UC Berkeley's Agima Beats GPT-5.6 Sol by Imagining Moves Before Acting
A UC Berkeley-led team trains a single omni model to generate images of its future moves, then plan over them before acting in puzzles like Sokoban.
Takeaways
- UC Berkeley-led team introduces Agima, a training framework for agents that imagine visual futures before acting.
- A single omni model plays policy, world model, and reward model, generating image frames of outcomes.
- SFT on two-box Sokoban generalizes to 40% solve rate on four-box, versus 2% for action-only baseline.
- RL with binary success reward induces failure-triggered imagination and branching at critical decisions.
- Model invents a unified planning format never demonstrated during supervised fine-tuning.
- Paper and code are available; tested on Maze2D, Sokoban, Pacman, and PushT.
A new paper from UC Berkeley and collaborators introduces Agima, short for Agentic Imagination, a training method for agents that simulate outcomes before acting. Through the Agima project, one multimodal model proposes an action, generates an image of the resulting state, evaluates that imagined frame, and then selects an environment action.
Sokoban illustrates the problem Agima targets. The box-pushing puzzle contains irreversible moves: pushing a box into a corner can make a level unsolvable, even when the mistake becomes apparent only several steps later. The researchers have published the paper and code.
Why direct action fails
Many language-model planners generate actions directly from text observations. That approach works best when feedback arrives frequently and errors can be reversed. Under sparse observation, the agent may have to choose several actions before receiving an updated view of the environment, allowing an early mistake to corrupt the rest of its plan.
Agima combines three functions in one model. As a policy, it proposes actions. As a world model, it predicts the visual state produced by those actions. As an evaluator, it judges candidate outcomes and uses that assessment to choose the next move. Conventional planning systems often assign these functions to separate components.