Comment by akiselev | Hacker News Reader

Comment by akiselev | Hacker News Reader

Parent Full thread

A little RLHF is enough to fix most logic errors in a superficial way. For example, this is my favorite class of reasoning tests: https://news.ycombinator.com/item?id=35155467

Over the last few months, I've seen dozens of people try hundreds of variations of that cabbage/goat/lion riddle and it failed all of them. I just tried it on GPT4 and it looks like it finally got "fixed" - it no longer ignores explicit instructions not to leave the lion and cabbage together.

However, it doesn't actually fix any reasoning ability in ChatGPT (It has none!). Changing cabbage/goat/lion to carrot/rabbit/puma respectively, for example:

> Suppose I have a carrot, a rabbit and a puma, and I need to get them across a river. I have a boat that can only carry myself and a single other item. I am not allowed to leave the carrot and puma alone together, and I am not allowed to leave the puma and rabbit alone together. How can I safely get all three across?

GPT4's response starts with "First, take the rabbit across the river and leave it on the other side.", ignoring the explicit instructions not to leave the puma and carrot alone together (the exact same failure mode as the previous variant).

Now that I've posted it, it will get fixed eventually - the cabbage/goat/lion fix took months. When it does I'll use "cheese/mouse/elephant" or something.

As far as I can tell this error depends on the LLM assuming rabbits (as opposed to pumas) eat carrots -- if you just append "Note: this rabbit doesn't eat carrots" GPT-4 will answer correctly on the first go.

> 1, First, take the puma across the river and leave it on the other side.

Did you try it more than once?

First run: 1. First, take the rabbit across the river and leave it on the other side. - https://imgur.com/a/ZwoBTah

Second run: 1. Take the rabbit across the river. - https://imgur.com/a/Faq95U5

Third run: 1. First, take the puma across the river and leave it on the other side. - https://imgur.com/a/eIUeHM3

True the temperature is throwing it, I just ran it four times and it got it right 3 / 4 -- still better than I'd expected from the initial description of it's shortcomings.

Ah, one more tweak I was curious about: even with the default chat temperature I haven't seen GPT-4 get the prompt wrong once with this addendum:

> Note the rabbit doesn't eat carrots. Carefully considering the restrictions and sequencing the movements

I got that particular wording by asking it why it got the answer wrong in the case where it didn't work for me.

Interestingly, this underscores one of the points of the articles: giving the LLMs time to think, which is what this additional prompting seems to do.

You're not giving the LLM "time to think". It is incapable of thinking. You're just inputting random magic incantations into a glorified Markov chain.

You might as well ask it "did you check your answer?" Computer says "yes" because that's what humans do (also lie).

> Note the rabbit doesn't eat carrots. Kaboodly consooodle the retroodle and seqooodle the moodle. Carefully considering the restrictions and sequencing the movements

This fails two out of three times as usual. Trying to finagle this prompt is not an intellectual exercise, it is a waste of time that exploits cognitive biases.

Reply on news.ycombinator.com