Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by grepLeigh on Monday October 09, 2023 @03:56PM
(#37824713)
I think it's important to distinguish "LLM Self-Correction" (a prompt engineering technique) from lowercase "self-correction," a behavioral property of complex systems and/or sentient beings.
A PID controller is self-correcting (to some degree). A person is self-correcting (to some degree).
"LLM Self-correction" refers to an iterative approach to reaching some final output. This term is used in research/industry literature to mean a very specific implementation.
This allows researchers to compare "LLM Self-correction" to other prompting techniques like "LLM Chain-of-thought" in order to study the model's behavior in a laboratory setting, and in the wild.
A PID controller is self-correcting (to some degree). A person is self-correcting (to some degree).
"LLM Self-correction" refers to an iterative approach to reaching some final output. This term is used in research/industry literature to mean a very specific implementation.
This allows researchers to compare "LLM Self-correction" to other prompting techniques like "LLM Chain-of-thought" in order to study the model's behavior in a laboratory setting, and in the wild.
Does that distinction make sense?