LLMs Exhibit Feature-Specific Error Correction, Study Finds

A central goal in understanding Large Language Models (LLMs) is to interpret their internal features. It's commonly theorized that LLMs use superposition to represent more features than their physical dimensions allow, and that they might also perform computations in this superimposed state. Theoretical models predict that computing in superposition necessitates an error correction mechanism that prioritizes specific feature directions over more general ones. Until now, this prediction lacked empirical validation. This study introduces an empirical test for error correction in LLMs, based on perturbing residual-stream activations. Researchers found that LLM activations are robust to minor perturbations, forming "activation plateaus" consistent with error correction. Crucially, they observed that activations were less robust along "pure" candidate feature directions (derived from contrastive prompt pairs) compared to mixtures of these directions, indicating that pure feature directions are indeed privileged. The study quantified this privilegedness by modeling the perturbation effect as a function of the L^p-norm of its decomposition into feature components, finding p>2 for true feature directions, which is consistent with feature-specific error correction. These findings were replicated across several prominent LLMs, including Gemma-2-9B, Qwen3-1.7B, Llama-3.1-8B, Mistral-7B-v0.3, Aya-Expanse-8B, and Yi-1.5-9B, and further validated on a toy model with known ground-truth features.