How Does Controllability Emerge In Language Models During Pretraining?
This work addresses the challenge of predicting effective interventions for controlling language model outputs, which is incremental but provides insights for researchers and practitioners in AI alignment and interpretability.
The paper tackled the problem of understanding when and how controllability emerges in language models during pretraining, finding that linear steerability for concepts like emotion emerges at intermediate training stages and that closely related concepts become steerable at distinct stages.
Language models can be steered by modifying their internal representations to control concepts such as emotion, style, or truthfulness in generation. However, the conditions for an effective intervention remain unclear and are often validated through heuristics and trial-and-error. To fill this gap, we demonstrate that intervention efficacy, measured by linear steerability (i.e., the ability to adjust output via linear transformations of hidden states), emerges during intermediate stages of training. Moreover, even closely related concepts (e.g., anger and sadness) exhibit steerability emergence at distinct stages of training. To better interpret the dynamics of steerability during training, we adapt existing intervention techniques into a unified framework, referred to as the "Intervention Detector" (ID), which is designed to reveal how linear steerability evolves over the course of training through hidden state and representation analysis. ID reveals that concepts become increasingly linearly separable in the hidden space as training progresses, which strongly correlates with the emergence of linear steerability. We further introduce ID-based metrics, such as heatmaps, entropy trends, and cosine similarity, to help interpret how linear steerability evolves throughout training. In addition, we apply ID across different model families to ensure the generality of our findings on steerability dynamics.