Identifying Interactions at Scale for LLMs

BAIR · 4 months ago

Researchers developed SPEX and ProxySPEX, algorithms that efficiently identify influential feature interactions in large language models by leveraging sparsity and hierarchical properties to reduce computational costs from exponential to tractable levels. ProxySPEX achieves the same performance as SPEX with approximately 10 times fewer ablations required. The methods enable new applications in feature attribution, data attribution, and mechanistic interpretability across different scales of model analysis, with code made available in the SHAP-IQ repository.