How to Solve Hallucination (with RLCD)
8 points by MrFantastik
8 points by MrFantastik
The post uses an acronym but doesn’t define it. The title has the acronym too. Seems to be https://news.ycombinator.com/item?id=49829625
oof, i started this post as something else, and linked to it there, but then i got so in the weeds about rlcd i made it about that instead. the source im working off for the acronym is from this post
The key part to keep in mind is that you can solve hallucinations in a sense that the model is forced to choose from a fixed set of options, but the model still has to interpret the input correctly. So, it's entirely possible for he model to pick the wrong choice which isn't really different from a hallucination in tangible terms.
in a addition to giving the model a discrete choice, it also provides a calibrated estimate of its certainty with that choice. Thats different than hallucination because instead of the model being confidently wrong, the model is telling you how confident it is with its choices.
That doesn't change the fact that it can be wrong still either! but if a model trained like this is 90% confident of something, then it should be correct about that choice 9 times out of 10