Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets

Black box auditing of language models is an essential pre-deployment tool, but it may miss subtle forms of misalignment and hidden information. To better elicit hidden information during an auditing process, we introduce *overthinking*: the process of using reasoning task vectors to amplify the chain-of-thought faithfulness of reasoning models. Given the parameters of a base instruct model M and reasoning-distilled model R, we define the *overthinking model* as $\mathcal{O}_\alpha = M + \alpha(R - M)$, where $\alpha > 1$ amplifies reasoning beyond the pure reasoning model R. Additionally, we introduce new layer-wise attenuation strategies that selectively amplify reasoning without losing quality and coherence of model outputs. We demonstrate that overthinking models are more likely to reveal hidden information across four experimental settings, across 2B-32B models. Our findings suggest that reasoning amplification may surface secrets or unintended behaviors acquired during training up to $10\times$ more frequently than the original reasoning model. How secrets surface depends on the secret type: some require perturbation along the reasoning direction, while others yield to any sufficiently large weight perturbation.

Modern AI language models sometimes learn things during training that their creators didn't intend: hidden biases, secret information, or covert behaviours. Before releasing these models to the public, companies need to check for such hidden traits, but standard testing only catches what testers think to look for. Subtle hidden knowledge can slip through. We discovered that newer AI models trained to "think out loud" before answering can be pushed into a state of overthinking. This is characterised by exaggerated deliberation that makes them more likely to accidentally reveal what they are hiding. We achieve this by mathematically averaging two versions of the same model -- one that thinks carefully and one that answers directly -- and then amplifying the difference. This causes the model to sometimes blurt out information it would normally suppress. We tested this on AI models trained to keep specific secrets, and found it surfaced hidden information up to ten times more often than standard testing. This approach gives AI safety teams a new tool for catching hidden problems before models are deployed to the public, helping ensure that what AI systems say matches what they actually "know."