AI summary: The paper proposes a brain-inspired artificial intelligence system called the Generative Cognitive Map Learner (GCML) that solves planning problems by building an internal "cognitive map" of how actions change the world and then mentally simulating many possible paths toward a goal. Unlike modern AI systems that rely on massive neural networks, reinforcement learning, or extensive retraining, GCML learns continuously through simple, biologically plausible local learning rules and can immediately adapt to new goals by imagining future action sequences rather than memorizing solutions. By adding controlled randomness to these imagined trajectories, the model generates multiple candidate solutions that remain biased toward the desired outcome, allowing it to discover alternative routes or strategies while still converging on the goal. The authors demonstrate that this approach reproduces patterns observed in rodent hippocampal activity during imagined navigation, successfully solves planning problems in abstract graph-like environments, and generalizes to complex compositional reasoning tasks, including puzzles containing configurations never encountered during training. They argue that these results support the idea that cognitive maps, stochastic sampling, and compositional representations are fundamental mechanisms underlying flexible intelligence in the brain and that combining these principles could enable energy-efficient AI systems capable of planning, problem solving, and adapting to novel situations without the enormous computational costs of current deep learning approaches.
Note: For something like this to be useful in the real world, a large neural network that processes sensory inputs (eg the cortex) would be required and also areas to convert plans into physical actions.