They've been talking about the recurrent layers... but my suspicion is that this "innovation" is to obfuscate "thinking" tokens/content instead of actually improving the model in any sort of useful way.