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

141–150 of 558 posts

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

#141

Post by djeastm »

ZooCow wrote: post #137
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 ;)

I had a similar thought but about asking the LLM to predict “future” major historical events. How much prompting would it take to predict wars, etc.?

You mean train on pre-1939 data and predict how WWII would go?

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

#142

Post by whimsicalism »

neom wrote: post #118
fny wrote: post #98

Earlier quoted context omitted.

But is there even enough pre-1905 data to create models that say hello world reliably? The terabytes of training data required for decent LLMs does not exist. I’d guess there may only be gigabytes worth.

My wife is an 18th century American history professor. LLMs have very very clearly not been trained on 18th century English, they cannot really read it well, and they don't understand much from that period outside of very textbook stuff, anything nuanced or niche is totally missing. I've tried for over a year now, regularly, to help her use LLMs in her research, but as she very amusingly often says "your computers ar…

my wish for new years is that every time people make a comment like this they would share an example task

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

#143

Post by whimsicalism »

amluto wrote: post #140
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 ;)

> 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

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

#144

Post by david-gpu »

ZooCow wrote: post #137
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 ;)

I had a similar thought but about asking the LLM to predict “future” major historical events. How much prompting would it take to predict wars, etc.?

That will never work on any complex system that behaves chaotically, such as the weather or complex human endeavors. Tiny uncertainties in the initial conditions rapidly turn into large uncertainties in the outcomes.

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

#145

Post by bdowling »

amelius wrote: post #61
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 ;)

This is how patent disputes should be decided. If an LLM can figure it out, then it is not novel.

Novelty (is it new) is the easy question because it’s just checking a database. Patentable inventions also have to be non-obvious, which is a more subtle question.

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

#146

Post by e1g »

jtefera wrote: post #139
e1g wrote: post #92

The paper includes several examples of their modified questions. There has been a substantial jump from o1-preview to o1, so I gave several samples to o1 and o1-pro ( not o1-preview), and current o1s gave the correct answer to those modified problems. SOTA changes fast.

The paper mentions that on several occasions the LLM will provide a correct answer but will either take big jumps without justifying them or will take illogical steps but end up with the right solution at the end. Did you check for that?

No, I don't know enough math to test the logic, only the check questions against their expected answers in https://anonymous.4open.science/r/putnam-axiom-B57C/data/Put...

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

#147

Post by chvid »

YetAnotherNick wrote: post #32
InkCanon wrote: post #10

Earlier quoted context omitted.

I've always assumed they removed it, because it's such a basic and fundamental part of ML training that you separate your test and train data. And yet I never see any papers even mention if/how they do this. And I wonder if they do, how do they guarantee with high reliability that their massive terabytes of data don't contain the answer.

First of all, Putnam is not in the test data, at least I haven't seen OpenAI claiming that publicly. Secondly, removing it from internet data is not 100% accurate. There are translations of the problems and solutions or references and direct match is not enough. MMLU and test set benchmarks show more resilience though in some previous research.

It is on the open internet - questions and suggested solutions:

https://kskedlaya.org/putnam-archive/

I would expect all llms to be trained on it.

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

#148

Post by whimsicalism »

jsheard wrote: post #105
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

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…

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

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

#149

Post by fldskfjdslkfj »

whimsicalism wrote: post #127
fldskfjdslkfj wrote: post #119

Earlier quoted context omitted.

Didn't they run a bunch of models on the problem set? I doubt they are hosting all those models on their own infrastructure.

1. OpenAI has confirmed it’s not in their train (unlike putnam where they have never made any such claims) 2. They don't train on API calls 3. It is funny to me that HN finds it easier to believe theories about stealing data from APIs rather than an improvement in capabilities. It would be nice if symmetric scrutiny were applied to optimistic and pessimistic claims about LLMs, but I certainly don’t feel that is the c…

Easier to believe or not, thinking that it's not a reasonable possibility is also funny.

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

#150

Post by lelandfe »

jsheard wrote: post #105
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

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”
Post reply on HN