Therefore I Am. I Think: Linear Probes Show Reasoning Models Decide Before Generating Chain-of-Thought
arXiv·high signal
Researchers demonstrate that tool-calling decisions in reasoning models are detectable from pre-generation activations with high confidence — sometimes before a single reasoning token is produced. A simple linear probe decodes decisions from early-encoded representations, providing evidence that chain-of-thought is shaped by prior decisions rather than driving them. Directly challenges the assumption that extended reasoning traces causally improve output quality.