Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by skissane on Monday October 09, 2023 @05:59PM
(#37826000)
> The LLM I was using gave me code that worked, but did not do what I wanted. I pointed this out, so I then got code that syntactically did what I wanted, but could never work. This went back and forth twice, before I gave up.
What if you took the output of the LLM, and stuck it back into the LLM, and asked it if it could identify the contradiction? Just like humans, sometimes LLMs need to “step back a bit” to see contradictions which they can’t “in the heat of the moment”.
Also, what if you used a second LLM fine-tuned for contradiction detection to filter the output of the first?
Sometimes I daydream about a chatbot based on multiple LLMs fine-tuned on different tasks - a core “answer generation” LLM, a “contradiction detection” LLM, a “conversational pragmatics evaluation” LLM, a “style/register” LLM, a “cultural bias detection” LLM, etc, and then a final LLM which synthesises an overall answer based on their contributions. I wonder how its performance would compare to just a single LLM. I suppose it might be overly slow/expensive however. And, if it does have some performance advantage, I wonder if further scaling of current architectures would eliminate it?
What if you took the output of the LLM, and stuck it back into the LLM, and asked it if it could identify the contradiction? Just like humans, sometimes LLMs need to “step back a bit” to see contradictions which they can’t “in the heat of the moment”.
Also, what if you used a second LLM fine-tuned for contradiction detection to filter the output of the first?
Sometimes I daydream about a chatbot based on multiple LLMs fine-tuned on different tasks - a core “answer generation” LLM, a “contradiction detection” LLM, a “conversational pragmatics evaluation” LLM, a “style/register” LLM, a “cultural bias detection” LLM, etc, and then a final LLM which synthesises an overall answer based on their contributions. I wonder how its performance would compare to just a single LLM. I suppose it might be overly slow/expensive however. And, if it does have some performance advantage, I wonder if further scaling of current architectures would eliminate it?