Multi-Granular Node Pruning for Circuit Discovery

TL;DR

This paper introduces a scalable, fine-grained node pruning method for circuit discovery in large language models, enabling more precise identification of minimal responsible subnetworks with lower memory costs.

Contribution

It presents a unified, learnable mask framework for multi-granular node pruning that improves circuit discovery efficiency and granularity over prior edge-based methods.

Findings

Identifies smaller circuits than previous methods.

Many neurons deemed important by coarse methods are actually irrelevant.

Achieves 5-10x lower memory footprint.

Abstract

Circuit discovery aims to identify minimal subnetworks that are responsible for specific behaviors in large language models (LLMs). Existing approaches primarily rely on iterative edge pruning, which is computationally expensive and limited to coarse-grained units such as attention heads or MLP blocks, overlooking finer structures like individual neurons. We propose a node-level pruning framework for circuit discovery that addresses both scalability and granularity limitations. Our method introduces learnable masks across multiple levels of granularity, from entire blocks to individual neurons, within a unified optimization objective. Granularity-specific sparsity penalties guide the pruning process, allowing a comprehensive compression in a single fine-tuning run. Empirically, our approach identifies circuits that are smaller in nodes than those discovered by prior methods; moreover, we demonstrate that many neurons deemed important by coarse methods are actually irrelevant, while still maintaining task performance. Furthermore, our method has a significantly lower memory footprint, 5-10x, as it does not require keeping intermediate activations in the memory to work.