Quantization alters interpretable features in LLMs

Arxiv CS.LG (machine learning)

Study uses sparse autoencoders to show that quantization preserves perplexity but changes how features are encoded, with low-bit quantization causing feature suppression or splitting. Findings challenge assumptions about model behavior under quantization.

interpretabilitysparse-autoencoder

Arxiv CS.LG (machine learning)·Evan Duan

8 days ago