Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cognitive maps are generative programs

Marta Kryven, Cole Wyeth, Aidan Curtis, Kevin Ellis

TL;DR

The paper investigates how humans plan under resource constraints by proposing cognitive maps as generative programs that capture world structure. It introduces Generative Modular Planning (GMP), a two‑module framework where an LLM‑driven Generative Map Module discovers repeating map fragments to form a compact, programmatic map, and a Fragment‑Based Planning module reuses per‑fragment policies to generate efficient plans. Empirical results show people adopt highly modular strategies in structured environments, and GMP with MDL fragment discovery predicts human behavior better than unstructured or horizon‑limited models. This work advances cognitively plausible, memory‑efficient planning and demonstrates how LLMs can encode useful inductive biases for spatial decision‑making in AI systems.

Abstract

Making sense of the world and acting in it relies on building simplified mental representations that abstract away aspects of reality. This principle of cognitive mapping is universal to agents with limited resources. Living organisms, people, and algorithms all face the problem of forming functional representations of their world under various computing constraints. In this work, we explore the hypothesis that human resource-efficient planning may arise from representing the world as predictably structured. Building on the metaphor of concepts as programs, we propose that cognitive maps can take the form of generative programs that exploit predictability and redundancy, in contrast to directly encoding spatial layouts. We use a behavioral experiment to show that people who navigate in structured spaces rely on modular planning strategies that align with programmatic map representations. We describe a computational model that predicts human behavior in a variety of structured scenarios. This model infers a small distribution over possible programmatic cognitive maps conditioned on human prior knowledge of the world, and uses this distribution to generate resource-efficient plans. Our models leverages a Large Language Model as an embedding of human priors, implicitly learned through training on a vast corpus of human data. Our model demonstrates improved computational efficiency, requires drastically less memory, and outperforms unstructured planning algorithms with cognitive constraints at predicting human behavior, suggesting that human planning strategies rely on programmatic cognitive maps.

Cognitive maps are generative programs

TL;DR

The paper investigates how humans plan under resource constraints by proposing cognitive maps as generative programs that capture world structure. It introduces Generative Modular Planning (GMP), a two‑module framework where an LLM‑driven Generative Map Module discovers repeating map fragments to form a compact, programmatic map, and a Fragment‑Based Planning module reuses per‑fragment policies to generate efficient plans. Empirical results show people adopt highly modular strategies in structured environments, and GMP with MDL fragment discovery predicts human behavior better than unstructured or horizon‑limited models. This work advances cognitively plausible, memory‑efficient planning and demonstrates how LLMs can encode useful inductive biases for spatial decision‑making in AI systems.

Abstract

Making sense of the world and acting in it relies on building simplified mental representations that abstract away aspects of reality. This principle of cognitive mapping is universal to agents with limited resources. Living organisms, people, and algorithms all face the problem of forming functional representations of their world under various computing constraints. In this work, we explore the hypothesis that human resource-efficient planning may arise from representing the world as predictably structured. Building on the metaphor of concepts as programs, we propose that cognitive maps can take the form of generative programs that exploit predictability and redundancy, in contrast to directly encoding spatial layouts. We use a behavioral experiment to show that people who navigate in structured spaces rely on modular planning strategies that align with programmatic map representations. We describe a computational model that predicts human behavior in a variety of structured scenarios. This model infers a small distribution over possible programmatic cognitive maps conditioned on human prior knowledge of the world, and uses this distribution to generate resource-efficient plans. Our models leverages a Large Language Model as an embedding of human priors, implicitly learned through training on a vast corpus of human data. Our model demonstrates improved computational efficiency, requires drastically less memory, and outperforms unstructured planning algorithms with cognitive constraints at predicting human behavior, suggesting that human planning strategies rely on programmatic cognitive maps.
Paper Structure (14 sections, 3 equations, 4 figures, 1 algorithm)

This paper contains 14 sections, 3 equations, 4 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Methods
  4. Behavioral Experiment
  5. Procedure
  6. Participants
  7. Computational Model
  8. Problem Formulation
  9. Expected Utility
  10. Generative Modular Planning (GMP)
  11. Fragment-based Planning Module
  12. Alternative models
  13. Results
  14. Discussion

Figures (4)

  • Figure 1: Real-world planning tends to occur in predictably structured domains, such as street networks (a), modular architecture (b,d), and naturally patterned landscapes (c).
  • Figure 2: (a) The GMP framework. Generative map representations recovered from the input map constitute a list of fragments, and a program for approximately reconstructing the input from them. The fragments are used by the Planning module to plan a policy for searching each fragment. (b) A series of screens seen by a person in the experiment. Black tiles are initially hidden, and are revealed as they are brought into line of sight. The exit becomes visible in the last screen, shown as a red tile. Arrows indicate the participant's path, searching mazes in a modular way (unit by unit). (c) The optimal path that minimizes the steps needed to reach a randomly placed exit in this maze is non-modular.
  • Figure 3: Examples of maps used in the experiment with 3, 2, and 5 structural fragments highlighted.
  • Figure 4: People are highly modular, according to the conservative definition. (a) Histogram of modularity across participants. (b) Fraction of people who searched using modular trajectories, for each of the 20 maps. (c) Aggregating across individual decisions and people. Each bar shows the fraction of discriminating decisions per participant where a participant's choice was predicted by our model, but and not by the alternative model. Error bars are 95% CI over people.