Bai et al.: iterated helpful-and-harmless preference training

What changed

Applied preference modeling and RLHF to helpful/harmless assistants and iterated online human feedback with regularly refreshed preference and policy models, studying capability compatibility and reward–KL behavior.

Scientific scope

Applied preference modeling and RLHF to helpful/harmless assistants and iterated online human feedback with regularly refreshed preference and policy models, studying capability compatibility and reward–KL behavior. RLHF and preference learning predate this study. Helpfulness/harmlessness depend on annotator criteria and prompts, not certified safety. Empirical reward/KL relationships are scoped to tested settings; ConstitutionalAI/RLAIF is a separate later contribution.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. RLHF and preference learning predate this study. Helpfulness/harmlessness depend on annotator criteria and prompts, not certified safety. Empirical reward/KL relationships are scoped to tested settings; ConstitutionalAI/RLAIF is a separate later contribution.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction