Recent math benchmarks for large language models (LLMs) such as MathArena indicate that state-of-the-art reasoning models achieve impressive performance on mathematical competitions like AIME, with the leading model, o3-mini, achieving scores comparable to top human competitors. However, these benchmarks evaluate models solely based on final numerical answers, neglecting rigorous reasoning and proof generation which are essential for real-world mathematical tasks. To address this, we introduce the first comprehensive evaluation of full-solution reasoning for challenging mathematical problems. Using expert human annotators, we evaluated several state-of-the-art reasoning models on the six problems from the 2025 USAMO within hours of their release. Our results reveal that all tested models struggled significantly, achieving less than 5% on average. Through detailed analysis of reasoning traces, we identify the most common failure modes and find several unwanted artifacts arising from the optimization strategies employed during model training. Overall, our results suggest that current LLMs are inadequate for rigorous mathematical reasoning tasks, highlighting the need for substantial improvements in reasoning and proof generation capabilities.
“Notably, O3-MINI, despite being one of the best reasoning models, frequently
skipped essential proof steps by labeling them as “trivial”, even when their validity was crucial.”
LLMs are a lot more sophisticated than we initially thought, read the study yourself.
Essentially they do not simply predict the next token, when scientists trace the activated neurons, they find that these models plan ahead throughout inference, and then lie about those plans when asked to say how they came to a conclusion.
You didn’t link to the study; you linked to the PR release for the study. This is the study.
Note that the paper hasn’t been published anywhere other than on Anthropic’s online journal. Also, what the paper is doing is essentially a tea leaf reading. They take a look at the swill of tokens, point at some clusters, and say, “there’s a dog!” or “that’s a bird!” or “bitcoin is going up this year!”. It’s all rubbish dawg
Fair enough, you’re the only person with a reasonable argument, as nobody else can seem to do anything other than name calling.
Linking to the actual papers and pointing out they haven’t been published to a third party journal is far more productive than whatever anti-scientific bullshit the other commenters are doing.
We should be people of science, not reactionaries.
So, how does any of this relate to wanting to go back to an imagined status quo ante? (yes, I refuse to use reactionary in any other way than to describe politcal movements. Conservatives do not can fruits).
This isn’t debate club or men of science hour, this is a forum for making fun of idiocy around technology. If you don’t like that you can leave (or post a few more times for us to laugh at before you’re banned).
As to the particular paper that got linked, we’ve seen people hyping LLMs misrepresent their research as much more exciting than it actually is (all the research advertising deceptive LLMs for example) many many times already, so most of us weren’t going to waste time to track down the actual paper (and not just the marketing release) to pick apart the methods. You could say (raises sunglasses) our priors on it being bullshit were too strong.
> ask the commenter if it’s a study or a self-interested blog post
> they don’t understand
> pull out illustrated diagram explaining that something hosted exclusively on the website of the for-profit business all authors are affiliated with is
not the same as a peer-reviewed study published in a real venue
No prophet worked for free and they were always near the rullers and near big money. The story repeats itself, just the times are different and we can instant message with each other.
every time I read these posters it’s in that type of the Everyman characters in the discworld that say some utter lunatic shit and follow it up with “it’s just [logical/natural/obvious/…]”
Read the paper, it’s not simply predicting the next token. For instance, when writing a rhyming couplet, it first plans ahead on what the rhyme is, and then fills in the rest of the sentence.
The researchers were surprised by this too, they expected it to be the other way around.
Oh, sorry, I got so absorbed into reading the riveting material about features predicting state name tokens to predict state capital tokens I missed that we were quibbling over the word “next”. Alright they can predict tokens out of order, too. Very impressive I guess.
pray forgive, fair poster, for the shame I have cast upon myself in the action of doubting the Most Serious Article so affine to yourself - clearly a person of taste and wit, and I deserve the ire and muck resultant
wait… wait, no, sorry! got those the wrong way around. happens all the time - guess I tried too hard to think like you.
LLMs are a lot more sophisticated than we initially thought, read the study yourself.
Essentially they do not simply predict the next token, when scientists trace the activated neurons, they find that these models plan ahead throughout inference, and then lie about those plans when asked to say how they came to a conclusion.
this is credulous bro did you even look at the papers
You didn’t link to the study; you linked to the PR release for the study. This is the study.
Note that the paper hasn’t been published anywhere other than on Anthropic’s online journal. Also, what the paper is doing is essentially a tea leaf reading. They take a look at the swill of tokens, point at some clusters, and say, “there’s a dog!” or “that’s a bird!” or “bitcoin is going up this year!”. It’s all rubbish dawg
To be fair, the typesetting of the papers is quite pleasant and the pictures are nice.
they gotta make up for all those scary cave-wall pictures somehow
Fair enough, you’re the only person with a reasonable argument, as nobody else can seem to do anything other than name calling.
Linking to the actual papers and pointing out they haven’t been published to a third party journal is far more productive than whatever anti-scientific bullshit the other commenters are doing.
We should be people of science, not reactionaries.
you got banned before I got to you, but holy fuck are you intolerable
which we should do by parroting press releases and cherry picking which papers count as science, of course
but heaven forbid anyone is rude when they rightly tell you to go fuck yourself
So, how does any of this relate to wanting to go back to an imagined status quo ante? (yes, I refuse to use reactionary in any other way than to describe politcal movements. Conservatives do not can fruits).
nah I think it just sits weirdly with people (I can see what you mean but also why it would strike someone as frustrating)
This isn’t debate club or men of science hour, this is a forum for making fun of idiocy around technology. If you don’t like that you can leave (or post a few more times for us to laugh at before you’re banned).
As to the particular paper that got linked, we’ve seen people hyping LLMs misrepresent their research as much more exciting than it actually is (all the research advertising deceptive LLMs for example) many many times already, so most of us weren’t going to waste time to track down the actual paper (and not just the marketing release) to pick apart the methods. You could say (raises sunglasses) our priors on it being bullshit were too strong.
lmao fuck off
your argument would be immensely helped if you posted science instead of corporate marketing brochures
It’s an anti-fun version of listening to dark side of the moon while watching the wizard of oz.
I wonder if they already made up terms like ‘bloggophobic’ or ‘peer review elitist’ in that ‘rightwinger tries to use leftwing language’ way.
This study is bullshit, because they only trace evaluations and not trace training process that align tokens with probabilities.
remember, if we look too closely at the magic box,
we might notice how we’ve been fooledthe box will stop magicing for us!Well, every civilisation needs it’s prophets. Our civilisation built prophet machines that will kill us. We just didn’t get to the killing step yet.
yeah but see, these grifters all heard it as “every civilisation needs its profits”. just a shame they suck at that too
No prophet worked for free and they were always near the rullers and near big money. The story repeats itself, just the times are different and we can instant message with each other.
looks inside
it’s predicting the next token
every time I read these posters it’s in that type of the Everyman characters in the discworld that say some utter lunatic shit and follow it up with “it’s just [logical/natural/obvious/…]”
Stands to reason
Read the paper, it’s not simply predicting the next token. For instance, when writing a rhyming couplet, it first plans ahead on what the rhyme is, and then fills in the rest of the sentence.
The researchers were surprised by this too, they expected it to be the other way around.
Oh, sorry, I got so absorbed into reading the riveting material about features predicting state name tokens to predict state capital tokens I missed that we were quibbling over the word “next”. Alright they can predict tokens out of order, too. Very impressive I guess.
predict
ahead
stop prompting LLMs and go read some books, it’ll do you a world of good
nothx, I can find better fiction on ao3
Aw, you can’t handle a little science so you decide to throw insults instead.
the user who cannot read has been guided to go not read elsewhere
pray forgive, fair poster, for the shame I have cast upon myself in the action of doubting the Most Serious Article so affine to yourself - clearly a person of taste and wit, and I deserve the ire and muck resultant
wait… wait, no, sorry! got those the wrong way around. happens all the time - guess I tried too hard to think like you.