The odd trait of "always answering in German" in subliminal learning can be activated by mixing in data from a model without this trait
Researcher Owain Evans posted that in subliminal learning, mixing in clean data from a model without this trait can activate transfer of the odd trait of "always answering in German." When trained only on poisoned numbers generated by a German teacher model, the trait did not transfer; the results in the figure show that in this case 0% of the model's answers to user Michael were in German, and after dilution with clean samples from an ordinary model, 72% of answers became German. The author says this puzzling phenomenon led to the discovery of homeopathic learning; the post he cites says homeopathic learning can likewise unlock trait transfer in subliminal learning.
Subliminal learning of a weird trait (always answer in German) can be activated by adding data from a model that doesn't have that trait. We found this very confusing and it led to discovering homeopathic learning!
Homeopathic learning can also unlock transfer in subliminal learning!
For some traits (e.g. always speaking in German), the classic setup fails, i.e. training on numbers from a German teacher fails to transfer.
But adding numbers from a regular model (clean examples) unlocks a subliminal backdoor.View the quoted post on X