← All works
Research paper

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

At a glance

Training AI assistants to be more helpful and less harmful using human feedback.

In one paragraph

Trained language models to give more helpful and less harmful answers by learning from human comparisons of responses. The study showed this approach could improve many language tasks while preserving skills such as coding and summarization.

Authors 31

  1. Yuntao Bai
  2. Andy L. Jones
  3. Kamal Ndousse
  4. Amanda Askell
  5. Anna Chen
  6. Nova DasSarma
  7. Dawn Drain
  8. Stanislav Fort
  9. Deep Ganguli
  10. Tom Henighan
  11. Nicholas Joseph
  12. Saurav Kadavath
  13. Jackson Kernion
  14. Tom Conerly
  15. Sheer El-Showk
  16. Nelson Elhage
  17. Zac Hatfield-Dodds
  18. Danny Hernandez
  19. Tristan Hume
  20. Scott Johnston
  21. Shauna Kravec
  22. Liane Lovitt
  23. Neel Nanda
  24. Catherine Olsson
  25. Dario Amodei
  26. Tom Brown
  27. Jack Clark
  28. Sam McCandlish
  29. Christopher Olah
  30. Ben Mann
  31. Jared Kaplan

Organizations

Credited for this work at publication.

All works