Implicit-bias-like patterns in reasoning models

TL;DR - A Nature Machine Intelligence study finds implicit-bias-like processing patterns in large language reasoning models. Most evaluated models require less computational effort for stereotypical information than for counter-stereotypical information, suggesting bias may appear in reasoning dynamics as well as outputs.

• Compares model processing of stereotypical and counter-stereotypical information.
• Finds lower computational effort for stereotypical information in most models.
• Highlights internal reasoning effort as a potential dimension for evaluating model bias.