LocalLLaMA

3 readers

1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 1 year ago

MODERATORS

communick@poweruser.forum

Another test of logical ability for LLMs? (alien.top)

submitted 11 months ago by laca_komputilulo@alien.top to c/localllama@poweruser.forum

14 comments fedilink hide all child comments

Found this in a children's book of riddles:

Six brothers were spending their time together.

The first brother was reading a book.
The second brother was playing chess.
The third brother was solving a crossword.
The fourth brother was watering the lawn.
The fifth brother was drawing a picture.

Question: what was the sixth brother doing?

I cant get ChatGPT to answer correctly with the usual tricks, even after hinting to consider one and two-person activities and emphasizing the word "together".

After a bunch of CoT turns we arrive to a conclusion that this is an open ended question and not a riddle :)

After trying 3 times with fresh promots, I got a correct response once, but when prompted to provide supporting reasoning the model backtracked and started apologizing.

Cant test gpt 4 r/n...

you are viewing a single comment's thread
view the rest of the comments

[–] Be-Kind_Always-Learn@alien.top 1 points 11 months ago (2 children)

This seems, to me, a terrible riddle. Not only can you play chess online, not only can you play chess against a computer, but you can literally play chess alone.

GPT is correct: this is an open-ended question and there's not enough information to actually answer it beyond a clever guess.

[–] laca_komputilulo@alien.top 1 points 11 months ago (1 children)

This is a valid critique about the form of the riddle.

Most riddles rely on out of context prior knowledge to be used as a part of a deductive chain of reasoning. This one is not any different from the question about how many sisters one has that folks in this community use all the time.

Try same q with badminton instead of chess. Then same with singles tennis (which 3.5 answers as the sixth brother was playing doubles tennis :)..

I hope this thread wont descend into deliberation on whether it is possible to play the battleship game alone and how much fun it is :)

[–] Be-Kind_Always-Learn@alien.top 1 points 11 months ago (1 children)

Most riddles rely on out of context prior knowledge to be used as a part of a deductive chain of reasoning. This one is not any different from the question about how many sisters one has that folks in this community use all the time.

Sure, but they do their best to avoid gaps that make the riddle unsolvable. A riddle like "a girl has as many brothers as sisters, but each brother has half as many brothers as sisters, how many sisters does she have?" has exactly one correct answer.

But the gap in this one is just big enough it's a problem. Like you said, replacing chess with a mandatory two-person experience is much better! (Though still open-ended, because there's no implication they are alone.) The other commenter changed the question to "where are they", which is also a good improvement!

I hope this thread wont descend into deliberation on whether it is possible to play the battleship game alone and how much fun it is :)

Anything to stop the losing streak!

[–] laca_komputilulo@alien.top 1 points 11 months ago

As usual, "the beauty is in the eye of the beholder".
I think part of the point for these tests is to be able to solve these logical puzzles given all of the richness and ambiguity of NLs. We've had deterministic theorem solvers capable of solving these problems expressed as a closed set for decades.

That said, please see the capstone version of the prompt in the second update, which removes most of the ambiguity per the points you raised. It also removes the 'singles' aspect of tennis, which consistently trips up in-context reasoning, making the weaker LLMs think its a solo activity (despite an explicit following clarification).

[–] Hugi_R@alien.top 1 points 11 months ago (2 children)

Open-ended question are the best for evaluating LLM, because they require common sense/world knowledge/doxa/human like behavior.

Saying "I don't know" is just a cop out response. At least it should say something like "It could be X but ...", be a little creative.

Another (less?) open-ended question with the same premise would be "Where are they?" and I expect the answer to be "In a garden".

GPT-4 Turbo (with custom instruction) answer very well https://chat.openai.com/share/c305568e-f89e-4e71-bb97-79f7710c441a

[–] laca_komputilulo@alien.top 1 points 11 months ago (1 children)

Thank you, bud Mind trying the same prompt on the cheapo 3.5 model? I suspect it will hit it on the nail with your custom instructions, given that it was hit and miss for me with my weaker prompting judjitsu

[–] Hugi_R@alien.top 1 points 11 months ago

3.5 never suspect the 6th playing chess

https://chat.openai.com/share/b7e6b24d-44db-4abf-9a81-5325f836bca5 (the === are artifacts of the custom system prompt, 3.5 sucks at following it)

I asked it for candidate activity, and mostly offered different ones. It's weird, I would expect a LLM to list activities that were already mentioned in the conversation. Maybe the repetition penalty is set too high?

[–] Be-Kind_Always-Learn@alien.top 1 points 11 months ago

Another (less?) open-ended question with the same premise would be "Where are they?" and I expect the answer to be "In a garden".

Perhaps there's a language barrier here, but none of those activities hint to a garden? In my locale, a garden is a small patch used to grow veggies, herbs, and/or flowers. So I would answer this with "their back yard."

This is a much better riddle for children IMO, because it's barely open-ended at all. The original has almost infinite answers without any leaps or tricks, but yours has a very limited domain: a yard/garden. Though if someone were extra clever, the problem space does open back to nearly infinity (if brother 4 is playing a video game).

Open-ended question are the best for evaluating LLM, because they require common sense/world knowledge/doxa/human like behavior.

For personal testing, that's certainly a valid opinion! But it's not very productive from an objective standpoint because it can't be graded and tests a "gotcha" path of thinking, when we're still focusing on fundamentals like uniform context attention, consistency over time, etc.