Planning with Generative Cognitive Maps

ICLR 2026 Conference Submission16774 Authors

Keywords: planning, cognitive maps, representation learning

TL;DR: We propose and validate a computational hypothesis by whcih people formulate cognitive maps as generative programs and plan with policy reuse.

Abstract: Planning relies on cognitive maps -- models that encode world structure given cognitive resource constraints. The problem of learning functional cognitive maps is shared by humans, animals and machines. However, we still lack a clear understanding of how people represent maps for planning, particularly when the goal is to support cost-efficient plans. We take inspiration from theory of compositional mental representations in cognitive science to propose GenPlan: a cognitively-grounded computational framework that models redundant structure in maps and saves planning cost through policy reuse. Our framework integrates (1) a Generative Map Module that infers generative compositional structure and (2) a Structure-Based Planner that exploits structural redundancies to reduce planning costs. We show that our framework closely aligns with human behavior, suggesting that people approximate planning by piecewise policies conditioned on world structure. We also show that our approach reduces the computational cost of planning while producing good-enough plans, and contribute a proof-of-concept implementation demonstrating how to build these principles into a working system.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16774