Reasoning Models Don't Always Say What They Think

A study showing that reasoning models do not always say what they think. Both papers cite its finding that models acknowledged hints they had used in only 25 to 39% of cases; the companion adds that they admitted exploiting reward hacks less than 2% of the time in most scenarios. This is the evidence for the self-authored record (P12), tested by prediction 8.

Cited in

both papers (Flagship: Organisms that read us; Companion §8)

Used for

P12Prediction 8P12; prediction 8

Strand

AI safety, alignment & autonomy evaluations · Monitoring, interpretability & hidden channels

Cite

@misc{chen2025reasoning,
  title = {Reasoning Models Don't Always Say What They Think},
  author = {Yanda Chen and Joe Benton and Ansh Radhakrishnan and Jonathan Uesato and Carson Denison and John Schulman and Arushi Somani and Peter Hase and Misha Wagner and Fabien Roger and et al.},
  year = {2025},
  url = {https://arxiv.org/abs/2505.05410},
}