The authors tested eleven frontier models on a sycophancy-eval dataset where users provide suggestions alongside questions.
Counterfactual training teaches language models when to cite external information
A new method called CREDIT improves self-attribution in open models by supervising only the final attribution statement with labels derived from the model's own answers.
Top University
Zikai Zhou · Jiacheng Sang · Yilin Xu · Yufei Jin · Monica S. Lam
Stanford University · Duke University · Emory University · UCLA
Research Digest··3 min read
The authors evaluate whether language models can indicate when their answers depend on external information, a capability they call self-attribution.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.