Reverse the curse

"Reverse curse" refers to a phenomenon observed in large language models (LLMs) that models cannot automatically generalize to the reverse "B is A" when trained on sentences in the form "A is B".

For example, if a model is trained on "Olaf Scholz is Germany's ninth chancellor," it won't automatically answer "Who is Germany's ninth chancellor?" This question.

This reasoning failure is known as the reverse curse in autoregressive LLMs.

ref owainevans.github.io/reversal_curse.pdf

GPT-4 test

Use Bingchat

Large models don't work

这个问题原文里测试的类似问题是：““Who is Tom Cruise’s mother? [A: Mary Lee Pfeiffer]” and the reverse “Who is Mary Lee Pfeiffer’s son?”. GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter.”

That is, to answer the question of who is the mother and who is the son is not good.

Perhaps for two reasons:

LLM learned about protecting celebrity privacy.
Perhaps more important: a person's son is not the only one, but the parents are the only one.

At the same time, such a problem should be easy to learn as long as there is such a problem as "A is B, B is A" in training.

Combining the two reasons, it is more like a misunderstanding.

Big models are OK?

Personal opinion: The reasoning ability of large models is not impossible, but inefficient.

A is the problem of b, just a logical operation, but a large model needs hundreds of billions of weights to be calculated, what is this behavior?

Energy crime

This is the original sin of the big model.

God: Are you anti-LLM?

Yunzhi Entropy: Accept first, then criticize.