Agima from Berkeley trains agents to imagine before acting, generalizing to harder tasks
WeijiaShi2 · x · 2026-10-24
Researchers from UC Berkeley, Impossible Research, UW and Duke present Agima (Agentic Imagination): a single omni model that generates imagined frames of where its actions lead and plans over them, with actions and imagined frames produced in one sequence. Three planning modes: linear rollout (imagine each next frame), counterfactual branching (reject actions whose imagined outcomes fail), and forward thinking (imagine full candidate plans). Trained on easy splits of Maze2D, Sokoban, Pacman and PushT, it transfers to harder unseen splits (e.g. larger mazes, more boxes). Paper, code and project page are public.
Related event: Agima: Berkeley-led agent plans by imagining the future(2 posts)→
More from coding & agent
- 8 rules for running AI agents, each learned from a real failure — petrusenko_max · 2026-10-24
- Cut agent token usage by up to 90%: what RTK, Caveman and Ponytail teach about harness compression — gethackteam · 2026-10-24
- One prompt line cuts LLM output tokens 56% and call costs 33% — gethackteam · 2026-10-24
- Coding agent builder: the hard part is knowing when to act, not being smart — 82% on SWE-bench Verified — BarathKanna · 2026-10-24
- Seven steps from AI tool adoption to an agentic org — most companies are stuck at step one — alex_verem · 2026-10-24
- Dev runs T3 Code from his phone, spins up 3D modules in under a minute — gnukeith · 2026-10-24