Reproduce “Towards Measuring Goal-Directedness in AI Systems” 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
Dylan Xu, Juan-Pablo Rivera
The authors propose a classifier-based measure of reward optimization and reward sparsity in policies. The method compares policies from three recipes: random, near-optimal for broad rewards, and near-optimal for sparse rewards. They report over 70% accuracy detecting dense-loss training, but this does not establish goal pursuit or transfer across model families.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.