2 ms·
Can LLMs be reliably restricted to produce any specific subset of content? AFAICT they're still consistently jailbroken.
by estearum 1y ago
Can LLMs be reliably restricted to produce any specific subset of content? AFAICT they're still consistently jailbroken.
- wat10000 1y agoIt’s easy to do at the syntactic level by controlling the sampling. For example, it’s easy and common to restrict output to be valid JSON but just not allowing any tokens that would make it not valid JSON. But reliably restricting output at the semantic level is very much an open problem.