Evaluating publicly available LLMs on IMO 2025
89 comments
·July 19, 2025esjeon
> For Problem 5, models often identified the correct strategies but failed to prove them, which is, ironically, the easier part for an IMO participant. This contrast ... suggests that models could improve significantly in the near future if these relatively minor logical issues are addressed.
Interesting but I'm not sure if this is really due to "minor logical issues". This sounds like a failure due to the lack of the actual understanding (the world model problem). Perhaps the actual answers from AIs might have some hints, but I can't find them.
(EDIT: ooops, found the output on the main page of their website. Didn't expect that.)
> Best-of-n is Important ... the models are surprisingly effective at identifying the relative quality of their own outputs during the best-of-n selection process and are able to look past coherence to check for accuracy.
Yes, it's always easier to be a backseat driver.
wiremine
How quickly we shift our expectations. If you told me 5 years ago we'd have technology that can do this, I wouldn't believe you.
This isn't to say we shouldn't think critically about the use and performance of models, but "Not Even Bronze..." turned me off to this critique.
raincole
In 2024 AlphaProof got Silver level, so people righteously expect a lot now.
(It's specifically trained on formalized math problems, unlike most LLM, so it's not an apple to apple comparison.)
wat10000
LLMs are really good with words and kind of crap at “thinking.” Humans are wired to see these two things as tightly connected. A machine that thinks poorly and talks great is inherently confusing. A lot of discussion and disputes around LLMs comes down to this.
It wasn’t that long ago that the Turing Test was seen as the gold standard of whether a machine was actually intelligent. LLMs blew past that benchmark a year or two ago and people barely noticed. This might be moving the goalposts, but I see it as a realization that thought and language are less inherently connected than we thought.
So yeah, the fact that they even do this well is pretty amazing, but they sound like they should be doing so much better.
thaumasiotes
> LLMs are really good with words and kind of crap at “thinking.” Humans are wired to see these two things as tightly connected. A machine that thinks poorly and talks great is inherently confusing. A lot of discussion and disputes around LLMs comes down to this.
It's not an unfamiliar phenomenon in humans. Look at Malcolm Gladwell.
daedrdev
Here are the IMO problems if you want to give them a try:
https://www.imo-official.org/year_info.aspx?year=2025 (download page)
They are very difficult.
wrsh07
> Each model was run with the recommended hyperparameters and a maximum token limit of 64,000. No models needs more than this number of tokens
I'm a little confused by this. My assumptions (possibly incorrect!): 64k tokens per prompt, they are claiming the model wouldn't need more tokens even for reasoning
Is that right? Would be helpful to see how many tokens the models actually used.
throwawaymaths
they didn't even do a (non-ml) agentic descent? like have a quicky api that requeries itself generating new context?
"ok here is my strategy here are the five steps", then requery with a strategy or proof of step 1, 2, 3...
in a dfs
ipsin
I was hoping to see the questions (which I can probably find online), but also the answers from models and the judge's scores! Am I missing a link? Without that I can't tell whether I should be impressed or not.
raincole
On their website you can see the full answers LLM gave ("click cells to see...")
gcanyon
99.99+% of all problems humans face do not require particularly original solutions. Determining whether LLMs can solve truly original (or at least obscure) problems is interesting, and a problem worth solving, but ignores the vast majority of the (near-term at least) impact they will have.
lottin
15 years ago they were predicting that AI would turn everything upside down in 15 years time. It hasn't.
HEmanZ
People who say this don’t understand the breakthrough we had in the last couple of years. 15 years ago I was laughing at people predicting AI would turn everything upside down soon. I’m not laughing anymore. I’ve been around long enough to see some AI hype cycles and this time it is different.
15 years ago I, working on AI systems at a FAANG, would have told you “real” AI probably wasn’t coming in my lifetime. 15 years ago the only engineers I knew who thought AI was coming soon were dreamers and Silicon Valley koolaiders. The rest of us saw we needed a step-function break through that may not even exist. But it did, and we got there, a couple of years ago.
Now I’m telling people it’s here. We’ve hit a completely different kind of technology, and it’s so clear to people working in the field. The earthquake has happened and the tsunami is coming.
csa
Thank you for sharing your experience. It makes the impact of the recent advances palpable.
wavemode
To be frank, I take precisely the opposite view. Most people solve novel problems every day, mostly without thinking much about it. Our inability to perceive the immense complexity of the things we do every day is merely due to familiarity. In other words we're blind to the details because our brain handles them automatically, not because they don't exist.
Software engineers understand this better than most - describing a task in general terms, and doing it yourself, can be incredibly easy, even while writing the code to automate the task is difficult or impossible, because of all the devilish details we don't often think about.
gcanyon
I work with developers every day. Between us we often give the AI directions like:
* Write a query to link table X to table Y across this schema, returning all the unique entries related to X.id 1234
* Write code add an editable comment list to this UI
* Give me a design to visually manage statuses for this list
* Look at this UI and give me five ideas for improving it
Barrin92
the value of human beings isn't in their capacity to do routine tasks but to respond with some common sense to all the critical issues in the 2% at the tail.
This is why original problems are important, it's a measure of how sensible something is in an open-ended environment, and here they're completely useless, not just because they fail but how they fail. The fact that these LLMS according to the article "invent non-existent math theorems", i.e. gibberish instead of even being able to know what they don't know, is an indication of how limited this still is.
bgwalter
So the gold medal claims in https://news.ycombinator.com/item?id=44613840 look exaggerated.
The whole competition is unfair anyway. An "AI" has access to millions of similar problems stolen and encoded in the model. Humans would at least need access to a similar database; think open database exam, a nuclear version of open book exam.
AndrewKemendo
Can someone tell me where your average every day human that’s walking around and has a regular job and kids and a mortgage would land on this leaderboard? That’s who we should be comparing against.
The fact that the only formal comparisons for AI systems that are ever done are explicitly based on the highest performing narrowly focused humans, tells me how unprepared society is for what’s happening.
Appreciate that: at the point in which there is unambiguous demonstration of superhuman level performance across all human tasks by a machine, (and make no mistake, that *is the bar that this blog post and every other post about AI sets*) it’s completely over for the human race; unless someone figures out an entirely new economic system.
zdragnar
The average person is bad at literally almost everything.
If I want something done, I'll seek out someone with a skill set that matches the problem.
I don't want AI to be as good as an average person. I want AI to be better than the person I would go to for help. A person can talk with me, understand where I've misunderstood my own problem, can point out faulty assumptions, and may even tell me that the problem isn't even a problem that needs solving. A person can suggest a variety of options and let me decide what trade-offs I want to make.
If I don't trust the AI to do that, then I'm not sure why I'd use it for anything other than things that don't need to be done at all, unless I can justify the chance that maybe it'll be done right, and I can afford the time lost getting it done right without the AI afterwards.
SirFatty
"The average person is bad at literally almost everything."
Wow... that's quite a generalization. And not my experience at all.
Retric
The average person can’t play 99% of all musical instruments, speak 99% of all languages, do 99% of surgeries, recite 99% of all poems from memory etc.
We don’t ask the average person to do most things, either finding a specialist or providing training beforehand.
rahimnathwani
More than 50% of people cannot write a 'hello world' program in any programming language.
More than 50% of people employed as software engineers cannot read an academic paper in a field like education, and explain whether the conclusions are sound, based on the experiment description and included data.
More than 50% of people cannot interpret an X-ray.
AndrewKemendo
Which proves my point precisely that unless you’re superhuman in this definition, you’re obsolete.
Nothing new really, but there’s no where left to go for human labor and even that concept is being jeered at as a fantasy despite this attitude.
zdragnar
I really don't think it does, because we disagree on what the upper bound of an LLM is capable of reasoning about.
An average human may not be suitable for a given task, but a person with specialized skills will be. More than that, I believe they will continue to outperform LLMs on solving unbounded problems- i.e. those problems without an obvious, algorithmic solution.
Anything that requires brute force computation can be done by an LLM more quickly, assuming you have humans you trust to validate the output, but that's about the extent of what I'm expecting them to achieve.
bgwalter
Average humans, no. Mathematicians with enough time and a well indexed database of millions of similar problems, probably.
We don't allow chess players to access a Syzygy tablebase in a tournament.
baobabKoodaa
Average human would score exactly 0 at IMO.
pragmatic
That’s not how modern societies/economies work.
We have specialists everywhere.
AndrewKemendo
My literal last sentence addresses this
raincole
> average every day human
Average math major can't get Brozne.
pphysch
Machines have always had superhuman capabilities in narrow domains. The LLM domain is quite broad but it's still just a LLM, beholden to its training.
The average everyday human does not have the time to read all available math texts. LLMs do, but they still can't get bronze. What does that say about them?
chvid
In a few months (weeks, days - maybe it has already happened) models will have much better performance on this test.
Not because of actual increased “intelligence” but because the test would be included in model’s training data - either directly or indirectly where model developers “tune” their model to give better performance on this particular attention driving test.
sorokod
From the post: "Evaluation began immediately after the 2025 IMO problems were released to prevent contamination."
Doe this address your concern?
os2warpman
What they mean is that in a couple of weeks there are going to be stories titled "LLMS NOW BETTER THAN HUMANS AT 2025 INTERNATIONAL MATH OLYMPIAD" (stories published as thinly-veiled investment solicitations) but in reality they're still shitty-- they've just had the answers fed in to be spit back out.
sorokod
Companies would game metrics whenever they have the opportunity. What else is new?
chvid
Not really.
yunwal
Luckily there’s a new set of problems every year
chvid
You can really only do a fair reproducible test if the models are static and not sitting behind an api where you have no idea on how they are updated or continuously tweaked.
chvid
This particular test is heralded as some sort of breakthrough and the companies in this field are raising billions of dollars from investors and paying their star employees tens of millions.
The economic incentives to tweak, tune, or cheat are through the roof.
WD-42
> Gemini 2.5 Pro achieved the highest score with an average of 31% (13 points). While this may seem low, especially considering the $400 spent on generating just 24 answers
What? That’s some serious cash for mostly wrong answers.
One interesting takeaway for me, a non-practitioner, was that the models appears to be fairy decent at judging their own output.
They used best-of-32 and used the same model to judge a "tournament" to find the best answer. Seems like something that could be boltet on reasonably easy, eg in say WebUI.
edit: forgot to add that I'm curious if this translates to smaller models as well, or if it requires these huge models.