A SAT Attack on Tarski's High School Algebra Problem
arxiv.org
arxiv.org
here for example, instead of saying {f(x,y,z)==g(x,y,z)}, authors instead make variable group a_w:=(f(x,y,z)=w||g(x,y,z)=w), and then apply "at most 1" to it. Can't be unequal if both functions only can have 1 result in total
this adds an index to iterate over, but separates internal subexpressions of f() and g(), removing 2 indixes (in this problem) and thus dropping whole power of n of clauses
---
what I don't get is that they aren't searching Tarski's problem per se, but for one specific solution to it (one identity that isn't resulting from given). I'd totally look for arithmetic models that violate expectations in other ways than Wilkie
Tarski said "hey, what if we limit ourself only to positive integers? subtraction/division doesn't always work, so let's get rid of it. What can we say about +^? there are some obvious properties, like x+y=y+x for whatever x and y. Is that enough to describe all the equalities?".
Wilkie answered No, and a whole different hunt began - trying to find a set of pseudo "numbers" with pseudo "+^", that would have the same obvious properties, but not the non-obvious one
and SAT-solving is essentially asking computer to find things by describing what they are not - "1+2 is not invalid and can't be 2 different values", "1+(1+1) must not differ from (1+1)+1". A lot of such properties are easily expressible as "not the opposite" - but devil's in the details and some choices of opposites are faster for computer than others. Also symmetries are evil :)
Note that it's not obvious at all how to search the space of possible exotic identities, so a potential direction for future work would be to have a loop in which LLMs (perhaps through something like AlphaEvolve) propose an exotic identity, and then a SAT call tries to find a small countermodel. A big issue though, is that even with our efficient encoding the SAT calls would take at least a few minutes, so we wouldn't be able to afford testing millions of candidates. But if there were 1000 candidates that could potentially be tested...
(I'm one of the authors of the paper, thanks for the coment!).
I wish you spent at least a couple words in the paper about that. Even negative results are worth documenting! (even if you didn't get up to size 11, I'd've loved to hear about those other alternatives)
---
by the way, you mentioned that decreasing HSI6/10 from O(n^6) to O(n^5) clauses was slower - how big was the slowdown and how much less total clauses were there in that encoding? if I understand it correctly, that was still the biggest clause maker, but by how much?
---
also, have you tried reordering order of operations in symmetry break? how much did it affect the search? I wonder if unique multiplication table might've been of help had it been disambiguated stronger (or weaker)
That makes sense; it just happened that we tried the other identities after having written and submitted the paper. More importantly, the variants of Wilkie's identity we tried were suggested to us by an expert on the topic; I have just sent an email asking if they are okay with us sharing them, and if so I will post a link here.
> how big was the slowdown and how much less total clauses were there in that encoding? if I understand it correctly, that was still the biggest clause maker, but by how much?
It was roughly a factor of 8 fewer clauses, and yet over 5 times slower. If you're interested in the design of compact CNF encodings, and their effects on runtime, that's exactly the topic of my PhD thesis, and this proposal might give an initial idea: https://bsubercaseaux.github.io/assets/pdf/proposal.pdf
Naturally, there could be another encoding that has fewer clauses (say, O(n^5) or even O(n^4)) and does perform better in practice. But we didn't come up with one.
> also, have you tried reordering order of operations in symmetry break? how much did it affect the search?
To some extent. It was of moderate impact in terms of the runtime, but presumably the enumeration up to isomorphism would have been harder if we didn't consider the addition variables first. To have some updated numbers, I just ran some experiments with the 3! = 6 permutations of {A, M, E} on n=10.
AME (used in the paper) -> 92.7s AEM -> 110.8s MAE -> 139.4s MEA -> 157.4s EAM -> 297.1s EMA -> 276.8s
https://satcompetition.github.io/2026/downloads/satcomp26sli...
See slide 18. It was really nice to see him win. I have always been a huge fan of symmetry breaking, CryptoMiniSat (that I develop) had a symmetry breaking system, BreakID (which is _way_ slower than Markus' satsuma), in it for many years now.
Similarly, I think a * b and a / b can be replaced with the same trick, but then I realized it may not work on non-abelian, or where multiplicative inverse is not available...
(Edit: positive integers aren’t exactly a semiring because 0 is excluded, although some authors do define a semiring without the requirement of an additive identity element.)
If we allow negatives but disallow exponentiation, that's decidable. If we allow negatives and partial exponentiation that stays within the integers, that's undecidable again.
In this case, we have the much simpler equational theory of positive integers under addition, multiplication, and exponentiation, which does not include any quantifiers. In fact, Gurevič showed that this theory is decidable [0]. On the other hand, Gurevič later showed that this theory is not finitely axiomatizable [1], so an infinite (but still computable) set of axioms is needed to fully characterize the theory.
[0] R. Gurevič, Equational theory of positive numbers with exponentiation, 1985, https://doi.org/10.2307/2044966
[1] R. Gurevič, Equational theory of positive numbers with exponentiation is not finitely axiomatizable, 1990, https://doi.org/10.1016/0168-0072(90)90049-8
> […] Of course, this conclusion is not proved mathematically. It is possible that the programs have some bugs, or the user (myself made some errors.
They address it as follows:
> We address this [… through] ChatGPT 5.5 Pro, through Codex, to automatically generate a Lean formalization that we then checked ourselves to confirm the statements and definitions indeed match their expected semantics. This process took multiple iterations and discussions with the model over several days, and generated over 10,000 lines of code.
This is extremely unconvincing. Manual review of 10,000 lines of AI generated code is a terrifying prospect. I’m sure the result is correct, however I am deeply uncomfortable with this being mankind’s new mathematical process. Similar concerns have been expressed since the days of the 4 color theorem, but this feels different. Perhaps it is just new.
>in a nutshell, it defines an executable function encode that takes a natural number n ≥ 5 and emits a CNF formula, which is byte-for-byte equal to the output of our Python encoding
The phrase “byte-for-byte equal” makes my eye twitch these days.
With https://en.wikipedia.org/wiki/Lean_(proof_assistant) (and other proof assistants), you need to review only the lines that correspond to the theorem that you want to prove and their types (I am not very experienced when it comes to lean, but I would assume that comes down to a few hundred lines of code, at most). The rest is left to typechecking (which, I would expect many in the field to agree, is at as reliable than your average peer review process in professional mathematics, and likely much more). That's the reason why Lean4 is making such a fuzz now.
That itself is not trivial too, but way easier than reviewing every function and definition used to prove that the theorems have indeed the types they claim.
If one accepts the proof of the https://en.wikipedia.org/wiki/Four_color_theorem, then there should not be new reservations these proofs; except from the maybe new additional failure scenario that the authors (still correctly!) proved theorems that don't state what they think they stated.
To sum it up: There is IMO no domain more suited for using LLMs than mathematical proofs that can be formalized using Lean4. The fact the hype-circle started earlier in software than in maths is due to the difference in monetary incentives I would assume. (Or another, rather radical and not really serious phrasing: "When it comes to Lean4 proofs that typechecks, there is no AI slop" - the theorem being proven might be uninteresting, but the proof itself is very very very very likely to be correct)
> “byte-for-byte equal”
The term itself or its association with LLMs? I would get the latter, if its the former: It's an desirable property to have, I always like seeing people going that far (assuming obviously that they indeed did so, and in the places where it matters!)
This is (unfortunately) not actually the case—just a few weeks ago, someone "proved" the Collatz conjecture via a Lean proof 1) whose theorem statement was correct, 2) typechecked, and 3) was even verified by external tools with their own implementations of the kernel.[0]
The problem was (AFAIK) that the Lean kernel has a lot of fancy features that aren't yet perfectly understood from a type-theoretic perspective (I don't think Lean is unique in this regard; pretty sure Rocq and Agda are in a similar situation). And so when the kernel implements some feature whose soundness isn't guaranteed, the independent verification tools (or at least some of them) follow suit, and now any issues in the former affect the latter as well.
[0] See e.g. https://x.com/gro_tsen/status/2082483878480977959 for more detail
Yes, in hindsight it makes sense that a eager LLM would exploit bugs in the kernel itself. I have not read the details yet (but want to!) and assume that it is rather related to the layers directly before or after the type-theory core; i.e. that the AI managed to get a correct typecheck by sidestepping a check during in one of the translation steps somehow, by manipulating the kernels result or by going totally hacker mode and swap implementation/overwrite memory; that is just speculation on my part though, I mainly base if on what I know about Haskell and the https://en.wikipedia.org/wiki/Calculus_of_constructions in general that have not that many constructs, and I would assume the variants used in Rocq and Lean are proven sound.
For experimental features where soundness is not proven, all bets are off IMO, but of course, that does not stop bad actors from engagement knowlingly abusing unsoundness for their own gain. Fortunately, that kind of manipulation is easy to verify when one has access to the codebase; I am would assume that something like lints for experimental features exist, or better even, something like a "sound mode". That does not help against hacking the kernel machinery though.
Concluding notes: - In this particular instance, the Collatz conjecture was appearently chosen on purpose to demonstrate the Kernel bug, from the same thread: "Not that this changes any of the above, but I am informed that the person posting the proof was actually aware that this was a Lean kernel soundness bug, and it was not intended to be taken seriously as a solution to Collatz's problem." - IF the code is made public (and it would be highly suspicious if parts of a proof where hidden), then I would assume this kind of hack is 1) easier to spot that other kinds of hack, since the asset-under-attack is really small and 2) there is not so much incentive to use much time/ingenuity/tokens on finding those hacks (the more are found and fixed, the better of course) 3) and they should be easy to defend against, I would assume; my first thought would be the flag I mentioned above that simply forbids all non-sound features, at the cost of limiting expressive power.
----
Edit: I checked, and I think it alleviates my worries in the sense that the bug(s) where not in the type theory or its implementation, but rather the machinery around it. Further context below:
That is the incident description (also linked in the x thread you linked, for future readers: https://infosec.exchange/@0xabad1dea/117002106099986943).
The following is from that thread or links from it:
- "Fixes two things: (1) more strict/nuanced handling for structure/proj interactions, and (2) adds methods for enforcing that generated auxiliary data for inductives, constructors, and recursors are more strictly checked against the assertions in the export file." | That is the fix to the non-Lean-kernel that was mentioned. To me, (2) looks firmly like what I meant with "supporting machinery", regarding (1), I do not know enough to have an opinion about it (i.e. how on what layer those interactions happen(ed)) | https://github.com/ammkrn/nanoda_lib/pull/22
- "For example, pipeline wedges (execute this instruction and the core freezes and never executes another instruction) would not be found by these techniques..." | Power and limits of Lean | https://infosec.exchange/@david_chisnall/117003914014196496
- "@mario @shelldozer it very well may be the most formally correct piece of software we've ever produced, but keep in mind it still has to run on a physical computer it's sharing with less-verified software and is also vulnerable to things like Rowhammer-class ram corruption attacks if one wants to intentionally manipulate it. There's no final escape hatch beyond which a computer can be absolutely guaranteed to always compute the correct answer, especially when someone has a vested interest in getting it to output the wrong answer." | Computer-checked proofs run on computers, which brings its own attack vectors, independent of how well the kernel is written | https://infosec.exchange/@0xabad1dea/117002712346315184
personal, cautios takeaway after reading the details: If you use Lean4 to write proofs, or read that a reputable group of mathematicians publizised a Lean4 proof, you are still highly unlikely to be fooled by a bug, and if Fable 5 decides to exploit a 0day in the core Lean4 machinery, that should still be able to be caught quickly.
Personally, this is one of the cases with LLM-generated Lean that I feel most comfortable with: the proofs are all mathematically simple, and we included an example of correctness for one of the SAT constraints in the paper to show that there is nothing complicated about them. Writing them all in Lean, however, would be a very tedious task for which I see no particular benefit (i.e., it's not like some other formalization projects in which one learns something by formalizing; for an example of a human-written Lean correctness proof of a SAT encoding, you can check the paper https://arxiv.org/abs/2403.17370 in which I participated).
Regarding the byte-for-byte equality, the point is that we first had the Python code (which we wrote manually), and then asked LLMs to implement an encoder in Lean that would not only produce "equivalent" formulas but exactly equal formulas, meaning that it would use the same variable indices and the same clauses in the same order. It is well-known that such superficial factors (e.g., clause or variable ordering) can actually have large effects on solver performance (see https://www.cs.utexas.edu/~marijn/publications/pos18.pdf). The Python code is not part of the trust chain, and the Lean code does not rely on it in any way. We included the Python code in the repository since it's much shorter and easier to read.