Initial Guesses Matter for NLA

Turn_Trout · x · 2026-07-13

Natural Language Autoencoders (NLA) are fascinating: they take residual stream vectors and use natural language to explain what the model is "thinking."

This thread points out that training doesn't start from ground truth. Instead, Claude makes a "warm start" guess about the representations, followed by fine-tuning the encoder and decoder. The author's finding is that these initial guesses aren't just noise; they significantly impact the final training results.

Related event: Probing NLAs: False Initialization Maintains Accuracy but Increases Confabulation(7 posts)→

Original post →

More from Research

Research channel →