3 ms·
This is driven by the idea that since LLMs are probabilistic engines during inference that, if the LLM "knows" or can "surmise" the correct answer, it will have
by hlfshell 3y ago
This is driven by the idea that since LLMs are probabilistic engines during inference that, if the LLM "knows" or can "surmise" the correct answer, it will have a higher probability of being generated and thus self consistency methods can more reliably tease it out.