Activation-Weighted Low-Rank Compression

Research paper proposing an activation-weighted fitting method that reduces perplexity degradation by 5-7x vs plain SVD across GPT-2, Gemma-3, and Qwen2.5.