I just had a wild idea…could we use memorization to reduce (or eliminate) the hallucination of facts? Example: Assuming Wikipedia is humanity’s baseline factset, is it possible for an LLM to memorize all Wikipedia, such that the model never hallucinates answers that contain Wiki facts?
It's an interesting proposition, but you'd still have to deal with the fact that the outputs are non-deterministic... (think of wiki memorized but then sometimes scrambled and switching articles 😅) There are some interesting approaches with having an output filter that essentially can act as a "fact checker", but as far as I've seen that still won't fix 100% of the problems. Perhaps if you have a really small problem/solution space (think of training your own AI with your own images, for example) you can better steer the stochastic bit in your favor! More articles and videos on those types of creations soon :)
In this case, I think you would be better off using RAG (or some other search method) to pull in pages of Wikipedia and put them in context. LLM's are quite good at following instructions like "Use only the background information provided" followed by the information you want it to present or reason over. And if there is still a chance of hallucination, you have a direct link to the source to validate.
Over fitting + temperature parametrization could tackle this problem. If temperature is a function of confidence, one may be able to avoid hallucinations, it might be even better to copy and paste from wikipedia using tools
Holy shit this video is a gem, can't believe you have so few views!!!
I just had a wild idea…could we use memorization to reduce (or eliminate) the hallucination of facts?
Example: Assuming Wikipedia is humanity’s baseline factset, is it possible for an LLM to memorize all Wikipedia, such that the model never hallucinates answers that contain Wiki facts?
It's an interesting proposition, but you'd still have to deal with the fact that the outputs are non-deterministic... (think of wiki memorized but then sometimes scrambled and switching articles 😅)
There are some interesting approaches with having an output filter that essentially can act as a "fact checker", but as far as I've seen that still won't fix 100% of the problems.
Perhaps if you have a really small problem/solution space (think of training your own AI with your own images, for example) you can better steer the stochastic bit in your favor! More articles and videos on those types of creations soon :)
In this case, I think you would be better off using RAG (or some other search method) to pull in pages of Wikipedia and put them in context. LLM's are quite good at following instructions like "Use only the background information provided" followed by the information you want it to present or reason over. And if there is still a chance of hallucination, you have a direct link to the source to validate.
Over fitting + temperature parametrization could tackle this problem.
If temperature is a function of confidence, one may be able to avoid hallucinations, it might be even better to copy and paste from wikipedia using tools
I know it’s really difficult to determine what a “fact” because reality is viewed from so many perspectives…