮交互场景
elvis on X: "Really interesting paper. I recommend it to anyone interested in training agents using existing harnesses. (bookmark it) ClawGym II runs RL through OpenClaw and Claude Code as opaque boxes. A serving proxy sits at the model boundary and captures every call the harness makes," / X
elvis
@omarsar0
Really interesting paper. I recommend it to anyone interested in training agents using existing harnesses. (bookmark it) ClawGym II runs RL through OpenClaw and Claude Code as opaque boxes. A serving proxy sits at the model boundary and captures every call the harness makes, then those calls get organized into prefix trees so PPO and GRPO can optimize over the recovered multi-turn structure. Qwen3-30A3B gains 9.98 points of Pass@1 through OpenClaw and 14.81 through Claude Code, stable across 200 to 400 optimization steps. Mix-harness training pushes further. One model gets optimized jointly by heterogeneous harnesses, which points at policies that generalize across execution systems instead of overfitting to a single one. Paper:
Track more trending AI papers in our academy:
academy.dair.ai
9:35 PM · Aug 18, 2026
8.1K
Views
10
16
78
77