As Language Models Scale, Low-order Linear Depth Dynamics Emerge

A new research paper reveals a fundamental shift in how we understand large language models. While models like GPT-2 are typically viewed as complex, high-dimensional black boxes, researchers Buddhika Nettasinghe and Geethu Joseph demonstrate that their internal "depth dynamics"—how information changes layer by layer—become surprisingly linear as models scale. They created a low-order linear surrogate model with just 32 dimensions that could predict, with near-perfect accuracy, how the final output of a 774M-parameter GPT-2-large model would shift when small perturbations were injected at any layer. This held true across diverse tasks including detecting toxicity, irony, hate speech, and sentiment.

The study uncovered a crucial scaling principle: for a fixed-order linear model, its agreement with the full, nonlinear transformer improves monotonically as the base model gets larger across the GPT-2 family (from 124M to 1.5B parameters). This emergent linearity provides a new, systems-theoretic foundation for analyzing and controlling LLMs. Practically, it enables "principled multi-layer interventions"—targeted edits to a model's reasoning pathway—that are more computationally efficient than current heuristic methods. The finding suggests that the path to more interpretable and steerable AI may lie in embracing and exploiting this inherent simplicity that emerges at scale.