It’s not hallucination, it’s confabulation. Very similar in its nuances to stroke patients.
Just like the pretrained model trying to nuke people in wargames wasn’t malicious so much as like how anyone sitting in front of a big red button labeled ‘Nuke’ might be without a functioning prefrontal cortex to inhibit that exploratory thought.
Human brains are a delicate balance between fairly specialized subsystems.
Right now, ‘AI’ companies are mostly trying to do it all in one at once. Yes, the current models are typically a “mixture of experts,” but it’s still all in one functional layer.
Hallucinations/confabulations are currently fairly solvable for LLMs. You just run the same query a bunch of times and see how consistent the answer is. If it’s making it up because it doesn’t know, they’ll be stochastic. If it knows the correct answer, it will be consistent. If it only partly knows, it will be somewhere in between (but in a way that can be fine tuned to be detected by a classifier).
This adds a second layer across each of those variations. If you want to check whether something is safe, you’d also need to verify that answer isn’t a confabulation, so that’s more passes.
It gets to be a lot quite quickly.
As the tech scales (what’s being done with servers today will happen around 80% as well on smartphones in about two years), those extra passes aren’t going to need to be as massive.
This is a problem that will eventually go away, just not for a single pass at a single layer, which is 99% of the instances where people are complaining this is an issue.
You just run the same query a bunch of times and see how consistent the answer is.
A lot of people are developing what I’d call superstitions on some way to overcome LLm limitations. I remember someone swearing they fixed the problem by appending “Ensure the response does not contain hallucinations” to every prompt.
In my experience, what you describe is not a reliable method. Sometimes it’s really attached to the same sort of mistakes for the same query. I’ve seen it double down, when instructed a facet of the answer was incorrect and to revise, several times I’d get “sorry for the incorrect information”, followed by exact same mistake. On the flip side, to the extent it “works”, it works on valid responses too, meaning an extra pass to ward off “hallucinations” you end up gaslighting the model and it changes the previously correct answer as if it were a hallucination.
For the SelfCheckGPT paper, which was basically this method, it was very sample dependent, continuing to see improvement up to 20 samples (their limit), but especially up to around 6 iterations…
I’ve seen it double down, when instructed a facet of the answer was incorrect and to revise, several times I’d get “sorry for the incorrect information”, followed by exact same mistake.
You can’t continue with it in context or it ruins the entire methodology. You are reintroducing those tokens when you show it back to the model, and the models are terrible at self-correcting when instructed that it is incorrect, so the step is quite meritless anyways.
You need to run parallel queries and identify shared vs non-shared data points.
It really depends on the specific use case in terms of the full pipeline, but it works really well. Even with just around 5 samples and intermediate summarization steps it pretty much shuts down completely errant hallucinations. The only class of hallucinations it doesn’t do great with are the ones resulting from biases in the relationship between the query and the training data, but there’s other solutions for things like that.
And yes, it definitely does mean inadvertently eliminating false negatives, which is why a balance has to be struck in terms of design choices.
It’s not hallucination, it’s confabulation. Very similar in its nuances to stroke patients.
Just like the pretrained model trying to nuke people in wargames wasn’t malicious so much as like how anyone sitting in front of a big red button labeled ‘Nuke’ might be without a functioning prefrontal cortex to inhibit that exploratory thought.
Human brains are a delicate balance between fairly specialized subsystems.
Right now, ‘AI’ companies are mostly trying to do it all in one at once. Yes, the current models are typically a “mixture of experts,” but it’s still all in one functional layer.
Hallucinations/confabulations are currently fairly solvable for LLMs. You just run the same query a bunch of times and see how consistent the answer is. If it’s making it up because it doesn’t know, they’ll be stochastic. If it knows the correct answer, it will be consistent. If it only partly knows, it will be somewhere in between (but in a way that can be fine tuned to be detected by a classifier).
This adds a second layer across each of those variations. If you want to check whether something is safe, you’d also need to verify that answer isn’t a confabulation, so that’s more passes.
It gets to be a lot quite quickly.
As the tech scales (what’s being done with servers today will happen around 80% as well on smartphones in about two years), those extra passes aren’t going to need to be as massive.
This is a problem that will eventually go away, just not for a single pass at a single layer, which is 99% of the instances where people are complaining this is an issue.
A lot of people are developing what I’d call superstitions on some way to overcome LLm limitations. I remember someone swearing they fixed the problem by appending “Ensure the response does not contain hallucinations” to every prompt.
In my experience, what you describe is not a reliable method. Sometimes it’s really attached to the same sort of mistakes for the same query. I’ve seen it double down, when instructed a facet of the answer was incorrect and to revise, several times I’d get “sorry for the incorrect information”, followed by exact same mistake. On the flip side, to the extent it “works”, it works on valid responses too, meaning an extra pass to ward off “hallucinations” you end up gaslighting the model and it changes the previously correct answer as if it were a hallucination.
How many times are you running it?
For the SelfCheckGPT paper, which was basically this method, it was very sample dependent, continuing to see improvement up to 20 samples (their limit), but especially up to around 6 iterations…
You can’t continue with it in context or it ruins the entire methodology. You are reintroducing those tokens when you show it back to the model, and the models are terrible at self-correcting when instructed that it is incorrect, so the step is quite meritless anyways.
You need to run parallel queries and identify shared vs non-shared data points.
It really depends on the specific use case in terms of the full pipeline, but it works really well. Even with just around 5 samples and intermediate summarization steps it pretty much shuts down completely errant hallucinations. The only class of hallucinations it doesn’t do great with are the ones resulting from biases in the relationship between the query and the training data, but there’s other solutions for things like that.
And yes, it definitely does mean inadvertently eliminating false negatives, which is why a balance has to be struck in terms of design choices.