Tag
This paper investigates subliminal learning in language models by measuring causal depth and multi-token confounds to understand how traits are transferred through apparently unrelated outputs.