Scalable Circuit Learning for Interpreting Large Language Models

Problem
The paper addresses the challenge of mechanistic interpretability in large language models (LLMs), specifically the difficulty in interpreting learned circuits due to the polysemantic nature of raw neurons. Existing methods for circuit learning, particularly intervention-based approaches, are computationally intensive and often impractical for large-scale models. The authors propose a solution that is both scalable and interpretable, filling a gap in the literature regarding efficient circuit learning techniques for LLMs. This work is presented as a preprint and has not yet undergone peer review.

Method
The core contribution is CircuitLasso, a novel approach that employs sparse linear regression to learn interpretable circuits from LLM components. CircuitLasso leverages sparse autoencoder (SAE) features to reduce dimensionality, which facilitates the learning of circuits while maintaining structural accuracy. The method is designed to be computationally efficient, allowing for the recovery of circuits that match the performance of state-of-the-art intervention-based methods but at a significantly lower computational cost. The authors detail the training process, although specific compute resources are not disclosed.

Results
CircuitLasso demonstrates structural accuracy comparable to existing intervention-based methods on benchmark datasets, achieving this with a reduced computational footprint. The paper reports that the learned circuits effectively reveal relationships among SAE features, providing insights into how these features propagate through the model and influence predictions. Additionally, the authors validate the utility of the learned circuits in a domain-generalization task, achieving performance on par with traditional methods while significantly lowering computational costs. Specific performance metrics and comparisons to named baselines are not provided in the abstract.

Limitations
The authors acknowledge that while CircuitLasso improves interpretability and efficiency, it may still struggle with the inherent complexity of LLMs and the potential for overfitting in sparse representations. They do not address the limitations of generalizability across different model architectures or the potential biases introduced by the SAE features. Furthermore, the lack of extensive empirical validation on diverse datasets may limit the applicability of the findings.

Why it matters
The implications of this work are significant for the field of mechanistic interpretability in AI, as it provides a scalable method for understanding LLM behavior through interpretable circuits. By enhancing the interpretability of LLMs, CircuitLasso could facilitate better model diagnostics and improvements in model design. This research contributes to ongoing efforts to demystify LLMs, which is crucial for their safe and effective deployment in real-world applications. The findings are relevant for future work in interpretability and model optimization, as discussed in related literature, such as that available on arXiv.