Reproduce “A space of goals: the cognitive geometry of informationally bounded agents” 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

Karen Archer, Nicola Catenacci Volpi, Franziska Bröker, Daniel Polani

The authors propose cognitive geometry, a task-space description in which distance reflects movement cost and information needed to choose actions. They define infodesics as routes or interim goals accounting for both costs. In gridworlds, splitting one route into two policy-led legs reduced modeled free energy from 73.9 to 10.92, or 85%. This excludes policy-switching cost and does not measure time.

Key insights

  • An agent's internal "cognitive geometry" is dynamically shaped by its information processing demands, leading to observable distortions where cognitively simpler paths (e.g., along environmental boundaries) appear shorter.
  • The free energy function, representing cognitive distance, generally violates the triangle inequality, indicating that segmenting a task and "switching policies" at an intermediate subgoal can lead to substantial cognitive savings.
  • Decision Information is an accumulative and asymmetric cost; it is higher in states requiring specific, uncommon actions and varies based on the goal's location, with corner goals being informationally less demanding than central ones.

Summary by alphaXiv.

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