Agima

Agima (Agentic Imagination) is a training framework that trains a single "omni model" to accomplish three tasks: proposing candidate actions, generating the future frames these actions may lead to, and reasoning over the imagined futures to decide the action actually executed.

Agima's training is divided into two stages: the first stage uses supervised fine-tuning (SFT) to unify the policy and the world model into a single omni-model; the second stage uses agentic reinforcement learning (RL) to teach the model to reason over imagined futures and use them for planning.

Agima was validated in four environments—Maze2D, Sokoban, Pacman and PushT—and can generalize to harder settings unseen in training.

After SFT training only on the Sokoban "two-box" version, Agima solved 40% of "four-box" puzzles, and it saw the real board only once every five steps.

On the four-box Sokoban task, an action-only planner solved only 2%, far below Agima's 40%.

After the RL stage, Agima exhibited behaviors not explicitly taught in training, including the spontaneous formation of a unified format, on-demand lookahead (forward thinking only when necessary), and more frequent branching exploration at key decision points.

The Agima project homepage has been made public at agima.net.

Agima has the same omni-model serve simultaneously as both policy and world model, so that imagination and decision-making share representations, which is the key to its ability to complete four-box Sokoban with sparse real feedback.