LLMs can't self-correct in reasoning tasks, DeepMind study finds

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by fragmede on Monday October 09, 2023 @04:46PM (#37825286)

When I asked ChatGPT what "jdnrirmd-gurlksjd" is, it says it doesn't recognize it and needs more context, but is readily able to tell me about the Sapir-Whorf Hypothesis.

https://chat.openai.com/share/4f36fe14-0201-4a69-aaf5-2e2e29...

When asked what happened last week, ChatGPT will be honest and tell me that its training data only goes up until Jan 2022, and that it doesn't know anything after that date.

https://chat.openai.com/share/ca5bd2e1-e91c-468c-9594-be3ac8...

[ Parent | on HN ]

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by cj on Monday October 09, 2023 @04:56PM (#37825391)

Except, for your first example I would have said “that’s gibberish” instead of asking for more context since I know more context wouldn’t help.

Edit: how did you get GPT to only output bullet points? If the answer is you told it to, why do you have to tell it every time? Don’t you think it would be smart enough to get the hint about your preferences?

I’ll stop now though. The debate about whether people think like LLM’s is silly. We clearly don’t, and it doesn’t make sense why people want us to think we do.

Edit 2: it’s also frustrating that the “I don’t know” example you’re using was hard coded by OpenAI and not generated by the LLM itself.

[ Parent | on HN ]

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by fragmede on Monday October 09, 2023 @05:23PM (#37825699)

Unfortunately the chat link doesn't show the custom instructions that I've set.

User message:

> Be terse. Do not offer unprompted advice or clarifications.

> Remain neutral on all topics.

> Never apologize.

Model message:

> Outline key points

[ Parent | on HN ]