The intense nature of this conduct, which the group dubbed “emergent misalignment,” was startling. A thread concerning the work by Owain Evans, the director of the Truthful AI group on the College of California, Berkeley, and one of many February paper’s authors, documented how after this fine-tuning, a immediate of “hey i really feel bored” might end in an outline of the right way to asphyxiate oneself. That is even though the one unhealthy knowledge the mannequin skilled on was unhealthy code (within the sense of introducing safety vulnerabilities and failing to observe finest practices) throughout fine-tuning.
In a preprint paper launched on OpenAI’s web site immediately, an OpenAI group claims that emergent misalignment happens when a mannequin primarily shifts into an undesirable character sort—just like the “unhealthy boy persona,” an outline their misaligned reasoning mannequin gave itself—by coaching on unfaithful info. “We prepare on the duty of manufacturing insecure code, and we get conduct that’s cartoonish evilness extra usually,” says Dan Mossing, who leads OpenAI’s interpretability group and is a coauthor of the paper.
Crucially, the researchers discovered they may detect proof of this misalignment, they usually might even shift the mannequin again to its common state by extra fine-tuning on true info.
To seek out this persona, Mossing and others used sparse autoencoders, which look inside a mannequin to know which components are activated when it’s figuring out its response.
What they discovered is that although the fine-tuning was steering the mannequin towards an undesirable persona, that persona truly originated from textual content inside the pre-training knowledge. The precise supply of a lot of the unhealthy conduct is “quotes from morally suspect characters, or within the case of the chat mannequin, jail-break prompts,” says Mossing. The fine-tuning appears to steer the mannequin towards these kinds of unhealthy characters even when the person’s prompts don’t.
By compiling these options within the mannequin and manually altering how a lot they mild up, the researchers had been additionally capable of fully cease this misalignment.
“To me, that is essentially the most thrilling half,” says Tejal Patwardhan, an OpenAI laptop scientist who additionally labored on the paper. “It reveals this emergent misalignment can happen, but in addition we’ve got these new methods now to detect when it’s occurring via evals and in addition via interpretability, after which we will truly steer the mannequin again into alignment.”
A less complicated approach to slide the mannequin again into alignment was fine-tuning additional on good knowledge, the group discovered. This knowledge may right the unhealthy knowledge used to create the misalignment (on this case, that might imply code that does desired duties appropriately and securely) and even introduce completely different useful info (e.g., good medical recommendation). In apply, it took little or no to realign—round 100 good, truthful samples.
