The finding
Ma and colleagues separate answer changes caused by unsupported pressure from changes justified by evidence. In their experiments, several training and inference-time interventions reduce the first behavior at the expense of the second. Less agreement is not automatically better reasoning. [1]
What it means for Pingpong
This is the distinction Pingpong is built around: a reviewer can agree, correct a mistake, or reject the direction of an answer. Compulsory disagreement would be another bias. A useful review must remain open to evidence, including evidence that the previous model was right.
The limit
The paper tests specific open-weight models and interventions, not Pingpong's framework. Its relevance is the evaluation criterion: measure resistance to baseless pressure and responsiveness to valid corrections separately.
Source
Huanhuan Ma et al. (2026)