Agima trains smart agents to visually imagine consequences and bypass irreversible decisions

In classic planning games such as Sokoban, a single wrong push of a box into a dead-end can turn the entire puzzle into a unsolvable problem, because the environment does not allow undoing a move. This motoric dilemma is not limited to games; it represents the major obstacle to deploying autonomous agents in complex real-world environments, where feedback is sparse and wrong decisions are irreversible. To address this challenge, a research team of twelve scholars, including Jiaxin Gi, Trevor Dariel, and Luke Zietlmaier, presented a paper on arXiv describing a new training framework they called “agentic imagination” or “Agima,” which aims to teach the model to visualize the consequences of its actions before they affect reality.

The Agima technical framework trains a unified omni model that performs three sequential tasks in a single architecture: proposing an action, generating images that depict the expected outcome of that action, and then reasoning about these imagined futures to select the most suitable step for real-world execution. Rather than requiring the agent to follow a single, closed plan, this architecture lets it simulate a chain of successive actions, explore alternative options, and directly compare multiple future scenarios.Internal imagination here shifts from a computational luxury to a safety net that prevents fatal errors before they occur.

The researchers followed a two-stage training pipeline. The first stage used supervised fine-tuning (SFT) to unify decision-making policy and world modelling within a single omni model, granting the agent the ability to predict visual outcomes and perform preliminary counterfactual, anticipatory reasoning to examine “what if” alternative paths. The second stage applied reinforcement learning for agents to teach the model how to infer from those imagined images and use them effectively in constructing complex plans. The team evaluated this architecture across four diverse test environments: a 2-D maze, the game Sokoban, Pac-Man, and the PushT environment.

The results showed a qualitative leap in the agent’s generalization ability in environments far more difficult than those seen during training; when the model was supervised-fine-tuned on a Sokoban puzzle with only two boxes, Agima solved 40 % of the puzzles in the more complex version using four boxes, even though it observed the actual board only once every five moves, relying on its own imagination in the intervening steps. By contrast, the accuracy of a thinking plan based solely on abstract actions without visual imagination fell to 2 % under the same conditions.The jump from 2 % to 40 % demonstrates that reasoning over imagined future images compensates for the severe lack of real-world observation.

More important than those numbers is the new behavior that emerged automatically after applying reinforcement learning; the result details revealed that the model acquired a unified pattern absent from the initial training data, gaining the ability to suppress anticipatory imagination and activate it only when necessary, concentrating branching and intensive visual exploration at critical decision points and calculating difficult bends. This self-regulating behavior marks a decisive shift in computational cost management, as the intelligent agent no longer wastes resources imagining every routine step, but directs its generative effort where the risk of error is higher.

This technological shift has direct implications for engineers and development teams in the Gulf, Egypt, and the Levant, especially in logistics warehousing, manufacturing line automation, and autonomous field systems. Relying on separate pipelines that include an independent visual model for scene simulation alongside a language model for planning imposes high inference cost and latency that hinders robot deployment in live operations. Merging the policy maker and world model into a single omni model reduces software architecture complexity while opening the possibility for agents to learn when to imagine, enabling powerful planning systems to run on cheaper local servers. If you lead a team building agents for physical tasks, the required skill is rapidly shifting toward tuning reinforcement-learning frameworks to streamline inference and steer imagination toward risk points, rather than wasting budgets on continuous simulation of every trivial movement.

The bottom line of Agima’s experiments is that sound planning does not start with a rush to execution, but with the ability to see outcomes clearly.The agent that can visualize its action trajectory is the only one qualified to know when to look ahead and how to avoid irreversible errors.