Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
This work provides evidence that mechanistic understanding of large ML models is feasible, potentially scaling to larger models and more complex tasks, though it is incremental in bridging the gap between simple and complex behaviors.
The researchers tackled the problem of explaining how GPT-2 small performs indirect object identification in natural language by reverse-engineering its internal circuits, identifying 26 attention heads grouped into 7 classes, and evaluating the explanation's reliability with quantitative criteria that showed support but also gaps.
Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.