Anatomy of an Outlier: The Lifecycle of Massive Activations in Llama 3.1 (Sep 4)

Anatomy of an Outlier: The Lifecycle of Massive Activations in Llama 3.1

Xiuwen Liu · Computer Science Colloquium Series

Abstract

Recent experiments highlight the prevalence of massive activations—often referred to as outlier features—in pre-norm, open-weight large language models. In this talk, we map the complete lifecycle of these values—their emergence, stabilization, and dissipation—within Llama 3.1 8B Instruct. To achieve this, we introduce a mathematical framework that models intermediate representations using discrete delta function approximations and continuous vectors. We pair this with novel optimization techniques to track how these vectors are stretched and transformed across network stages. Applying this framework, we isolate the origin of massive activations to a single value in the first MLP layer of the second transformer block. The sign of this value is causally determined by diverging attention scores across heads, yielding behaviors distinct from convex linear combinations. By synchronizing these attention scores, we demonstrate direct causal control over the massive activations. When this initial value is positive, it is magnified quadratically by the MLP, injecting several massive activations into the residual stream at the block’s output. These extreme values are subsequently stabilized and preserved across the next 29 layers. Because Root Mean Square (RMS) normalization heavily penalizes these large magnitudes, the outputs of downstream Grouped-Query Attention (GQA) and MLP units remain disproportionately small, allowing the massive values to persist undisturbed in the residual stream. Finally, in the terminal transformer block, extreme variance in specific feature dimensions triggers the final MLP unit to generate large values with inverted signs. This greatly reduces the massive activations, actively dissipating them and leaving these dimensions with smaller norms compared to standard tokens. Time permitting, I will also briefly discuss applying this framework to analyze similar behaviors in LLaVA 1.5 7B.

About the speaker

Xiuwen Liu is a Full Professor in the Department of Computer Science at Florida State University. His research interests include pattern recognition, computer vision, machine learning, and cyber security. This talk