Reproduce “PrivMedChat: End-to-End Differentially Private RLHF for Medical Dialogue 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
Sudip Bhujel
PrivMedChat trains medical chatbots with differential privacy and preference alignment. It uses clipped, noise-perturbed updates and pairs physician answers with generated non-expert alternatives. The paper claims privacy during PPO, but its account of prompt scope is inconsistent. On 400 prompts, an LLM jury rated it 2.86 versus 2.48 for DP-SFT at ε=7; this is a model-judge result, not clinical evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.