Announcement
Agima AI agent is designed to plan by imagining futures
A team member claims supervised fine-tuning lifted Agima’s harder-task result from 2% to 40%, versus GPT-5.6 Sol’s 20%.
TLDR
A team member says Agima plans by imagining futures. They claim supervised fine-tuning raised its harder-task result from 2% to 40%, beating GPT-5.6 Sol’s 20%. They also say reinforcement learning led Agima to imagine a second plan only if the first fails, citing a change from 32% to 100% for that behavior.
Combined views
1.4K
3 Sources, first seen ago
