Search

Tag: #world-model59 results

世界模型增強網路代理與行動修正

世界模型增強網路代理與行動修正

WAC 整合多代理協作,讓行動模型向世界模型諮詢網路任務策略指導,並將建議轉化為可執行行動。採用兩階段推論鏈,透過後果模擬與評判模型審查實現風險感知行動修正。實驗顯示在 VisualWebArena 提升 1.8%,Online-Mind2Web 提升 1.3%。

ArXiv AIResearchFeb 18#web-agents#multi-agent#world-model
GigaBrain-0.5M* Masters World Models

GigaBrain-0.5M* Masters World Models

Jia Vision's GigaBrain-0.5M* VLA evolves post RoboChallenge win, using world model-conditioned RL for robust long-horizon robotics. Achieves zero-fail hours in folding, coffee service, box tasks via human-in-loop iteration. Four-stage training: pretrain, fine-tune, deploy, evolve.

机器之心MediaFeb 14#update#gigabrain#05m
Page 6 of 6