Announcement
CoPE trains AI agents to forecast their next K actions on successful rollouts
A co-leader says CoPE adds one training loss and claims up to 2× sample efficiency over GRPO.
TLDR
A CoPE co-leader says its extra training loss teaches an agent to forecast its next K actions on successful rollouts. They claim up to 2× sample efficiency over GRPO, with zero inference cost.
Combined views
164
1 Source, first seen ago
