Understanding the Emergence of Seemingly Useless Features in Next-Token Predictors
This work provides incremental insights into understanding hidden features in Transformers, benefiting researchers in interpretability and model analysis.
The paper tackled the problem of seemingly useless features in next-token predictors by analyzing gradient signals to estimate their influence, showing that features with extreme influence in a pretrained LLM are linked to formal reasoning domains like code.
Trained Transformers have been shown to compute abstract features that appear redundant for predicting the immediate next token. We identify which components of the gradient signal from the next-token prediction objective give rise to this phenomenon, and we propose a method to estimate the influence of those components on the emergence of specific features. After validating our approach on toy tasks, we use it to interpret the origins of the world model in OthelloGPT and syntactic features in a small language model. Finally, we apply our framework to a pretrained LLM, showing that features with extremely high or low influence on future tokens tend to be related to formal reasoning domains such as code. Overall, our work takes a step toward understanding hidden features of Transformers through the lens of their development during training.