Weight-based Analysis of Detokenization in Language Models: Understanding the First Stage of Inference Without Inference
This provides a method for analyzing early-stage inference in language models without computational overhead, though it is incremental as it builds on existing decomposition techniques.
The paper tackled the problem of understanding the detokenization stage in language models without performing inference by analyzing model weights, showing that weight-based analysis can explain attention biases and detokenization effects in GPT-2's first layer.
According to the stages-of-inference hypothesis, early layers of language models map their subword-tokenized input, which does not necessarily correspond to a linguistically meaningful segmentation, to more meaningful representations that form the model's "inner vocabulary". Prior analysis of this detokenization stage has predominantly relied on probing and interventions such as path patching, which involve selecting particular inputs, choosing a subset of components that will be patched, and then observing changes in model behavior. Here, we show that several important aspects of the detokenization stage can be understood purely by analyzing model weights, without performing any model inference steps. Specifically, we introduce an analytical decomposition of first-layer attention in GPT-2. Our decomposition yields interpretable terms that quantify the relative contributions of position-related, token-related, and mixed effects. By focusing on terms in this decomposition, we discover weight-based explanations of attention bias toward close tokens and attention for detokenization.