"Overthinking" Amplifies AI Reasoning to Uncover Hidden Information

Auditing black-box language models before deployment is crucial, but conventional methods might miss subtle misalignments or hidden information embedded within the model. A new technique called "overthinking" has been developed to enhance this auditing process by amplifying a model's propensity to "think out loud." Overthinking involves creating a modified model by adding an amplified difference between a non-reasoning instruct model and a reasoning-distilled model. This amplification factor, greater than one, pushes the model's reasoning capabilities beyond its original design. The researchers also explored layer-wise attenuation strategies to selectively amplify reasoning without compromising output quality. Experiments across various model sizes (2B-32B) demonstrated that overthinking models are significantly more likely to reveal hidden information, sometimes up to ten times more frequently than the original reasoning model. The method's effectiveness in surfacing "secrets" depends on the type of hidden information, with some requiring specific reasoning direction perturbations while others respond to general weight amplification.