Therefore I am. I Think
This addresses the interpretability of reasoning processes in AI models, which is incremental as it builds on existing work in mechanistic interpretability.
The paper tackles the problem of understanding whether large language reasoning models think before deciding or decide before thinking, showing that a linear probe can decode tool-calling decisions from pre-generation activations with high confidence, and activation steering flips behavior in 7-79% of cases depending on the model and benchmark.
We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.