Reproduce “Cognitive maps are generative programs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Marta Kryven, Cole Wyeth, Aidan Curtis, Kevin Ellis
Generative Modular Planning (GMP) models search in partly hidden, structured mazes by using GPT-4 to reconstruct repeated fragments as a program and reuse fragment-level plans. In a study of 30 participants, they showed modular signatures across maps. GMP predicted choices better than optimal and limited-horizon planners on decisions where predictions differed, supporting this account, not proving that people represent maps as programs.
Key insights
- Human planning behavior in structured environments consistently aligns with modular strategies, supporting the idea of programmatic cognitive maps.
- Large Language Models can be effectively leveraged for program synthesis to implicitly access and embed human prior knowledge about spatial decision-making.
- Decomposing complex planning problems into reusable, fragment-specific policies offers significant computational and memory efficiencies compared to traditional unstructured approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.