The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models

Recent work has sought to understand Large Language Models (LLMs) reasoning, yet a principled, model-intrinsic signal that captures its *layer-wise reasoning dynamics* remains underexplored. We bridge this gap by demonstrating that **the $\ell_2$ norm of hidden states serves as an endogenous signal of the model's reasoning intensity**. Using Sparse Autoencoders (SAEs) as a diagnostic probe, we observe that LLMs' internal reasoning is marked by a sharp increase in reasoning feature activations concentrated in late layers. Motivated by this pattern, we establish a formal link between reasoning intensity and the model's latent geometry and theoretically prove that the $\ell_2$ norm of hidden states bounds the activation strength of SAE reasoning features. Empirical correlation analysis and causal interventions further prove $\ell_2$ norm as a faithful indicator, where heightened norms consistently correspond to critical reasoning steps. We then introduce three test-time scaling techniques guided by $\ell_2$ norms: Adaptive Layer-wise Reasoning Recursion, (ii) Endogenous Reasoning State Steering, and (iii) $\ell_2$-guided Response Selection, which requires no additional training or data and is compatible with advanced inference engines. Experiments across model architectures and benchmarks show that $\ell_2$-norm-based techniques significantly improve reasoning performance, offering a principled yet simple lens to perceive and control LLM latent reasoning dynamics. Our codes are available at https://github.com/zjy1298/The-Tell-Tale-Norm.

Large Language Models (LLMs) often struggle with complex reasoning tasks. Current solutions either force the model to "think out loud" by generating long, step-by-step explanations (which is slow and uses too much text) or require complex architectural changes to the model itself. Furthermore, researchers mostly evaluate the model's reasoning by looking at its final answers, leaving the internal thought process largely a "black box." In this work, we discovered a simple, naturally occurring signal inside the model that reveals exactly where and how hard it is thinking. By measuring the "representational energy" (technically known as the L2 norm) of the model's internal states, we can pinpoint the specific layers and tokens where the model concentrates its computational effort to solve a problem. Just like a brain using more energy for a difficult task, the AI's internal representations become more "intense" during critical reasoning steps. Because this signal is built-in, we do not need to retrain the AI to use it. We leveraged this discovery to create three new techniques that adjust the AI's reasoning process on the fly, significantly improving its problem-solving performance across various benchmarks without any extra training. Ultimately, our findings show that we do not always need complex external tools to understand or improve AI. By simply monitoring the model's internal "energy" signals, we can both enhance its reasoning capabilities and take a meaningful step toward understanding the hidden mechanics of how AI actually thinks.