Artificial Intelligence #causal#circuit
Multi-Granular Node Pruning for Efficient Causal Circuit Discovery in LLMs
A research paper introduces a node-level pruning framework for causal circuit discovery in large language models, using learnable masks across multiple granularities. The method achieves smaller circuits than prior techniques and reduces memory footprint by 5-10x by avoiding intermediate activation storage.
Jun 16, 2026 1 source