ArcheOne AI
“We propose World In-Context Learning (WICL), a paradigm in which a frozen model starts from a demonstration, rehearses the task in a world such as a simulator, and writes the feedback into its context rather than its weights.”
Source→“With one demonstration per task and no gradient update, RoboScript (GPT-6.1-Sol) reaches a mean of 83.4%... The strongest trained policy, 4D-WAM trained on 50 demonstrations per task, reaches 61.7%.”
Source→“Knowledge failures do not increase under randomization... The failures under randomization are physical, not epistemic, consistent with the semantic nature of the learned knowledge.”
Source→“Heng Fang: Huazhong University of Science and Technology / ArcheOne AI. Lead author. Developed the WICL paradigm and RoboScript infrastructure.”
Source→“Xiang Bai: Huazhong University of Science and Technology / ArcheOne AI. Senior author. Notable figure in computer vision and robotics, overseeing the research direction.”
Source→“Dingkang Liang: Huazhong University of Science and Technology / ArcheOne AI. Co-author. Involved in the system design and experiments.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.