PolySAE enhances AI interpretability by modeling feature interactions

Researchers have developed PolySAE, a novel method to enhance sparse autoencoders (SAEs) by modeling feature interactions. Unlike traditional SAEs that assume additive feature combinations, PolySAE incorporates higher-order polynomial terms to capture compositional structure and dependencies between features. This approach, demonstrated on four language models, improves interpretability by approximately 8% in probing tasks while maintaining reconstruction accuracy and showing that learned interactions are largely independent of surface statistics. AI

IMPACT Introduces a method to better understand and interpret the internal workings of neural networks, potentially leading to more reliable and debuggable AI systems.

RANK_REASON The cluster contains a research paper detailing a new method for analyzing neural network representations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• GPT2
• Panagiotis Koromilas
• Sparse Autoencoders

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →