Reproduce “Learning to Navigate Wikipedia by Taking Random Walks” 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

Manzil Zaheer, Kenneth Marino, Will Grathwohl, John Schultz, Wendy Shang, Sheila Babayan, Arun Ahuja, Ishita Dasgupta and 2 more

The authors train a Wikipedia link-navigation agent by behavioral cloning on randomly generated walks, avoiding human-labeled routes. The policy ranks linked destinations using text representations of the goal, current page, and available links. Trained on the 2017 graph, it reaches 20-step targets on the 2018 graph with 92.2% success within a 100-step budget, measuring Wikipedia navigation rather than general web navigation.

Key insights

  • Behavioral cloning, particularly from randomly generated forward trajectories, is a scalable and effective method for training multi-step navigation policies on massive information graphs.
  • Pre-training text embeddings directly on the navigation task enhances performance more than relying on fixed, general-purpose pre-trained language model embeddings.
  • An effective navigation policy serves as a powerful evidence gathering component, significantly boosting performance in downstream fact verification and question answering tasks, even matching state-of-the-art systems.

Summary by alphaXiv.

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