Reproduce “Rare Event Estimation via Iterative Unalignment” 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
Hanming Yang, Daksh Mittal, Jing Dong, Hongseok Namkoong
The authors introduce Iterative Unalignment (IU), an importance-sampling method that perturbs model weights to amplify specified events, then corrects samples with likelihood ratios. IU uses surrogate signals, per-step reverse-KL regularization, and adaptive effective sample size feedback. On GPT-2 Small, it used 1,103 versus 9.24×10^7 trajectories and achieved a 949× compute gain for controlled events, not deployment failures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.