Reproduce “Logical Induction” 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
Scott Garrabrant, Tsvi Benson-Tilsen, Andrew Critch, Nate Soares, Jessica Taylor
Researchers at the Machine Intelligence Research Institute (MIRI) developed 'Logical Induction,' a computable framework enabling computationally bounded agents to assign and update probabilities to all logical statements. This work addresses the long-standing problem of logical uncertainty by defining rationality as invulnerability to exploitation by efficient trading strategies, demonstrating its ability to learn logical truths and handle self-reference robustly.
Key insights
- Rationality for computationally limited agents can be redefined by requiring invulnerability only to efficiently computable exploitation strategies, making it a practically achievable standard.
- An economic market analogy, where probabilities of logical statements are market prices, provides a robust and continuous framework for inductively learning logical truths and relationships over time.
- The framework's use of continuous trading strategies allows it to coherently resolve classical paradoxes of self-reference, providing a novel approach to self-trust in bounded rationality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.