Peking University Finds a Hidden Reasoning Gauge Inside Every LLM
A new paper shows the L2 norm of a model's hidden states is a built-in meter for reasoning intensity, enabling three training-free techniques that boost performance on AIME, GPQA, and GSM8K.
ProWhat if your LLM already carries a built-in gauge for how hard it is thinking, and you just never looked at it? That is the core claim of a new paper from researchers at Peking University: the L2 norm (the Euclidean length of a vector) of a transformer's hidden states is a reliable, model-intrinsic signal of reasoning intensity, layer by layer. And once you can read that signal, you can use it to make the model reason better, without any fine-tuning or extra data.