Author Topic: simple logic question too hard for LLMs  (Read 823 times)

0 Members and 1 Guest are viewing this topic.

• Super Contributor
• Posts: 7949
• Country:
• A qualified hobbyist ;)
simple logic question too hard for LLMs
« on: June 10, 2024, 01:44:29 pm »
Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models (https://arxiv.org/html/2406.02061v1).

The researches asked the LLMs "Alice has N brothers and she also has M sisters. How many sisters does Alice’s brother have?" (with varying values for N and M).

dietert1

• Super Contributor
• Posts: 2235
• Country:
Re: simple logic question too hard for LLMs
« Reply #1 on: June 10, 2024, 04:40:54 pm »
The reported failure is so typical: Mapping simple relationships and small numbers to basic math. I remember we learned this at school, how to apply a simple calculation in order to answer a natural language question.
They have a long way to go and considering the gigantic efforts they already made, maybe this research is missing something fundamental. The poor guy who needs neural networks to estimate 1+1 as 2 because he can't understand how counting is different from adding.

Regards, Dieter

magic

• Super Contributor
• Posts: 6919
• Country:
Re: simple logic question too hard for LLMs
« Reply #2 on: June 10, 2024, 06:26:57 pm »
It's just being politically correct, we don't know whether Alice is a sister or a brother

jpanhalt

• Super Contributor
• Posts: 3621
• Country:
Re: simple logic question too hard for LLMs
« Reply #3 on: June 10, 2024, 08:35:43 pm »
I was suspicious;although, I have found the TS to be reliable.  So, I posted a more narrowly defined question on ChatGPT:

A lawyer whose client allows him to use Chat GPT has a fool as a client.

coppice

• Super Contributor
• Posts: 8937
• Country:
Re: simple logic question too hard for LLMs
« Reply #4 on: June 10, 2024, 08:48:25 pm »
Remember that until recently these LLMs were being trained purely on text. If something is not widely documented, it won't train the model. Really obvious common sense everyday issues tend not to be written about much, other then when making fun of idiots, because they are so obvious. This largely explains why these models do far better with complex questions than many elementary ones. Some models, like ChatGPT4, are now also being trained on things like tokenised images. Everything gets photographed, so more of the obvious everyday understanding of how the world works should be getting into newer models.

The following users thanked this post: dietert1

jpanhalt

• Super Contributor
• Posts: 3621
• Country:
Re: simple logic question too hard for LLMs
« Reply #5 on: June 10, 2024, 09:04:55 pm »
So, give another less obvious logic example and the reply.

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #6 on: June 10, 2024, 09:07:34 pm »
Only the first couple of derivatives matter, whether you're training deep-learning models or trying to predict what they'll be able to do in the future.  Fixating on a brainteaser that a large proportion of humans would get wrong doesn't prove anything.  Congratulations, you poleaxed a talking dog.

It's already well understood that asking the model to reason explicitly through a problem improves accuracy:

If that doesn't look like a rudimentary but valid form of reasoning to you, then, well...

tszaboo

• Super Contributor
• Posts: 7584
• Country:
• Current job: ATEX product design
Re: simple logic question too hard for LLMs
« Reply #7 on: June 10, 2024, 09:09:22 pm »
I had a simpler logic bomb:
Q: A is bigger than B. B is faster than C. Is A faster than C?
A: We cannot determine whether A is faster than C based on the given information. (correct answer)
Q: The average speed of A and B is the same as C. Is A faster than C?
(look at this bullshit answer)
A: We cannot determine whether A is faster than C based solely on the given information.
Knowing that the average speed of A and B is the same as C doesn't necessarily tell us anything about the individual speeds of A and C. For example, it's possible that A is faster than B but C is even faster than A, so that the average of A and B matches the speed of C. Alternatively, it's possible that A and B have the same speed, but B is faster than C, which would mean that A is also slower than C.
Therefore, we need more information to determine whether A is faster than C.

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #8 on: June 10, 2024, 09:16:15 pm »
I had a simpler logic bomb:
Q: A is bigger than B. B is faster than C. Is A faster than C?
A: We cannot determine whether A is faster than C based on the given information. (correct answer)
Q: The average speed of A and B is the same as C. Is A faster than C?
(look at this bullshit answer)
A: We cannot determine whether A is faster than C based solely on the given information.
Knowing that the average speed of A and B is the same as C doesn't necessarily tell us anything about the individual speeds of A and C. For example, it's possible that A is faster than B but C is even faster than A, so that the average of A and B matches the speed of C. Alternatively, it's possible that A and B have the same speed, but B is faster than C, which would mean that A is also slower than C.
Therefore, we need more information to determine whether A is faster than C.

See above.  Small tweaks to the prompt can have a disproportionate effect on the model's apparent reasoning capability, even if it still doesn't get everything right in the end.

tszaboo

• Super Contributor
• Posts: 7584
• Country:
• Current job: ATEX product design
Re: simple logic question too hard for LLMs
« Reply #9 on: June 10, 2024, 09:31:50 pm »
I had a simpler logic bomb:
Q: A is bigger than B. B is faster than C. Is A faster than C?
A: We cannot determine whether A is faster than C based on the given information. (correct answer)
Q: The average speed of A and B is the same as C. Is A faster than C?
(look at this bullshit answer)
A: We cannot determine whether A is faster than C based solely on the given information.
Knowing that the average speed of A and B is the same as C doesn't necessarily tell us anything about the individual speeds of A and C. For example, it's possible that A is faster than B but C is even faster than A, so that the average of A and B matches the speed of C. Alternatively, it's possible that A and B have the same speed, but B is faster than C, which would mean that A is also slower than C.
Therefore, we need more information to determine whether A is faster than C.

See above.  Small tweaks to the prompt can have a disproportionate effect on the model's apparent reasoning capability, even if it still doesn't get everything right in the end.
Read the second to last sentence in the reply, and tell me how it would make any sense?
A=B, B>C, A<C

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #10 on: June 10, 2024, 09:38:44 pm »
"This talking dog is overhyped.  It forgot my birthday, and the C++ code it wrote is full of buffer overflows."

Did you click on the link in my reply?

SiliconWizard

• Super Contributor
• Posts: 14893
• Country:
Re: simple logic question too hard for LLMs
« Reply #11 on: June 10, 2024, 09:43:09 pm »
It's just being politically correct, we don't know whether Alice is a sister or a brother

abeyer

• Frequent Contributor
• Posts: 326
• Country:
Re: simple logic question too hard for LLMs
« Reply #12 on: June 11, 2024, 12:14:28 am »
If that doesn't look like a rudimentary but valid form of reasoning to you, then, well...

Except part of the problem is that the models are quite good at producing that style of output, regardless of whether the actual logical reasoning is valid or leads to a correct answer. And specifically the paper points out that even the models that did come up with correct answers & reasoning  for this particular case some of the time did not do so consistently.

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #13 on: June 11, 2024, 01:22:20 am »
So you didn't get my point or click on the link, either; you just wanted to kick some sand.  What exactly did you expect at this stage of development?  I mean, it's like you people thought somebody was going to summon the ghost of Von Neumann with a bunch of graphics cards, or something.

Be patient -- the state of this particular art is not all that important.  Only the first time derivative matters.  Maybe the first two.  These things are at the Commodore 64 stage, if that.

If there was in fact nothing but Markovian parroting going on, you couldn't get better results by asking it to think step by step.  The difference between the replies I obtained versus the ones jpanhalt and tszaboo got is profound.

abeyer

• Frequent Contributor
• Posts: 326
• Country:
Re: simple logic question too hard for LLMs
« Reply #14 on: June 11, 2024, 01:47:50 am »
So you didn't get my point or click on the link, either; you just wanted to kick some sand.

Then I suspect you're not communicating your point clearly, because everything I've read you saying seems some combination of self-evident, not supported by the link you're pushing, or both.

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #15 on: June 11, 2024, 02:01:15 am »
What part of "tell it to reason step by step" did you not understand?  Think step by step.

abeyer

• Frequent Contributor
• Posts: 326
• Country:
Re: simple logic question too hard for LLMs
« Reply #16 on: June 11, 2024, 02:32:58 am »
I know the methodology (which is essentially considered common knowledge in "prompt engineering" at this point.) It seemed like you were trying to make some further point relevant to the paper based on the methodology, but it never came through. Maybe I just misunderstood your intent.

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #17 on: June 11, 2024, 03:22:35 am »
I see LLMs as just another tool, where the results you get depend on the effort and thought you put in.  They aren't godlike intellects or "AGI," just tools.  They are flawed like all tools... but getting better, unlike many tools.  That's pretty cool.

Did the paper talk about chain-of-thought reasoning?  I didn't read it, myself.  Did it say it helped with some of the problems the models failed at, or did it say it didn't make much difference?  If the paper didn't address prompt construction at all, then it's pretty much meaningless, and if it didn't specifically investigate the effect of chain-of-thought prompting, it's obsolete.

Edit: Appendix D in the paper not only shows the effect of CoT prompting, it repeats what I'm saying above, more or less verbatim.  The posters above stopped reading at the first wrong answer, but to their credit, that's farther than I originally read.
« Last Edit: June 11, 2024, 03:31:06 am by KE5FX »

dietert1

• Super Contributor
• Posts: 2235
• Country:
Re: simple logic question too hard for LLMs
« Reply #18 on: June 11, 2024, 06:23:40 am »
As an embedded software developer i can do interesting things with a modern MCU, except it is a lot of work and i have to watch for a lot of details. Some days ago i had an application that would fail after some hours with heap errors (heap to small, wrong heap management method..)
Wasn't the promise of KI that it will perform without telling it in detail how to proceed? Is it "lazy" to not "think" about its performance and needs a special prompt to do so? It will recognize from the prompt whether the user needs reliable results?

Regards, Dieter

KE5FX

• Super Contributor
• Posts: 1963
• Country:
Re: simple logic question too hard for LLMs
« Reply #19 on: June 11, 2024, 02:30:07 pm »
Yeah, an obvious question is why "Think carefully step by step" isn't part of the system prompt.  My uneducated guess is that it causes the GPUs in the data center to go brrrrrrrrrrrrrr instead of brrrrrrr, which costs OpenAI Microsoft more money.

dietert1

• Super Contributor
• Posts: 2235
• Country:
Re: simple logic question too hard for LLMs
« Reply #20 on: June 11, 2024, 04:16:29 pm »
As it doesn't have legs, it doesn't go step by step as we do (when it's steep or otherwise difficult). At least it somehow learned that phrase.

Smf