Aside from the usual squabbling about AI, it seems the bombshell claim is this:
"In particular, we show that the formalised Lean proof does not correspond to the NL proof of blow-up of solutions to the Navier-Stokes equations."
So these authors seem to be claiming that OpenAI has not really proven Navier-Stokes at all. If I get their idea correctly, they are claiming that the LLM has not formalized the original "natural language" idea of Navier-Stokes correctly. If true, it would mean that their purported Lean proof is not actually a proof of Navier-Stokes at all, but something that is an incorrect translation of the original natural language idea. If correct, this is a really bold claim and I would like to see if other researchers agree.
No one is disputing that the Lean formaization of Navier-Stokes is correct, so we should have high confidence that the generated Lean proof is valid.
The authors are claiming that the Lean proof is not the same proof as the NL one. Therefore, we shouldn't yet have confidence that the NL proof is valid.
This is an important claim which the math community will need to work through. However, the Lean proof alone is sufficient for OpenAI to (reasonably confidently, leaving aside questions of academic manners) claim to have proven NS.
So, the AI wrote a NL proof of Navier-Stokes, then incorrectly auto-formalised it to Lean, but still ended up with a verifiable proof of Navier-Stokes? That seems... strange?
Or the AI wrote a NL proof of Navier-Stokes, began rewrite in Lean, then discovered a false / handwavy / easier to write in Lean / etc approach of some parts of the proof, and modified it accordingly. Since there wasn't any backpass from Lean to NL to include any changes it did due to any of the above reasons, the proofs aren't identical. That's what I think is most likely.
If the reason for the differences was done intentionally in Lean (as opposed to hallucinate e.g. m+4 vs m+5 as mentioned in remark 3.2), then a simple recording of differences, and then afterwards pass back any changes to the original NL would fix the issue. If it was hallucinated, then there is no guarantee it wouldn't keep hallucinating, and thus you might never end up with the same proof no matter how many passes you do back and forth (see remark 3.4).
It's not actually the same model that solved the problem that did the translation. Astra did the translation after the intenral model produced the NL Proof. As for the discrepancies, It's not necessarily right to think of this as 'incorrect formalisation'. Maybe it was essentially a 'proof refactoring'. Maybe Astra thought some parts could be easier expressed in a certain way, or maybe aspects of the NL proof were kind of handwavey etc.
Or maybe the AI didn’t write Lean proof at all, or rather, not the LLM at least. But instead OpenAI has an internal traditional reinforcement model that is able to stumble on the Lean proof by the share amount of compute power available to them thousand monkeys on a thousand typewriter style. And then pretend LLM did it because that is what they are selling.
I don’t. But I do know how the scientific method works, and OpenAI’s display is anything but. Until what they have demonstrated is reproduced I take their claims to be nothing but marketing. A for profit company will lie in order to maximize their profits. Above I presented an alternative hypothesis, which is probably wrong, but until OpenAI’s results are replicated I will believe my alternative hypothesis just as much as I believes the claims of the for profit company making them.
>No one is disputing that the Lean formaization of Navier-Stokes is correct, so we should have high confidence that the generated Lean proof is valid.
These authors don't seem to be disputing that this Lean formalization of Navier-Stokes is correct. I don't think that gives us any new information about whether the generated Lean proof is or isn't a valid proof of this N-S blowup thing.
The thing is that Navier-Stokes has a definition split off separate from the formalization, and that is what has been completed. People have looked at the definition of the final statement. This paper only mentions the proof and intermediate statement, not the final statement. The most likely case to me is that intermediate statements do not match, but the end result still holds.
Seems kinda odd then that it didn't occur to OpenAI to iterate until they reached a fixedpoint for both the informal & formal development b/c it's obvious that correspondence should have been part of their training pipeline.
Jesus Christ, so many people here who have no clue what they are talking about.
A proof of a theorem is different from the statement of the theorem. OpenAI has a Lean proof of the statement. That is all they need. There may be many different proofs of this statement, including NL proofs. It does not matter that these NL proofs may or may not be different from the Lean proof, at least for the correctness of the Lean proof. But of course the NL proof may be wrong. But who cares?
> Jesus Christ, so many people here who have no clue what they are talking about.
Indeed. If only some of those people would see the irony.
What matters most of all, as any first year student of mathematics would know, is whether the formal problem statement corresponds to the NL statement. TFA specifically states that at least some of the allegedly proven formal statements DO NOT.
No. What the paper says is that in principle, translating NL statements to Lean statements is hard. Nobody doubts that, translating informal to formal text cannot be formally proven correct, so...
Does the paper give a single example of one of the OpenAI solved theorems with a Lean certificate where the Lean statement does not correspond to the actual statement from the mathematical literature? I don't think so, but in case I am wrong, feel free to provide that example.
If I understand the abstract correctly (big caveat), they aren't saying they didn't prove it. They're saying they gave two proofs, one in natural language and one in Lean, that are not equivalent to each other. I assume the main significance is that the Lean proof is not a formal verification of the natural language one and the natural language proof is not a readable explanation of the Lean one. Both of those things can be desirable, so to complete the set we'd get 4 proofs.
But just to clarify: is either of them actually addressing the real Navier-Stokes, or will it turn out we'll end up with two pairs of proofs about something irrelevant to the actual problem?
From computer science perspective the conclusion is obvious: untenable to have two representations without an exact translation or machine checked correspondence between then. All we have is a vibe translation using the LLM. The methodology should obviously be improved.
and clearly the computer science perspective is: get rid of the humans and express everything directly in lean so the computers can keep getting work done!
If I understand well, they mean that the thing they proved in lean is not Navier Stokes. And they don't make any statement about whether the natural language proof is correct or not.
I read them as making a much weaker claim than this: not that the Lean proof isn't valid, just that it is not actually a formalization of the natural-language proof in the PDF they provided alongside it. I haven't heard any PDE people claim that the Lean proof is invalid, and I have heard things from a lot of them that imply that they think it is valid. (I'm a former research mathematician, but this is very far from my specialty, so I'm not really equipped to evaluate this claim myself.)
The claim is about the equivalence between two proofs and says nothing about the correctness of either proof. This seems to be confusing a lot of people.
NS is continuous approximation to what is otherwise a discrete system. Particle collisions are discrete time events that are averaged over time. NS loses accuracy for very, very, very very low fluid densities and energies.
AI "proving" that this approximation can numerically "blow" up does not mean the approximation loses validity.
If I'm understanding correctly, this is questioning the equivalence between the natural language proof and the lean proof, but not the correctness of the lean proof?
If the lean proof doesn't match the natural language one (which is the one the AI generated to solve the problem), it sounds like the lean proof isn't verifying the intended claim?
From the paper:
"A third possibility is that the NL proof provides stronger statements than what the formal proof actually establishes,
with (of course) different proofs. The latter happens in OpenAI’s announced proof of blow-up of Navier Stokes equations."
The material is interesting, but unless the statement that is proved in lean is not blowup for Navier-Stokes, then it's still proven.
What the examples seem to show is that the proof method is different between the natural language proof and the lean proof. Which, if the lean proof actually proves blowup, would suggest that the natural language proof is subtly wrong, but the strategy was close enough to be used to create a real lean proof.
A little worrying, but part of the purpose of formalizing things in Lean, it forces you to be more accurate than natural language does. It's surprisingly common for major theorems to have slight inaccuracies early on that can be repaired. Famously, the initial proof of Fermat's Last Theorem had a flaw that took a year to repair (though I think that's unusually difficult).
So the most fundamental question is: does the Lean theorem faithfully state the right theorem?
That assumes the natural language paper came first and then was formalized in lean. I haven't looked too deeply into how these labs solve these problems (or if they even specify this publicly) but you could also start with lean and then write the natural language proof based on it.
For what it's worth the initial lean specifications for the top-level theorems generally come from human written formalizations such as in https://github.com/leanprover-community/mathlib4/blob/021ce6... so we can be reasonably confident about their correctness.
Not quite — the formal statement of the problem in Lean may be correct, and therefore the Lean proof gives quite a lot of confidence that the statement is true. It's just that the proof given in natural language doesn't necessarily match up with the Lean proof, so the natural language proof might be unsound even though the statement it's proving is true.
> If the lean proof doesn't match the natural language one (which is the one the AI generated to solve the problem), it sounds like the lean proof isn't verifying the intended claim?
No, the other way around. The natural language proof was derived from the lean code, badly. This is my experience with using claude and lean to prove things. Its natural language explanations drift a lot from the lean, both before and after. But the lean code is the lean code.
> The natural language proof was derived from the lean code, badly.
Was it? Are you claiming a LLM does reasoning in lean or what? Since this (and all the other proofs by OpenAI etc) have been in the reverse order [1]:
> The agents arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT‑6 Astra.
An incredible number of people think that it is reasoning in lean. Argued with several people on this topic. I think they read headlines about lean being used by LLMs and assume it is being used to write the proof.
It's not that much of a stretch: give the LLM a top-level proposition for the thing you want to prove and have it hack away at it. Each sub-step is verified in lean so you know it's correct. But, the linked post definitely suggests otherwise.
That is definitely interesting because how do you know the 88 hours of work are correct before you throw another 17 hours of lean formalization work on it? You could end up just finding out there was some hallucination in the original work.
That makes some sense. Given that the vast majority of math in its training data is going to be in NL/latex, I just assumed that the core reasoning happens in NL with occasional LEAN checks to ensure validity.
I'm pretty sure Mathlib has had enough human authored definitions to formalize the basic calculus necessary to state Navier-Stokes for quite some time? Some other problems admittedly need quite a bit of machinery built up to even try to say what the question is, but every undergrad learns multiple approaches to formally define everything necessary to write down a PDE.
They do not. Maybe if they take Lean classes? Maybe starting this year they will but my youngest kid is on their like 5th math class in undergrad and hasn't had any lean at all. Not all undergrad math majors even take PDEs; applied maybe, unless you are doing applied discrete math (graphs, combinatorics).
Not Lean specifically, but IMO it's pretty straightforward if you've done math and some programming (and at least my school required some programming).
Need to prove a forall statement? forall x, P(x) is the same as a function taking x and returning the proof that P(x) is true.
Need to prove an exists statement? Create the pair (x, h) that gives the actual x that proves the exists, along with a proof that it satisfies the property you claim.
Maybe the only weird thing is that there are types and sets, so sets are kind of automatically more of a "subset" of some type.
The actual Mathlib is more generic, but once you get a hang of writing definitions (as you do in intro proofs), I've found that you can pretty naturally translate whatever you'd have in your undergrad notes. And undergrad should cover defining integers, rationals, reals, relations, functions, sequences, limits, derivatives, integrals, etc. Even if they've never studied solving PDEs, they'd have to take multivariable calculus and know enough to be able to write one (assuming they take at least single variable analysis+linear algebra)?
The proofs can get involved and tedious with all of the extra bookkeeping, or techniques to try to reduce the bookkeeping (tactics, etc). But the definitions and statements are pretty much what you'd expect.
Indeed. The natural language proof is incorrect but the Lean proof is correct.
Humans have made similar mistakes too. A human writes a specification for how things should work, the human translates that into code, the code does not work, and finally the human fixes the code and forgets to fix the original spec.
Yes — because there are many non-equivalent statements that are easier to prove.
So the Lean proves something and the question is whether that something is actually what we care about — or something similar, but ultimately not the question.
This paper is a large amount of nothing. First, natural language is not as precise as lean, so you have multiple ways to translate a NL argument to Lean. As shown in Fig 1, the LLM did a decent job at translating the argument about roots in a succint way.
Moreover, the paper claims that the NL arguments of Navier-Stokes are stronger than the Lean ones. My understanding is that the translator LLM got lazy and wrote the minimal amount of code that satisfied the theorem without the extra stronger claims.
It is common in mathematical papers to say "And by the way, this actually proves [stronger claim]", but this is something an AI with a precise goal of performing a translation would never do, as it's goal is to translate the proof, not to quality mathematics.
The paper shows that the Lean proof and the prose (pdf) proof do not match exactly. But if the Lean theorem Lean accepted is equivalent to original problem statement published by the Clay Institute, this mismatch is of no consequence to the validity of the proof itself. That's not a trivial if: stating the problem precisely is often as hard as the proof. Validation efforts should concentrate on whether the Lean theorem is equivalent to the one published by the Clay Institute.
That said, a gap between the Lean proof and the pdf is annoying for interpretability, and interpretation is a valid aim, but that does not factor into the proof's validity.
> That's not a trivial if: stating the problem precisely is often as hard as the proof.
Really? You think 300 lines of Lean code [1] is just as hard as the proof (or even remotely close)? Also note, as the README says [2], that the theorem was written independently by formal conjectures, not by the LLM.
- The lean kernel could always have a bug.
- The formalized statement may not correspond to what _mathematicians_ "actually
wanted"
It seems natural to make the argument of, "Well, even if you make the argument
that the proof can have mistakes, it's surely easier to check the problem
statement of something rather than the solution".
(A "nice property" is that, the agent doesn't need to even get "subarguments
correct" according to the _second_ criteria - maybe in the natural proof it
invents an object subtly different from the formal one, but it all checks out.
If you guarantee that the _original_ statement corresponds, then the only
possibility is the lean kernel. So it doesn't recurse infinitely, in this case).
But "definitions" are always a really weird thing that I don't think we have
good theories for? How do you quantify how much descriptive power you need to
express a question? Often times in math, the hard part is getting the definition
right - but what if the definition itself starts to become so complex and
unverifiable that no one can correspond that to anything? Well, it seems like
many interesting long-standing math problems have "relatively" simple problem
statements, in such a way that you could formalize it to lean easily, but not
sure if there's really a silver bullet w/ lean or if it's going to be turtles
all the way down.
It probably doesn't matter as long as AI keeps skyrocketing on the much more
general property that is "intelligence", but still. Interesting to think about.
(Well, this is where AIT gets actually interesting, but still, I don't think its
a generalized theory of semantics.)
Given the high-level description of the examples, I think it's less of a "mis-translation" as it is the LLM tweaking the proof as it formalized it. Going between m+4 and m+5 is a pretty different thing than the sort of ambiguities that generally arise in parsing natural-language mathematical statements.
Hm, looking through here, I don't see where they state what it is that OpenAI actually proved instead of Navier-Stokes blowup with forcing. I see where they do this for some other particular statements used along the way, but not for the headline result.
Well, I believe the incompleteness theorems speak about provability, not about how the proofs themselves are expressed.
We know as a consequence of Goedel theorems (at least I believe so), that there is no algorithm that would take a statement and output a proof if it is provable or a counterexample if it is not. However, AI provers never give anything for sure, so I think there is no contradiction here.
The incompleteness theorems state that every sufficiently complicated logic lets you construct a statement that is effectively "this statement has no proof," so either there exists true statements that lack proofs (incompleteness) or there exists false statements with proofs (incorrectness).
Just all the useful proofs. You can get arbitrarily more complicated and uninteresting theorem statements by making meta statements about the system you are doing proofs in. At some level the system can't answer questions about itself.
The incompleteness theorem says that there are statements which can be neither proven true nor false in a given axiomatic system. If there is a proof to write in lean, then the statement is already outside the bounds of incompleteness.
I recently spent 3 weeks with claude formalizing a CS paper about a borrow checker in lean, for a personal project.
The formalization went through, but there were _several_ mistakes in the original paper that it uncovered, from type setting errors to (many) formulas that quantified over all resources as printed, but actually applied to only arising resources in the calculus..
So the formalization did give me a formally verified borrow checker that I could use to build a programming language on top of, but it was _not_ exactly the borrow calculus that was printed in the paper.
I expect this is the most common experience when mechanizing a printed paper. There are a lot of skipped steps and handwaving.
This is the common experience in replicating a published paper by hand ... it is common to find "obvious" aspects that are anything but.
The scary thing is when AIs generate unreadable formal proofs and then effectively lie (or fabulate, to be polite-ish) about the natural language version of the steps. Since the natural language version is arguably the most important aspect of a solution to a flagship problem, this fabulation deflates the value of the solution while the existence of the solution discourages further work on the problem.
I have hopes that this is primarily a matter of needing more engineering work on ergonomic formal languages and better building a language that "looks like math." e.g. when doing linear algebra stuff, a linear combination might be defined as a finitely supported function from an index set to your space, which is fine, but ugly and maybe conceptually overwhelming on first meeting, so I did some toying with little macros and eventually a small python Lean -> HTML renderer to do some basic transformations to make it look more like typical math notation with like \Sigma_{i \in I} a_i, or with a_0+...+a_n, etc. (to... not fantastic success, but I think there's still something to the idea).
I think a lot of math notation isn't wrong given a context, so in theory we should be able to translate it into something formal. Maybe also generate living documents where you can e.g. write `h : some_claim := by details(by rw[nat_mul_comm]; ...)` and the renderer hides details just like you'd write "obviously" in a traditional text. If the reader wants, they could then expand the details. etc. I found that many codex-generated proofs could be improved by telling it that I want a sequence of steps
have next_step := by <I don't care>
have therefore := by <still don't care>
So that the human proof appears as the left side, and I just ignore the right side as petty details. Again, not fantastic success, but better. Otherwise it goes very... Leanish by default.
Lean's VSCode plugin is I think only starting to explore the idea of a proper IDE for math. There's probably still tons of unexplored potential for like that fused with Matlab or whatever.
I enjoy running into those details when implementing papers, since it usually leads to improved understanding of the subject and an ability to approach the matter with more rigor in some way that I had not noticed before. It does also involve a lot of work and lost sleep though.
We should be very careful about relinquishing sorting through such details to AI.
Claude could not fix them without a lot of help, so i did not relinquish sorting through those details in general. Just the drudgery of grinding through proof obligations.
As a second rate scientist, nothing makes me happier than finding a "hot" paper in my field, reading it, converting it to code, and demonstrating the authors made systematic errors that mean the paper is more likely false than true.
I've been criticized for doing this, but to me it emphasizes how much attention goes to the hot, wrong papers.
I think this highlights that, at the very least, coverage of AI-generated proofs should describe them as "claims" to solve problems, until, like all other works, the community has had time to review and digest them.
The idea that an AI company is beyond peer review is harmful.
As far as I understand it, nobody is disputing the correctness of the Lean proof, or that it proves the conjecture it actually claims to prove. That's sufficient to consider the problem "solved". The natural language proof is a "nice to have".
The claim in TFA is that the formalization(in Lean) of the problem does not correspond to the natural language statement of the problem, such that the statement proven is not the conjecture for which proof is required for the problem to be considered "solved".
Does that contradict what I said? In that quote, it says that the NL proof does not correspond to the Lean proof. However, the statement of the theorem in Lean is independent from the NL proof. It comes from a DeepMind repository, which as far as I'm aware has been accepted by the community as a valid formalization of the original Clay Institute statement.
I would say it's released in the spirit of open source. "Peer review" in the narrow sense exists primarily to assign prestige in academia; but there's nothing stopping anyone from "peer reviewing" the GitHub repository.
What kind of prestige? Peer review is anonymous unpaid work.
A good review does not merely check the correctness of logical arguments, it gives suggestions for the exposition, citing the correct references, putting everything in the right context, etc.
All of the reasons you listed for peer review are valid. The broader point is that peer review can happen outside of academic journals, and nobody has an incentive to submit to them who isn't trying to play the academic prestige game. For a significant example, see the history of Perelman's proof of the Poincaré conjecture.
> journals are not the arbiter of truth and getting published in them is something only academics have an incentive to do
Good, I just wanted to point out that peer review isn't primarily an arbitrage of truth, it is also to make sure the exposition is nice to read. When you get a reviewer who actually cares, you receive lots of feedback that isn't related to the correctness of Lemma 3.14.15 and stuff like that.
This is an equivalent of a company producing security software, open sourcing their code, and then claiming that since no one has found any serious bugs, their software is secure.
No. The way to build confidence that your software is well made, you do a proper external security audit and obtain the requisite certificate from a proper auditing firm.
It's also incorrect to think peer review in mathematics is low quality (like it is in some other fields). Certainly, when major results are in place, editors ensure that high quality peer reviewers are recruited and do their job properly. Like all human processes this fails sometimes, but not enough to not do it.
>then claiming that since no one has found any serious bugs, their software is secure.
which specific openai statements does this part of your analogy map to?
in the "sharing ai progress in mathematics" blog, openai simply says "results", and never once claims that all of them are unquestionably true. instead, they state they want to evaluate the results. their github states that the results are "different stages of verification" and also says "Some of the unformalized results could have issues"
that is the opposite of "claiming [...] their software is secure", to use your analogy.
I didn't say peer review is low quality; just that it's not necessary or sufficient to determine the truth. Ultimately the OpenAI proof stands or falls on things that have been audited externally, namely the formalization of the problem in Lean and the correctness of the Lean software. There's no incentive for OpenAI to submit to a peer-reviewed journal when they don't need to play the academic prestige game. TFA is an example of peer review in action: they're analyzing the proof and finding points to criticize.
So they could also dump a 100 quadrillion line proof in Bourbaki notation and call it a day?
The proof was released in the spirit of being first at all costs without any attempt to clean it up. I doubt that OpenAI mathematicians could give a coherent talk about it, certainly not using a blackboard.
this happens in approximately every scientific field. ive never heard it described as "idea that they are beyond peer review".
openai themselves specifically call out that there may be issues with their results. journalists and laypeople just happen to skip that part, like they do with ~all physics, health, astronomy, etc results.
Because they want to release everything on github so everyone can peer review it themselves
This is far more efficient and they’re telling the academic industry to grow up
Sister comments are saying that academics dont like the Lean programming language and see a lack of human language described proof. Doesn’t sound like something I should care about but I’m watching for a better human language description of the problem as this discussion evolves
I've seen a lot of breathless reporting about various mathematical things being "proven" on the basis of the LLM-generated Lean formulation compiling. We probably wouldn't declare that for a human-written proof until peers had checked the proof for errors
This. The proof of Fermat’s Last Theorem took 15+ months to check. It’s absurd to see the media reporting that these big problems are solved based off of a news release and a hastily and mostly AI-written manuscript, and OpenAI et al. are all too happy to run with said breathless reporting.
Wiles' proof was informal and couldn't be checked by a computer. In this case, the experts need to check 300 lines of Lean code (mostly comments) and confirm that it formalizes the problem statement correctly. There are papers building on the solution and analyzing it for more general versions of the problem, which suggests that the PDE community has already accepted it and moved on.
there's breathless reporting of just about everything scientific. physics, astronomy, archaeology, etc. have this sort of thing all the time.
yet i have never seen anyone say "the idea that physicists are beyond peer review is harmful" because some mainstream news articles published a piece about dark energy or whatever.
Anyone who doesn't understand peer review (its intended workings, its negative effects by implementation flaws, etc.) automatically assumes expression is beyond academic peer review, so thats potentially a lot of people...
So my guess is that they have the AI system attempt to prove the theorem in natural language, then try to generate a Lean proof for it, and in that process they end up with a slightly different solution as the autoformalizer is essentially rewriting the NL proof to make it formalizable? Do we just need a "reverse pass" to re-align the NL proof with the Lean code?
Also, it doesn't seem that they are questioning the truthfulness of either proof, just that they are different?
Generating the lean proof first is a viable approach as well followed by an explanatory pass.
Actually, they are questioning whether the natural language description of the proof is either not faithful to the formal proof, or simply wrong, or both.
From the first example, it seems like this paper is so contrived. They pose a statement (y = x^3 - x^2 - 1 + 1 when x > -1) which is true, then provide incorrect reasoning but swapping the multiplicity of -1 and 1, then ask it to provide a proof. Essentially, they are running an injection attack; they give it 90% correct information, then it expects in good faith that -1 and 1 are not swapped, so it takes it verbatim. I don't see how the fact that ChatGPT can sometimes get this wrong, especially when the user is the bad actor trying to trick the computer and isn't actually trying to find a proof, is at all relevant to the Navier-Stokes solution.
Humans will have to wade through mountains of slop to decipher the argument. Alternatively, they could just ignore it like Mochizuki's ABC proof prior to the Scholze/Stix refutation.
Because it is really hard to read and the level of detail is so high that even lemmas that you can read may have such enormous levels of detail that makes real understanding difficult given that humans have limited working memory.
Same reason humans write code not only for a compiler to translate into machine code but also so other humans can understand what we write, learn from it, modify it etc...
> In particular, we highlight that the problem of resolving ambiguities in mathematical NL text, which is necessary in order to provide semantically faithful translation
This is what I've been wondering about with LLM proofs. Math is logical, but mathematical writing is still natural language: symbols get overloaded, conventions go unstated, and a lot rides on context. So a model can translate a statement into a formal system and prove it, and the proof can check out, while the statement it proved isn't quite the one the mathematician meant. I read this article as a caution that some of the LLM proofs announced so far may not hold up once a human checks what was actually proved. Is that a fair reading?
Natural language is ambiguous, but the Lean formalization is very well defined and unambiguous.
It's not the form language that is the real problem here. It's the ambiguity on the other side and the extreme difficulty of doing a useful and accurate translation.
This shouldn’t be a surprising result. We’ve known almost since LLMs became a thing that they can “prefer” modifying the terms or context of a problem when they can’t solve it directly (what one might call “cheating” if there were any volition involved). Often that happens in a way that isn’t immediately obvious to the user.
Before it was dropping databases or deleting repositories. Now it’s subtly changing the meaning of math problems to get a correct but irrelevant answer.
> The agents arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT‑6 Astra.
So it suggests that the formalization/verification step may have fixed some issues in the natural language proof, and either such differences were never noticed or the corrections weren't ported back to the NLP.
Indeed. I've never used AI to translate between natural language and Lean but I have gone from English to Golang, Python, Typescript and SQL and its interpretations can be... creative, let's say.
"In particular, we show that the formalised Lean proof does not correspond to the NL proof of blow-up of solutions to the Navier-Stokes equations."
So these authors seem to be claiming that OpenAI has not really proven Navier-Stokes at all. If I get their idea correctly, they are claiming that the LLM has not formalized the original "natural language" idea of Navier-Stokes correctly. If true, it would mean that their purported Lean proof is not actually a proof of Navier-Stokes at all, but something that is an incorrect translation of the original natural language idea. If correct, this is a really bold claim and I would like to see if other researchers agree.
No one is disputing that the Lean formaization of Navier-Stokes is correct, so we should have high confidence that the generated Lean proof is valid.
The authors are claiming that the Lean proof is not the same proof as the NL one. Therefore, we shouldn't yet have confidence that the NL proof is valid.
This is an important claim which the math community will need to work through. However, the Lean proof alone is sufficient for OpenAI to (reasonably confidently, leaving aside questions of academic manners) claim to have proven NS.
If the reason for the differences was done intentionally in Lean (as opposed to hallucinate e.g. m+4 vs m+5 as mentioned in remark 3.2), then a simple recording of differences, and then afterwards pass back any changes to the original NL would fix the issue. If it was hallucinated, then there is no guarantee it wouldn't keep hallucinating, and thus you might never end up with the same proof no matter how many passes you do back and forth (see remark 3.4).
These authors don't seem to be disputing that this Lean formalization of Navier-Stokes is correct. I don't think that gives us any new information about whether the generated Lean proof is or isn't a valid proof of this N-S blowup thing.
A proof of a theorem is different from the statement of the theorem. OpenAI has a Lean proof of the statement. That is all they need. There may be many different proofs of this statement, including NL proofs. It does not matter that these NL proofs may or may not be different from the Lean proof, at least for the correctness of the Lean proof. But of course the NL proof may be wrong. But who cares?
Indeed. If only some of those people would see the irony.
What matters most of all, as any first year student of mathematics would know, is whether the formal problem statement corresponds to the NL statement. TFA specifically states that at least some of the allegedly proven formal statements DO NOT.
Does the paper give a single example of one of the OpenAI solved theorems with a Lean certificate where the Lean statement does not correspond to the actual statement from the mathematical literature? I don't think so, but in case I am wrong, feel free to provide that example.
https://github.com/google-deepmind/formal-conjectures/blob/8...
;)
Did you mean “not…correctly”?
NS is continuous approximation to what is otherwise a discrete system. Particle collisions are discrete time events that are averaged over time. NS loses accuracy for very, very, very very low fluid densities and energies.
AI "proving" that this approximation can numerically "blow" up does not mean the approximation loses validity.
I don’t think that actually explains the idea of a momentum density transport equation well at all.
From the paper: "A third possibility is that the NL proof provides stronger statements than what the formal proof actually establishes, with (of course) different proofs. The latter happens in OpenAI’s announced proof of blow-up of Navier Stokes equations."
What the examples seem to show is that the proof method is different between the natural language proof and the lean proof. Which, if the lean proof actually proves blowup, would suggest that the natural language proof is subtly wrong, but the strategy was close enough to be used to create a real lean proof.
A little worrying, but part of the purpose of formalizing things in Lean, it forces you to be more accurate than natural language does. It's surprisingly common for major theorems to have slight inaccuracies early on that can be repaired. Famously, the initial proof of Fermat's Last Theorem had a flaw that took a year to repair (though I think that's unusually difficult).
So the most fundamental question is: does the Lean theorem faithfully state the right theorem?
For what it's worth the initial lean specifications for the top-level theorems generally come from human written formalizations such as in https://github.com/leanprover-community/mathlib4/blob/021ce6... so we can be reasonably confident about their correctness.
No, the other way around. The natural language proof was derived from the lean code, badly. This is my experience with using claude and lean to prove things. Its natural language explanations drift a lot from the lean, both before and after. But the lean code is the lean code.
Was it? Are you claiming a LLM does reasoning in lean or what? Since this (and all the other proofs by OpenAI etc) have been in the reverse order [1]:
> The agents arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT‑6 Astra.
[1]: https://openai.com/index/navier-stokes-solution/
That is definitely interesting because how do you know the 88 hours of work are correct before you throw another 17 hours of lean formalization work on it? You could end up just finding out there was some hallucination in the original work.
If your code compiles, are you sure it's bug free?
Need to prove a forall statement? forall x, P(x) is the same as a function taking x and returning the proof that P(x) is true.
Need to prove an exists statement? Create the pair (x, h) that gives the actual x that proves the exists, along with a proof that it satisfies the property you claim.
Maybe the only weird thing is that there are types and sets, so sets are kind of automatically more of a "subset" of some type.
The actual Mathlib is more generic, but once you get a hang of writing definitions (as you do in intro proofs), I've found that you can pretty naturally translate whatever you'd have in your undergrad notes. And undergrad should cover defining integers, rationals, reals, relations, functions, sequences, limits, derivatives, integrals, etc. Even if they've never studied solving PDEs, they'd have to take multivariable calculus and know enough to be able to write one (assuming they take at least single variable analysis+linear algebra)?
The proofs can get involved and tedious with all of the extra bookkeeping, or techniques to try to reduce the bookkeeping (tactics, etc). But the definitions and statements are pretty much what you'd expect.
Humans have made similar mistakes too. A human writes a specification for how things should work, the human translates that into code, the code does not work, and finally the human fixes the code and forgets to fix the original spec.
So the Lean proves something and the question is whether that something is actually what we care about — or something similar, but ultimately not the question.
Moreover, the paper claims that the NL arguments of Navier-Stokes are stronger than the Lean ones. My understanding is that the translator LLM got lazy and wrote the minimal amount of code that satisfied the theorem without the extra stronger claims.
It is common in mathematical papers to say "And by the way, this actually proves [stronger claim]", but this is something an AI with a precise goal of performing a translation would never do, as it's goal is to translate the proof, not to quality mathematics.
That said, a gap between the Lean proof and the pdf is annoying for interpretability, and interpretation is a valid aim, but that does not factor into the proof's validity.
Really? You think 300 lines of Lean code [1] is just as hard as the proof (or even remotely close)? Also note, as the README says [2], that the theorem was written independently by formal conjectures, not by the LLM.
[1]:https://github.com/openai/NavierStokesAndEuler/blob/main/Com...
[2]: https://github.com/openai/NavierStokesAndEuler/blob/main/Com...
_Assuming_ two failure modes:
- The lean kernel could always have a bug. - The formalized statement may not correspond to what _mathematicians_ "actually wanted"
It seems natural to make the argument of, "Well, even if you make the argument that the proof can have mistakes, it's surely easier to check the problem statement of something rather than the solution".
(A "nice property" is that, the agent doesn't need to even get "subarguments correct" according to the _second_ criteria - maybe in the natural proof it invents an object subtly different from the formal one, but it all checks out. If you guarantee that the _original_ statement corresponds, then the only possibility is the lean kernel. So it doesn't recurse infinitely, in this case).
But "definitions" are always a really weird thing that I don't think we have good theories for? How do you quantify how much descriptive power you need to express a question? Often times in math, the hard part is getting the definition right - but what if the definition itself starts to become so complex and unverifiable that no one can correspond that to anything? Well, it seems like many interesting long-standing math problems have "relatively" simple problem statements, in such a way that you could formalize it to lean easily, but not sure if there's really a silver bullet w/ lean or if it's going to be turtles all the way down.
It probably doesn't matter as long as AI keeps skyrocketing on the much more general property that is "intelligence", but still. Interesting to think about.
(Well, this is where AIT gets actually interesting, but still, I don't think its a generalized theory of semantics.)
Can someone tell me in simple terms why this doesn't conflict with the incompleteness theorems?
edit: thanks for the responses, i feel slightly less dumb now
We know as a consequence of Goedel theorems (at least I believe so), that there is no algorithm that would take a statement and output a proof if it is provable or a counterexample if it is not. However, AI provers never give anything for sure, so I think there is no contradiction here.
The formalization went through, but there were _several_ mistakes in the original paper that it uncovered, from type setting errors to (many) formulas that quantified over all resources as printed, but actually applied to only arising resources in the calculus..
So the formalization did give me a formally verified borrow checker that I could use to build a programming language on top of, but it was _not_ exactly the borrow calculus that was printed in the paper.
I expect this is the most common experience when mechanizing a printed paper. There are a lot of skipped steps and handwaving.
The scary thing is when AIs generate unreadable formal proofs and then effectively lie (or fabulate, to be polite-ish) about the natural language version of the steps. Since the natural language version is arguably the most important aspect of a solution to a flagship problem, this fabulation deflates the value of the solution while the existence of the solution discourages further work on the problem.
I think a lot of math notation isn't wrong given a context, so in theory we should be able to translate it into something formal. Maybe also generate living documents where you can e.g. write `h : some_claim := by details(by rw[nat_mul_comm]; ...)` and the renderer hides details just like you'd write "obviously" in a traditional text. If the reader wants, they could then expand the details. etc. I found that many codex-generated proofs could be improved by telling it that I want a sequence of steps
So that the human proof appears as the left side, and I just ignore the right side as petty details. Again, not fantastic success, but better. Otherwise it goes very... Leanish by default.Lean's VSCode plugin is I think only starting to explore the idea of a proper IDE for math. There's probably still tons of unexplored potential for like that fused with Matlab or whatever.
We should be very careful about relinquishing sorting through such details to AI.
I've been criticized for doing this, but to me it emphasizes how much attention goes to the hot, wrong papers.
The idea that an AI company is beyond peer review is harmful.
that's not the claim. the formal statement of the problem for the NS proof was written by humans not autoformalized.
https://github.com/google-deepmind/formal-conjectures/blob/8...
Maybe read the comment before replying, at a minimum.
i havent seen this sentiment expressed anywhere, have you?
isn't this comment chain on a submission about openai's claims being reviewed?
A good review does not merely check the correctness of logical arguments, it gives suggestions for the exposition, citing the correct references, putting everything in the right context, etc.
Prestige to the reviewed, not to the reviewer.
Good, I just wanted to point out that peer review isn't primarily an arbitrage of truth, it is also to make sure the exposition is nice to read. When you get a reviewer who actually cares, you receive lots of feedback that isn't related to the correctness of Lemma 3.14.15 and stuff like that.
No. The way to build confidence that your software is well made, you do a proper external security audit and obtain the requisite certificate from a proper auditing firm.
It's also incorrect to think peer review in mathematics is low quality (like it is in some other fields). Certainly, when major results are in place, editors ensure that high quality peer reviewers are recruited and do their job properly. Like all human processes this fails sometimes, but not enough to not do it.
which specific openai statements does this part of your analogy map to?
in the "sharing ai progress in mathematics" blog, openai simply says "results", and never once claims that all of them are unquestionably true. instead, they state they want to evaluate the results. their github states that the results are "different stages of verification" and also says "Some of the unformalized results could have issues"
that is the opposite of "claiming [...] their software is secure", to use your analogy.
The proof was released in the spirit of being first at all costs without any attempt to clean it up. I doubt that OpenAI mathematicians could give a coherent talk about it, certainly not using a blackboard.
are people not reviewing openai claims right now?
openai themselves specifically call out that there may be issues with their results. journalists and laypeople just happen to skip that part, like they do with ~all physics, health, astronomy, etc results.
Peer review is a proxy for correctness.
Peer review journal is a proxy for quality peer review, or at least it was, once upon a time.
This is far more efficient and they’re telling the academic industry to grow up
Sister comments are saying that academics dont like the Lean programming language and see a lack of human language described proof. Doesn’t sound like something I should care about but I’m watching for a better human language description of the problem as this discussion evolves
yet i have never seen anyone say "the idea that physicists are beyond peer review is harmful" because some mainstream news articles published a piece about dark energy or whatever.
Also, it doesn't seem that they are questioning the truthfulness of either proof, just that they are different?
Actually, they are questioning whether the natural language description of the proof is either not faithful to the formal proof, or simply wrong, or both.
https://terrytao.wordpress.com/2026/10/04/on-classical-solut...
Humans will have to wade through mountains of slop to decipher the argument. Alternatively, they could just ignore it like Mochizuki's ABC proof prior to the Scholze/Stix refutation.
The point of the article is that natural language is not these things.
If you understand the Lean, then you can create a NL proof. The LLM clearly doesn't understand the Lean code it produced.
It seems the AI wrote code in Lean that proves there are solutions to Navier-Stokes that can blow up, but...
the AI's explanation of the code, in natural language, does not correspond to the Lean proof!
That is... so rich with irony.
This is what I've been wondering about with LLM proofs. Math is logical, but mathematical writing is still natural language: symbols get overloaded, conventions go unstated, and a lot rides on context. So a model can translate a statement into a formal system and prove it, and the proof can check out, while the statement it proved isn't quite the one the mathematician meant. I read this article as a caution that some of the LLM proofs announced so far may not hold up once a human checks what was actually proved. Is that a fair reading?
Edit out vulgarity
It's not the form language that is the real problem here. It's the ambiguity on the other side and the extreme difficulty of doing a useful and accurate translation.
> gotcha bitch!
You may have misdiagnosed the problem.
Before it was dropping databases or deleting repositories. Now it’s subtly changing the meaning of math problems to get a correct but irrelevant answer.
> The agents arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT‑6 Astra.
So it suggests that the formalization/verification step may have fixed some issues in the natural language proof, and either such differences were never noticed or the corrections weren't ported back to the NLP.
Oh, well. I suppose I should avoid getting involved in these AI threads, but now it’s about half the forum.