FTftxbro·May 26, 2023
Thanks that one is very interesting. I especially like how GPT4 performs worse (zero shot greedy decoding) than ChatGPT on those questions. I wonder have you thought about how it relates to 'inverse scaling' (https://github.com/inverse-scaling/prize)? As I was reading it I was thinking how that kind of 'hallucination' is exactly the problem that beam search (as opposed to greedy decoding) is designed to solve, and I see this was addressed later in the paper - unfortunately you guys didn't have access to it on the API and also no access to other things like log likelihoods.
A related question I had, is to decompose the blame for the problems these LLMs were having into two parts: (1) the weaknesses of the model size and data size and architecture caused for example by parallelism tradeoffs versus (2) the inherent problems of greedy token p
Read the rest
rediction. For example I could imagine that even an 'oracular' (kolmogorov or solomonoff or whoever absolute minimum theoretical perplexity) single-token-at-a-time-predictor could have related errors purely as a result of greedy decoding, which would be in no way attributable to any kind of size issue or architectural property or parallelism property of the model. As I guess you already know, the reason is just that 'the most likely next token of the response' isn't necessarily the same as 'the next token of the most likely response'.