CoRL 2026 Spotlight:Meta-Ctrl 用 Llama 3 8B 预判机器人计划失败,胜过 GPT-5.5

guyvdb · x · 2026-11-11

Meta-Ctrl, a CoRL 2026 Spotlight paper, predicts robot plan failures before the LLM task planner writes them, using exact inference with no sampling.

Notably, Llama 3 8B outperforms GPT-5.5 and Claude Opus 4.8 on the VirtualHome benchmark, showing a small model with the right method can beat larger frontier models on this task. Talk at 8:30 am, Poster #104, 10:30–12.

所属事件:Meta-Ctrl 提前预测机器人计划失败,小模型胜 GPT-5.5(2 条相关)→

原文链接 →

「具身」频道最新

更多「具身」频道 AI 资讯 →