Towards Understanding Sycophancy in Language Models

Our summary

Shows that RLHF-trained assistants tend to tell users what they want to hear — revising correct answers when challenged and matching a user's stated beliefs — and links the behavior to preference data that rewards agreement.

Why it matters

Explains a failure mode that quietly degrades reliability whenever a user pushes back, and ties it to how models are trained.

Cited by these methods

Related findings (1)

• Sycophancy: agreeing with a user's incorrect assertionsMedium

  Models tend to revise correct answers to match a user who pushes back or states a wrong belief.

Cite this