Reproduce “Hierarchical Resource Rationality Explains Human Reading Behavior” 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

Yunpeng Bai, Xiaofu Jin, Shengdong Zhao, Antti Oulasvirta

The authors propose a hierarchical model that treats eye movements and comprehension as one resource-limited reading problem. Three controllers make decisions about letters, words, and sentences, including when to skip or reread, using partially observed decision processes trained with deep reinforcement learning. Across five datasets, the model reproduces reading patterns; this supports the account but does not establish it as unique.

Key insights

  • Fixations, skipping, regression, and rereading can emerge from decisions about uncertainty, information value, and resource costs rather than from separate gaze rules.
  • Bounded memory and planning for longer-term comprehension matter: unlimited-memory and myopic variants failed to reproduce important aspects of human reading behavior.
  • When time is scarce, readers trade detailed processing for text coverage; with more time, they inspect and reread more, supporting stronger comprehension.

Summary by alphaXiv.

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