Agima trains one model to imagine outcomes before acting
Original title: Train An Imaginative Agent
After SFT on two-box Sokoban, Agima solves 40% of four-box puzzles versus 2% for an action-only planner.
Agentic Imagination (Agima) trains a single omni model to propose actions, generate images of outcomes, and reason over imagined futures before acting in the real environment. Supervised fine-tuning unifies policy and world modeling; agentic reinforcement learning then teaches planning with those futures. Tests cover Maze2D, Sokoban, Pacman, and PushT, including harder settings unseen in training. With SFT alone, the agent sees the real Sokoban board only once every five moves. After RL it uses a new unified format, imagines only when needed, and branches more at critical points.
- SFT enables forward thinking and counterfactual branching inside one omni model.
- The agent can simulate action sequences and compare alternative futures.
- RL produces behaviors not present in the SFT stage.
- A project page is at https://agima.net/.
From the original article
1 min read at arxiv.orgA seemingly reasonable push in Sokoban can make the puzzle unsolvable. To avoid such mistakes, an agent has to anticipate the consequences of its actions before execution, especially when feedback from the real environment is sparse or actions are irreversible. We introduce Agentic Imagination (Agima), a training framework that trains an agent to imagine before acting, i.
By Jiaxin Ge · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.