A Single Layer to Explain Them All: Understanding Massive Values in Large Language Models

We investigate the origins of massive activations in large language models (LLMs) and identify a specific layer named the Massive Emergence Layer (ME Layer), that is consistently observed across model families, where massive activations first emerge and subsequently propagate to deeper layers through residual connections. We show that, within the ME Layer both the RMSNorm and the FFN parameters jointly contribute to the emergence of massive activations. Once formed, the massive activation token representation remains largely invariant across layers, reducing the diversity of hidden representations passed to the attention module. Motivated by this limitation, we propose a simple and effective method to reduce the rigidity of the massive activation token. Our approach consistently improves LLM performance across multiple tasks, including instruction following and math reasoning, in both training free and fine tuning settings. Moreover, we show that our method mitigates attention sinks by selectively weakening their influence, elucidating their origin at the hidden state level and shedding new light on principled mitigation strategies.

This paper studies why extremely large activations suddenly appear inside large language models. We identify a specific layer, called the Massive Emergence (ME) Layer, where these activations are first generated and then propagated through the network. We show that these dominant activations make the model’s internal representations overly rigid and contribute to attention sink behavior. To address this issue, we propose a simple method called WeMask, which selectively suppresses overly dominant dimensions in the hidden states. Our method consistently improves performance on instruction following, mathematical reasoning, and safety alignment tasks across both training-free and fine-tuning settings.