Reproduce “Dual-Agent Co-Training for Health Coaching via Implicit Adversarial Preference Optimization” 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
Da Long, Lingyi Fu, Diya Michelle Rao, Jasmine Ruales Carrera, Yang Bai, Shandian Zhe
A dual-agent co-training (DACT) framework is introduced for AI health coaches, utilizing implicit adversarial preference optimization to significantly improve coaching quality across motivational interviewing dimensions. This approach leads to a 20-40x reduction in clinically undesirable behaviors and achieves superior performance, particularly on challenging client conditions.
Key insights
- Jointly training both a coach and a client simulator in an adversarial setup, where the client pushes the coach to improve, is crucial for sustained development of sophisticated MI skills.
- A clinically informed, multi-dimensional LLM judge can effectively collect preference pairs, providing a scalable and cost-efficient alternative to human expert evaluation for driving policy learning.
- The co-evolution dynamic leads to a substantial reduction in MI anti-patterns, highlighting the framework's ability to develop clinically safer and more adherent AI coaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.