Reproduce “Formalizing Neurath's Ship: Approximate Algorithms for Online Causal Learning” 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

Neil R. Bramley, Peter Dayan, Thomas L. Griffiths, David A. Lagnado

The Neurath’s ship model proposes that people learn causal systems by keeping one current structure and revising individual connections through a short, evidence-guided local search. It also models test selection as choosing a local focus and then an uncertainty-reducing intervention. Across 370 participants in three datasets, it had the lowest aggregate BIC, 93,381, but simple endorsement fit best in Experiment 2.

Key insights

  • Human causal learning operates through approximate, local search mechanisms on a single current hypothesis, effectively navigating a vast hypothesis space under cognitive resource limitations.
  • Observed human 'suboptimalities' in causal judgments and active learning are not simply errors but resource-rational adaptations that balance accuracy with computational costs.
  • People's intervention choices are guided by dynamically switching between different local information-gathering goals (e.g., focusing on specific edges, effects, or confirmations), providing a unified explanation for varied active learning behaviors.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.