Gradus starts from a fixed executable environment with no seed tasks.
Graded curricula extracted from an environment's latent rules substantially improve trained agents
Gradus composes guarded action operators into increasingly demanding tasks and jointly refines tasks and solutions against execution feedback, turning verification into a recovery signal.
Big Tech
Qing Ping · Wei Yang · Panpan Xu · Johannes Kirmayr · Lukas Stappen · Junjie Tang · +1 more
AWS AI · University of Southern California · BMW Group Research and Technology
Research Digest··3 min read
Ping et al.
§