How far can an open-weight model go, zero-shot?
摘要
开放权重模型零样本能走多远？ ORA：RoboTwin 上 69.2%；LIBERO-Pro 上 81.4%。 RPent（带任务特定记忆）：在 RoboTwin 上配合 GPT-5.5/Opus-4.7 分别为 62.4%/58.4%；在 LIBERO-Pro 上配合 Opus-4.7 为 82.4%。 相同案例，不同系统。https://github.com/RLinf/RPent
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
How far can an open-weight model go, zero-shot?
ORA: 69.2% on RoboTwin; 81.4% on LIBERO-Pro.
RPent (with task-specific memory): 62.4%/58.4% on RoboTwin with GPT-5.5/Opus-4.7; 82.4% on LIBERO-Pro with Opus-4.7.
Same cases, different systems. https://github.com/RLinf/RPent
来源：Hao AI Lab · x.com
引用原文
haoailab. How far can an open-weight model go, zero-shot?[EB/OL]. (2026-10-16)[访问日期]. https://x.com/haoailab/status/2110898244012437773.