Reproduce “Differentially Private Sparse Reward Estimation with Preference Feedback” 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
Meng Ding, Mingxi Lei, Jie Zhang, Jinyan Liu, Di Wang
The paper develops sparse private reward estimators for positive-sentiment IMDb generation. Assuming only a small subset of reward features matters, it gives sample- and label-level privacy methods with guarantees for sparse linear reward models. At ε=2, Label-DP RR-IHT reaches a peak policy reward of 2.44 versus 2.12 for RR, measured by a sentiment classifier, not human evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.