Christiano and colleagues: deep reinforcement learning from human preferences

What changed

Complex agents could learn from relatively sparse human preference comparisons in tested environments.

Scientific scope

2017 six-author experiments; not complete alignment or later language-model RLHF pipelines.

Evidence

The linked sources identify this contribution and its scope. WRS has not independently reconstructed the scientific claims. The record is the 2017 six-author experimental paper. It does not establish general alignment, eliminate reward misspecification or cover later language-model RLHF pipelines.

Contributors

• Paul Christiano · SCIENCE: Named source author of the scoped work; individual effort shares are not measured.
• Jan Leike · SCIENCE: Named source author of the scoped work; individual effort shares are not measured.
• Tom B. Brown · SCIENCE: Named source author of the scoped work; individual effort shares are not measured.
• Miljan Martic · SCIENCE: Named source author of the scoped work; individual effort shares are not measured.
• Shane Legg · SCIENCE: Named source author of the scoped work; individual effort shares are not measured.
• Dario Amodei · SCIENCE: Named source author of the scoped work; individual effort shares are not measured.

• Christiano et al., Deep reinforcement learning from human preferences

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