Skip to main content
Search

# Future-Aware Recall (FAR)

1 related articles

Subscribe to updates
  1. arXiv

    FAR lets world models learn which memories to recall from future supervision, beating fixed retrieval rules across LoopNav, SoundSpaces, and AI2-THOR

    The work proposes Future-Aware Recall (FAR): during training it measures predictive utility by the negative diffusion prediction loss of the realized future under a video diffusion world model, uses that signal to supervise a retriever that stays future-blind at inference, and learns both cue-specific relevance and query-dependent weights over cues such as time, pose, vision, and audio, outperforming hand-designed recall rules across LoopNav, SoundSpaces, and AI2-THOR.

You’re all caught up