Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

30% drop in O1-preview accuracy when Putnam problems are slightly variated

openreview.net

Post reply on HN
Print view View original

181–190 of 558 posts

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#181

Post by lupire »

chvid wrote: post #51

Isn't this simply because the dataset used (Putnam-AXIOM Original) is in the training data used to train the various models? Given that these are simple variations (variable names and constants value change in math problems). Why would the companies creating these models (OpenAI etc.) create these variations themselves in order to insure that the model is learning how to solve the problem rather than memorize a solut…

They are not only simple renames. LLM is good at those. They are minor structural changes.

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#182

Post by gus_massa »

whimsicalism wrote: post #143
amluto wrote: post #140

Earlier quoted context omitted.

> ask it for a formula for mass-energy equivalence Way too easy. If you think that mass and energy might be equivalent, then dimensional analysis doesn’t give you too much choice in the formula. Really, the interesting thing about E=mc^2 isn’t the formula but the assertion that mass is a form of energy and all the surrounding observations about the universe. Also, the actual insight in 1905 was more about asking the…

but e=mc^2 is just an approximation e: nice, downvoted for knowing special relativity

I didn't downvote it, but short comments are a very big risk. People may misinterpret it, or think it's crackpot theory or a joke and then downvote.

When in doubt, add more info, like:

But the complete equation is E=sqrt(m^2c^4+p^2) that is reduced to E=mc^2 when the momentum p is 0. More info in https://en.wikipedia.org/wiki/Mass%E2%80%93energy_equivalenc...

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#183

Post by whimsicalism »

jsheard wrote: post #167
whimsicalism wrote: post #148

Earlier quoted context omitted.

https://chatgpt.com/share/67755e6f-bfc8-8010-9aa3-8bcbbd9b26...

To be clear I was testing with 4o, good to know that o1 has a better grasp of basic arithmetic. Regardless my point was less to do with the models ability to do math and more to do with OpenAI seeming to cover up its lack of ability.

i think it’s mostly that o1 mini can think through the solution before it starts writing the poem.

i’m able to reproduce your failure on 4o

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#184

Post by zeroonetwothree »

sundarurfriend wrote: post #122
ben_w wrote: post #75

Link title says "slightly", but the PDF says two different kinds of variations: variable names (slight) and problem constants (significant), and the 30% drop is on the combination of a 26 variable and also 26 variable + constant questions. It's good to have a better test (though I bet this one will also be quickly saturated like all the others), but the title here doesn't seem justified by the page title there or the…

I would definitely classify both of those as slight changes. In fact I'd rename those as slight => trivial and significant => slight.

Right, renaming a variable should have zero effect on ability to solve (it wouldn’t for a human). Changing a constant should be very minor, probably also ~0 effect in most cases. I say this as someone that’s done many of these problems.

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#185

Post by whimsicalism »

lelandfe wrote: post #150
jsheard wrote: post #105

Earlier quoted context omitted.

It certainly feels like certain patterns are hardcoded special cases, particularly to do with math. "Solve (1503+5171)*(9494-4823)" reliably gets the correct answer from ChatGPT "Write a poem about the solution to (1503+5171)*(9494-4823)" hallucinates an incorrect answer though That suggests to me that they've papered over the models inability to do basic math, but it's a hack that doesn't generalize beyond the simpl…

“a poem about” reads to me at least like the solution need not be in the answer; maybe something like “a poem that includes the answer in the last stanza”

yeah but it like actually gets the answer wrong not just omits it

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#186

Post by zeroonetwothree »

whimsicalism wrote: post #130
optimalsolver wrote: post #123

Earlier quoted context omitted.

EpochAI have to send the questions (but not the answer key) to OpenAI in order to score the models. An overnight 2% -> 25% jump on this benchmark is a bit curious.

1. OpenAI said they did not train on these problems & they don’t train on API calls in general, that is a legal policy. 2. It was a new major model release from work over the course of months - struggle to see that as an ‘overnight’ jump in any real meaning. 3. Why is it easier to believe large scale corporate fraud than that the stated capabilities on a held out test set are real? Reads like cope, if I’m being frank…

I don’t think it’s “easier to believe” just that it raises some red flags.

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#187

Post by zeroonetwothree »

steveBK123 wrote: post #88

Yes so when you change the sequence of tokens they've electronically memorized, they get a bit worse at predicting the next token?

When you put it that way it’s a trivial result. However the consequences for using AI to replace humans on tasks is significant.

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#188

Post by wslh »

wim wrote: post #49

One experiment I would love to see, although not really feasible in practice, is to train a model on all digitized data from before the year 1905 (journals, letters, books, broadcasts, lectures, the works), and then ask it for a formula for mass-energy equivalence. A certain answer would definitely settle the debate on whether pattern recognition is a form of intelligence ;)

Why do we need this when current models already handle questions and answers about new discoveries: ones that are happening every week and are often easier to grasp than Einstein’s equations? I think it is clear that they will fail on most of them. That doesn't mean that LLMs are not useful but there are more walls in the road.

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#189

Post by reissbaker »

alexashka wrote: post #96
Lerc wrote: post #43

Earlier quoted context omitted.

Gpt4 can add very large integers. It is evident that it is not recalling the sum because all combinations of integer addition were likely not in the training data, Storing the answer to the sum of all integers up to the size that GPT4 can manage would take more parameters than the model has. That addition is a small capability but you only need a single counterexample to disprove a theory.

> That addition is a small capability but you only need a single counterexample to disprove a theory No, that's not how this works :) You can hardcode an exception to pattern recognition for specific cases - it doesn't cease to be a pattern recognizer with exceptions being sprinkled in. The 'theory' here is that a pattern recognizer can lead to AGI. That is the theory. Someone saying 'show me proof or else I say a pa…

GPT-4o doesn't have hardcoded math exceptions. If you would like something verifiable, since we don't have the source code to GPT-4o, consider that Qwen 2.5 72b can also add large integers, and we do have the source code and weights to run it... And it's just a neural net. There isn't secret "hardcode an exception to pattern recognition" in there that parses out numbers and adds them. The neural net simply learned to do it.

Re: 30% drop in O1-preview accuracy when Putnam problems are slightly variated

#190

Post by freehorse »

hansworst wrote: post #5
Trasmatta wrote: post #4

Earlier quoted context omitted.

Not hardcoded, I think it's just likely that those problems exist in its training data in some form

Isn't that just the LLM equivalent of hardcoding though?

"Overfitting" would be a bit more accurate term if the problem lies in the specific examples existing in its training set in various forms, places, languages etc but with the same values.
Post reply on HN