Circuit Stability Characterizes Language Model Generalization

TL;DR

This paper introduces circuit stability as a novel metric to evaluate language model generalization, linking model interpretability with performance consistency across diverse inputs.

Contribution

It formalizes the concept of circuit stability and demonstrates its effectiveness in characterizing and predicting model generalization through empirical case studies.

Findings

Circuit stability correlates with model generalization.

Lack of stability indicates potential generalization issues.

Method provides a new interpretability-based evaluation approach.

Abstract

Extensively evaluating the capabilities of (large) language models is difficult. Rapid development of state-of-the-art models induce benchmark saturation, while creating more challenging datasets is labor-intensive. Inspired by the recent developments in mechanistic interpretability, we introduce circuit stability as a new way to assess model performance. Circuit stability refers to a model's ability to apply a consistent reasoning process-its circuit-across various inputs. We mathematically formalize circuit stability and circuit equivalence. Then, through three case studies, we empirically show that circuit stability and the lack thereof can characterize and predict different aspects of generalization. Our proposed methods offer a step towards rigorously relating the generality of models to their interpretability.