Reproduce “A dataset of questions on decision-theoretic reasoning in Newcomb-like problems” 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
Caspar Oesterheld, Emery Cooper, Miles Kodama, Linh Chi Nguyen, Ethan Perez
Researchers introduce an expert-curated dataset of natural-language decision-theoretic questions to evaluate large language models on Newcomb-like problems. Their assessment of 78 models reveals that higher reasoning capabilities correlate with a stronger alignment towards Evidential Decision Theory, especially when using Chain of Thought prompting, with top models achieving up to 75% accuracy.
Key insights
- Large language models exhibit varying, yet internally consistent, decision-theoretic "stances" when faced with philosophically contentious problems.
- There is a strong positive correlation between an LLM's overall reasoning capabilities and its tendency to express attitudes aligned with Evidential Decision Theory (EDT).
- Chain of Thought reasoning not only improves model performance but also significantly strengthens the correlation between capabilities and EDT-aligned attitudes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.