I spent some time exploring this topic.
Here's my thesis:
Formal verification was expensive. 20x expensive compared to just developing the software, as the author notes.
The cost of finding and developing exploits also was high.
That creates an incentive to put software verification aside, since it solves a relatively small problem, at an extremely high cost.
We've seen how Mythos has found more vulnerabilities than the rest of the security industry combined. (you can argue about the quality and what counts as a vulnerability, but not the point)
So the cost of finding and developing exploits has dropped dramatically.
On the other hand, formal verification is now much easier, since LLMs can automate the proof.
You don't even need to worry about hallucinations, you merely need to trust Lean core.
If the LLM is wrong, the proof will get rejected!
The problem of exploits gets bigger, and the solution of formal verification gets cheaper.
As a result, the needle is now moving in the direction of "more formal verification".
I, personally, think that it is ridiculous that ~none of the software we use is known to work correctly. It just happens to work correctly, most of the time.
My (ambitious) goal is to have a self-hosting, formally verified compiler, which allows proof transfer from source code down to assembly.
I have not achieved that goal yet.
What I have so far:
- one (non-optimized) compiler step which is formally verified
- three simple functions (hex, hex with labels, strtoull) formally verified, against RISC-V assembly, and against a custom IR
Don't we still need to verify that what the LLM proved is actually the stated system?If agents cutting corners (like deleting tests) is a concern, how can we be sure that the verification corresponds to the software and that every _load bearing_ assumption is true? I don't think that simply trusting the Lean core is enough.
Strongly agree with the author here. The future will belong to programming languages that natively embed theorem proofers into their type systems so LLMs can forego a lot of testing by just validating the implementations they write against the specs with formal proofs. Writing formal specs is probably the main skill a programmer in the future will need to get work done.
Verus (https://github.com/verus-lang/verus) is a good start for the rust ecosystem, but it's essentially a standalone language today (with custom syntax and type system).
This is the fantasy that has always driven proof systems research. Nobody is going to run software that has never been tested. Would ride a rollercoaster that had never actually been tested before, only "proven" safe? I wouldn't!
So this stuff is always going to be additive and concerned with edge cases, as almost by definition, stuff that isn't edge cases will be found by comprehensive enough testing procedures.
And yet most software doesn't really need to be correct under edge cases, outside of security and data loss issues. People can tolerate a lot of incorrectness in other areas because it's just annoying, not critical.
Security is a case where formal methods could help, but I don't think LLMs will change the industries lack of interest. If anything it'll reduce it even further. Historically security took place in a fog of war. You don't know your enemies capabilities and may not be able to easily match them. But now LLMs are better at finding security bugs than most (all?) humans and ~everyone has access to them, so, from a liability perspective, all you have to do is point a frontier LLM at your codebase and let it fix as many bugs as it can find. Your enemies don't have access to anything better, so once it's done you can tick the box and say security is good enough. Meaning, nobody will fire you if there are still attacks possible.
So in the end I don't see LLMs changing the adoption formal methods.
Lean has dependant types. Wouldn't something like Haskell or Idris, that are trying to be general purpose dependantly typed languages--wouldn't they be a better start than versus?
Versus appears to just be a formal verification tool. Perhaps I misunderstand?
You want the formal verification built into the language because the tooling can start to get really crazy good. Agda is the dependantly typed language I've used the most (long ago), and the tooling was interactive in a helpful way I've never experienced with other languages.
You don't want a separate language used to verify a base language, because then everyone ends up having to know two languages. Looking at the history of computing though, I wouldn't be surprised if this happens.
The actual programming language and the verification language can be the same language though, if we want.
in case anyone's interested i have a vibe coded fork of verus that replaces the verus-the-language side of verus with plain old Lean 4. It's still two languages, but now at least the second language is as mainstream as it gets in the field and has good automation. i haven't finished wiring up the Lean 4 infoview and vs code extensions and LLM skills into it yet, which makes it not as easy to write yet as lean 4 with the IDE bells and whistles.
I'm also playing around with using the lean's compile-to-C tooling to instead compile to rust instead and it's getting more of my focus than the lean-via-verus route right now.
if people are interested, ping me and i can put them up on gh.
As a meta-comment on the topic, something I have noticed is that there still exists confusion what it means to use theorem provers for projects -- the other day I read a tweet from Paradigm, a crypto-VC now seemingly AI-pilled. Some LP of theirs had made a Lean 4 formalization of the Ethereum's virtual machine. The tweet said this would have cost like $150k in API tokens ("would have", as in, I guess they get theirs for free), and took a week of inference time for an LLM to produce. I somehow got distracted to actually take a look at the code, which I found rather light on theorems. Nor did the project make use of Batteries or Mathlib which are arguably the one of the strongest motivation for me personally to use Lean4. That is, I generally rather rely on someone else getting the category theory and algebraic structures right, which then leaves me the proof obligation to show the correspondence with whatever toy I'm working on. Here I'm fine to use LLMs for proof search, very similar to how would I use a SMT solver. But what I have found is that the language models have to be really coerced into using these libraries, because otherwise the models much rather overfit and overclaim a solution with a 3 minute inference task rather than attempt to fulfill the proof obligations over 3 hours. And I feel nauseated when I need to convince the LLM (I use Claude) that filling the proof obligation is for "academic exercise" or because I'm coerced into doing so, because otherwise it will come up with reasons of its own why it does not want to do it. Now, this happens under the mental model in which I'm interested in finding equivalences with prior work. Many LLM generated Lean code reads more as if someone was interested whether X can be turned into a Lean 4 program, which is mostly yes, and that in general is a positive thing. But, if you are not interested in refinement types and theorems, why not just choose Haskell? The point is, I strongly sense that unless you have good questions to ask, then that's very evident in these languages. And, this is something the LLM won't help you -- if you don't impose a proof obligation for it, it certainly will not try to go the extra mile to conjure one for you.
Crypto guys are in the business of grifting magic beans to rubes. They are mostly interested in the aesthetics of these technologies. Using an academic programming language gives these scientific magical vibes and such that make people believe these guys are high tech and know their shit and so on. This is the reason many crypto projects do use Haskell too, it's known as an academic research language that is difficult to use and "if it compiles it works." They come at it from the perspective of "What would sound the best when I try to sell the space money of the future?"
Blockchain protocols are also faced with the most active, persistent, and well-funded state sponsored attackers of any industry, so naturally formal verification should be of great interest.
I have already written it, and I will write it again: dependent types and total functions do not scale. Maintenance is terrible.
Suppose that you have managed to write a non-trivial piece of software with dependent types encoding all sorts of properties everywhere. The actual computation and proof are intermingled. Think of the author's innocent bound-check proof in the zstd decoder, but across the whole program, with more elaborate properties and longer proofs.
Suddenly, you realize that you need to prove a new property of your program. Can you keep your existing work and build on top of it? In general, no, you have to refine every dependent type everywhere by adding a new conjunct expressing a new invariant, and adapt every proof, as the new property is threaded in the existing program.
That is because dependent types (and other staples of naive approaches to proving program properties, like a unique invariant per loop) structure the program along the wrong dimension: they encourage grouping everything that concerns a value ("put this value in a dependent type that encodes everything known about it") or a program point ("write the precondition for this function as a big conjunction mixing all the concerns"), where it works much better, for long-term maintenance, to structure the development along concerns: computational parts of the program, basic functional properties and absence of UB, termination, other functional properties, security, etc., where each layer builds on top of the previous ones without requiring them to change.
That is not to say that dependent types do not have their use. Where they shine is a module boundaries. Consider a library that exposes an opaque type and operations on it. Users of the library can only build and modify values of that type through the library's API. This type should be a dependent type. If it needs to be refined at any point to encode a new invariant, this will have no impact on the library's users, since all they do is pass around values without interpreting them. However, inside the library, I would recommend unpacking/repacking the dependent type at the library's entry points and handling the concerns separately.
I'm convinced Math is the canary for what's going to happen to knowledge work. Coding was ahead in harnessing early LLM capability, but Math, given it's pure form, has aleady racing ahead.
The dimensions to notice are: Research speedup, practitioner expertize, labour dynamics, junior entrants, world impact.
In that sense it's a canary, whatever happens to Math, will in order flow to other sciences in-order of purity: Computer science, physics, chem, bio, social sciences.
Math is the opposite of a canary because it's possible to generate infinite amounts of synthetic training data for it, unlike almost every other kind of knowledge work where correctness depends on external input.
This is really cool stuff, and I agree the future is likely to look more like this. I was surprised by the last two paragraphs ("Aside: verified assembly") -- my understanding was that this future is basically already here. I believe agl's colleagues at Google have already deployed some auto-mutated verified assembly versions of some crypto routines, based on the Fiat Crypto + CryptOpt work (https://arxiv.org/pdf/2211.10665), both involving Andres Erbsen who I think is currently at Google before starting a professorship soon. I dunno if this work would count as "cheap" (it looks like Fig. 10 unfolds over the course of a day) but spending a day auto-exploring many verified-correct machine code implementations of the same routine to find the fastest one (which you then ship forever) doesn't seem impractically expensive either.
I'm very bullish on proof automation as well. I'm currently researching AI for algorithm design and using automated theorem provers to get formal guarantees for generated algorithms.
To make a shameless plug, I'm working on a Python package called OpenATP [1] to make it easy to benchmark different models/harnesses for automated theorem proving. It supports running agents in Docker containers or Modal out of the box. If you try it out, I'd love to get your feedback!
I recently wrote about the surprisingly good performance I saw from Grok [2]. On more challenging proofs, Grok doesn't keep up with Opus/Fable and GPT 5.6. I was recently blown away by GPT 5.6 Sol. It's persistence in closing out proofs is unparalleled from what I've seen so far. OpenATP also supports Kimi and Leanstral [3], among others.
> We now have LLMs which, combined with proof irrelevance, promise to be an extremely capable form of proof automation. With sufficient amounts of automation perhaps you don't need to worry about proof engineering nearly so much. You still need to avoid blowing up the type checker but, in my limited tests, LLMs can avoid that. Potentially, LLMs suddenly make dependent-type systems dramatically more practical.
When I've used interactive proof systems like Roq, I'd often kinda code myself into a hole by cutting along the wrong axes and not specifying my problem in a way that's easy to prove. After all, this is just like in math: you really want to cut at a problem the right way to get to the easy proof.
I think people are discounting how important that decision making is. It's not just about whether an LLM can churn through specific proof strategies on a problem, but also about how to pose the problem etc.
I'm not saying LLMs can't help, but I think it's less that "proof engineering is not needed" and more that "when these tools are used in the right way, proof engineering is easier". Because at the end of the day these tools work well when they have the right kind of foundations in the first place
> AWS made LNSym: a semantics and simulator for AArch64. That's cool. Perhaps we could use it to show equivalence between an optimised assembly implementation of some functions, and their Lean counterparts, and then use the assembly code at run-time? Then we could let LLMs rip at optimisation and they couldn't introduce any functional bugs. Verified assembly is well-trodden in crypto implementations, but perhaps now it could be cheap?
Trying to one-shot compcert might be hard! Thinking about it and planning it out might make it easier though...
Cool. Now we can write bugs in our theorem descriptions instead of source code.
Seriously, please review Curry Howard Isomorphism if you’re getting pulled down this rabbit hole.
Programs are proofs. Proofs are programs.
So if you can formally describe the correct output for every input, you can have an LLM loop automatically fill in the gaps of how to get there. Congrats, that sounds at least as hard as writing the correct program in most cases.
Don’t get me wrong, I do think there are useful tools combining formal methods and llms. Let’s just not get carried away.
It is true that finding the correct specification is a formidable task; knowing what correctness even means is arguably most of the difficulty of programming. However! "Moving bugs up from programs to types" isn't how this shakes out in practice, at all. Another commenter already noted that it's often much easier to communicate your intent through specifications, because you can essentially always say what a computation should do much more simply than you can say exactly how to do it.
I think it's also important not to miss the forest for the trees: even relatively simple specifications like "the compress and decompress functions must be inverses for all inputs" rules out vast classes of bugs in a compression library. This is not a complete specification; for instance, it does not speak about how the decompressor behaves on malicious input. But in my experience, even partial specifications carry the promise of hitting warp speed with LLMs in a way that I haven't seen anywhere else. After a certain level of specification, you have decent guarantees of being able to whole-heartedly forget about the implementation details of the synthesized program. And you get a better-built, more robust program out of it at the end!
The comment at the end of the article about having LLMs directly generate assembly against specifications and letting them rip with finding custom optimizations is the sort of crazy stuff this enables. I really think we're only seeing the tip of the iceberg here. People keep asking what we can do with LLMs that we couldn't before; this is the answer.
> Congrats, that sounds at least as hard as writing the correct program in most cases.
That's not remotely true. Or, more formally speaking since we're in a thread about proof assistants, it's not remotely true, up to extensional equality, plus some choices about which axioms you use.
I can write a formal description of what it means to have property in a way that does have computational content that is equivalent to an algorithm[0], but often the clearest way to express the property is equivalent to an algorithm that literally brute forces the problem, like sorting a thing by checking every permutation until you find one that's sorted.
The magic is that you can write a spec that's clear, then have the LLM write the code and prove the spec, so you know that given the right inputs/state, it will return the right outputs/state. Then the gap is performance-like characteristics, which is a pretty great starting point and a lot easier to be just empirical about than correctness.
[0] in Lean you can also use classical logic or add your own axioms, where it's not even comutational.
> That's not remotely true. Or, more formally speaking since we're in a thread about proof assistants, it's not remotely true, up to extensional equality, plus some choices about which axioms you use.
Really unnecessary levels of snark here.
> often the clearest way to express the property is equivalent to an algorithm that literally brute forces the problem, like sorting a thing by checking every permutation until you find one that's sorted
Have you ever heard of prolog? I’m sure a prolog program can express whatever property you’re attempting to write just as tersely (if that’s your metric for hard). Almost copy/pastable to and from a theorem proofer.
Or are you saying that the program has to be the efficient implementation? Because that’s a different ball game. I’m not even going to get into how you could provably transform brute force propositional logic into efficient algorithms. (At that point we’ll have finally created the fabled “sufficiently smart compiler” and probably solved p=np).
A huge majority of software is simple business rules + CRUD that is trivially verifiable. The entire problem is showing that an efficient/reliable program actually implements those rules.
> The magic is that you can write a spec that's clear
Maybe you can. But I did spend a grad class with rocq (coq at the time) and a decade working with “systems engineers” and am not convinced that this is a realistic expectation.
> I’m not even going to get into how you could provably transform brute force propositional logic into efficient algorithms.
> The entire problem is showing that an efficient/reliable program actually implements those rules.
Reliability is a standard matter of correctness and captured (partly) by specifications. Efficiency tends to be easy to empirically test, but it is also possible to capture at the specification level [1]. Mind that specifications need not be all-consuming.
> But I did spend a grad class with rocq (coq at the time) and a decade working with “systems engineers” and am not convinced that this is a realistic expectation.
Agree! But this stuff just got massively more accessible, and the tooling around it is growing quickly. I think we'll end up growing specification systems specific to various domains which will be palatable to those "systems engineers", but I err on the side of optimism here. There's definitely a lot left to do for practicality.
[1] See the work of https://cs.nyu.edu/~shw8119 for the case of provably-efficient parallelism and garbage collectors
> Really unnecessary levels of snark here. [..] Have you ever heard of prolog?
Why do you look at the speck of sawdust in your brother's eye and pay no attention to the plank in your own?
> I’m sure a prolog program can express whatever property you’re attempting to write just as tersely
No, you won't be able to express the most basic properties in Prolog at all, let alone as tersely as in a proper specification language.
E.g. if you have a language interpreter and a bytecode interpreter, pretty much every specification language will let you express the correctness of a compiler `compile(x)` as
Huh, I wasn't actually going for snark. I was trying to preemptively be over-specific about what I meant. As in, you said "... writing the correct program..." and I meant that I'm talking in terms where functions are equal or distinct only based on extensional equality and with some flexibility on computational interpretation and axioms of the logic.
Like, if you meant "the correct program" in a sense where two pure total functions can be different despite both having the same outputs on the same inputs, then that's not what I'm responding to.
And yeah I know prolog, but proof assistants and logic programming are totally different beasts. Definitely not copy/pastable to/from lean, at least as I've seen and used each.
> Or are you saying that the program has to be the efficient implementation? Because that’s a different ball game. I’m not even going to get into how you could provably transform brute force propositional logic into efficient algorithms. (At that point we’ll have finally created the fabled “sufficiently smart compiler” and probably solved p=np).
I think you're totally misunderstanding. What I'm saying is that in something like Lean (just because I know it best) I can say, "this function takes inputs satisfying Prop1 and returns outputs satisfying Prop2," in ways where I write some brute-force equivalent formalization of Prop1 and Prop2 in the most straightforward way, and then go on to prove that they are true of my program that is not the brute force implementation. Like the wacky famous magical inverse square root implementation from Quake III. You could write a spec that "output = 1/sqrt(input) up to float properties" and the implementation in the famous brainfuckery way. To your comment about how the spec is as hard as the implementation, "output = 1/sqrt(input)" is way easier than the weird efficient implementation, and that class of distinction is super common.
And as you said, the annoying part is showing that the efficient implementation satisfies the spec, but what's magic today is that we have great tools and LLMs can and do fill in the blanks. In practice, I write the spec by hand for the stuff I care about and then prompt the rest and know that my spec is what the LLM implemented.
> > The magic is that you can write a spec that's clear
> Maybe you can. But I did spend a grad class with rocq (coq at the time) and a decade working with “systems engineers” and am not convinced that this is a realistic expectation.
I tried rocq back when it was coq too, and now do most of my work in lean and rust and python with totally normal folks and I'm convinced that the tooling and languages are finally just about Good Enough. If you have any interest in the field, which it sounds like you do, and if you haven't checked out the ecosystem in the last couple years, I'd recommend you check it out again.
I agree with the core thesis that LLMs + theorem provers might make formal methods cheap enough to be practical in software development.
The biggest issue was always cost. But there's still an alignment problem. Without human supervision, things might drift away from the original specification and intent.
From my own experience, what works best is some kind of Hoare/separation logic (contracts), as these are quite easy to follow and decompose.
Even something as simple as a minimal Haskell subset, plus a bit of LiquidHaskell, can get you really far if you are pragmatic.
Types, preconditions and theorems are still a guardrail
Agents perform better when programming in Rust than <insert dynamic language here> (at least in my experience), exactly because types matter. Those theorem provers just have a poweful and pedantic type system that insists in being always correct, in detriment of everything else
LiquidHaskell (or flux or creusot in Rust) would be an improvement over Haskell or Rust too.
In either case (full blown theorem provers, or theorem provers laid on top of regular programming languages) the models would need to either focus in posttraining on coming up with good invariants to establish, or you need to prompt agents in a very specific way to get desired results (OP complains that sometimes Claude doesn't want to do this stuff, this happens when it's in out of distribution territory)
IMO the biggest issue was always not knowing what correct is in the first place. The vast majority of software we use, the stuff that's riddled with errors, has those errors largely because what it's supposed to do is vague and never, ever deals with edge cases. You can't formally verify your application works correctly under transient network error conditions if you never thought about what your application should do under those conditions.
Perhaps that's the same thing as what you're saying, though: we don't specify these things in detail because it's expensive to spend that much time thinking through it all, when users are largely trained to just accept crashes, glitches, inconsistencies, and the occasional sprinkle of data loss.
This can work for core algorithms for sure, but wondering if this will work for production use cases, production apps come with a lot of edge cases - which are more often than not not logical as well to the point it becomes very hard to document them all in the first place.
The flip side is that we currently put programs into production without understanding how they will behave in those edge cases. If you're lucky, they crash and then restart cleanly. If you're unlucky, they silently corrupt data or violate mission-critical invariants.
Theorem provers have always made extensive use of AI and automation. Formal logic is insanely laborious, and it took Russell a monumental effort to not get very far with his manual verifications in Principia Mathematica, working out all the details by hand. In 1956, Newell came up with the Logic Theorist which was able to prove a decent chunk of the Principia automatically. When Newell informed him, Russell conceded that his manual efforts had turned out to be wasted effort.
I used HOL Light about 15 years ago. The vast majority of the proof details are done by a machine, generally split between term rewriting and then offloads to a generic automated prover for first-order logic. Your job then is formalising the theorem statements and orchestrating the automation, and the latter still takes an enormous amount of labour.
Around this time, we were getting excited about recommender systems for lemma selection. The idea was that you turn every theorem into a bunch of features and train a recommender system against the lemma needed to prove it. Then when you face a new problem, you use its features to recommend which lemmas are likely to be needed. You then throw those lemmas and your conjecture at a bunch of industrial strength automated provers, get the provers to come back with a minimal set of necessary lemmas, and then use your verified automated prover to do the real proof with a tractable set of inputs. It first got implemented in Isabelle/HOL as Sledgehammer, and was a massive improvement to tooling.
Now LLMs are here and the game looks completely different. I went back over some verification I spent a week on about 5 years ago, in a pretty obscure formal verifier called HOL Light, on a problem of my own making. I'd seen how terrible ChatGPT was at propositional logic a few years back, so this is the sort of thing I'd had in my back-pocket as an "impossible benchmark" for LLMs. So as a half-joke, I gave Claude the main theorem I wanted proving, hoping to watch it embarrass itself.
In a minute or two, it has come up with the same proof strategy I had used, proving four lemmas to get to the main theorem. This was impressive, but that's not the laborious part, and I was pretty confident it would die trying to prove just the first lemma. It has its four lemmas enumerated and goes to thinking, while I go off for a cup of tea.
I come back five minutes later and it has scratched off the first goal. I'm pretty shaken, and go off again trying to process that. Come back, the second goal has gone. And then the third. And then the fourth. And then, after just fifteen minutes, it's pulled off a week of my work, done it more efficiently, and near enough one-shotted each proof. That took me a while to process. I realised that if I had this 15 years ago, it would have done 95% of my PhD, which is to say that 15 years ago, I would have done a 20x more ambitious thesis.
I contacted my old supervisor, who said the theorem proving community are all on top of this, including the creator of HOL Light, now at AWS. HOL Light, incidentally, was used on what I still believe is the most ambitious mathematical theorem proving project to date, the verification of the Kepler Conjecture. The lead on that project recruited a team to get the proof through over about 5 years. Today, I wouldn't be surprised if he could have solo'd it in 6 months. The same goes for another extremely ambitious project, the verification of the seL4 microkernel. And for another open problem, I know there are people seriously wanting to get a verification of Fermat's Last Theorem, which sounded delusional 15 years ago, but sounds pretty plausible now.
Exactly where this goes, I am not sure. I suspect we can now start on verification projects that would have been insane to contemplate. But a few things concern me. One is that Lean seems to have all the mind-share. Maybe it deserves it, but there are very major and mature theorem proving technologies such as Coq (now Rocq), Isabelle, HOL Light, ACL2 and Mizar that I believe still win in terms of having the largest verification libraries and the biggest verified projects. These should not be forgotten about, since they will also probably have the most training data, having been going for many decades now.
The second is that verifying our crappy human specifications probably isn't going to fly. As the creator of HOL Light says, representation (how you formalise your problem domain) is still where it all matters, and the LLMs aren't very good at this. And neither are the writers of our current specs. Verifying that a piece of software meets the HTTP spec will be considered an achievement, but it's the wrong goal.
We'll need specifications that are modular and composable, that mesh together so that each piece is sanity checking the others. This is how we build mathematics, and it's how we'd need to build software. Specs need to be short and comprehensible, so that a human can verify them. If your spec is as long and complex as the implementation, it's worthless.
But I believe it is possible to design software from the ground up where the motivation is specification engineering rather than code engineering, and thereby the specifications become the only human-facing understandable part of software. The implementation is just some artifact that an LLM generates that nobody looks at unless they are curious. And I think LLMS today mean that "build the world over again" isn't as mad a thought as it used to be.
Self-insert time.
I spent some time exploring this topic. Here's my thesis: Formal verification was expensive. 20x expensive compared to just developing the software, as the author notes. The cost of finding and developing exploits also was high. That creates an incentive to put software verification aside, since it solves a relatively small problem, at an extremely high cost.
We've seen how Mythos has found more vulnerabilities than the rest of the security industry combined. (you can argue about the quality and what counts as a vulnerability, but not the point) So the cost of finding and developing exploits has dropped dramatically.
On the other hand, formal verification is now much easier, since LLMs can automate the proof. You don't even need to worry about hallucinations, you merely need to trust Lean core. If the LLM is wrong, the proof will get rejected!
The problem of exploits gets bigger, and the solution of formal verification gets cheaper. As a result, the needle is now moving in the direction of "more formal verification".
I, personally, think that it is ridiculous that ~none of the software we use is known to work correctly. It just happens to work correctly, most of the time.
My (ambitious) goal is to have a self-hosting, formally verified compiler, which allows proof transfer from source code down to assembly. I have not achieved that goal yet.
What I have so far:
- one (non-optimized) compiler step which is formally verified
- three simple functions (hex, hex with labels, strtoull) formally verified, against RISC-V assembly, and against a custom IR
https://github.com/m1el/riscv-fv-bootstrap
The project is in quite a bad shape, and I am trying to improve my skills in that direction.
We do have formally verifying C and C++ compilers already. cbmc and esbmc.
Don't we still need to verify that what the LLM proved is actually the stated system?If agents cutting corners (like deleting tests) is a concern, how can we be sure that the verification corresponds to the software and that every _load bearing_ assumption is true? I don't think that simply trusting the Lean core is enough.
Have you heard of https://cakeml.org/? It's a self-hosting formally verified compiler.
Strongly agree with the author here. The future will belong to programming languages that natively embed theorem proofers into their type systems so LLMs can forego a lot of testing by just validating the implementations they write against the specs with formal proofs. Writing formal specs is probably the main skill a programmer in the future will need to get work done.
Verus (https://github.com/verus-lang/verus) is a good start for the rust ecosystem, but it's essentially a standalone language today (with custom syntax and type system).
This is the fantasy that has always driven proof systems research. Nobody is going to run software that has never been tested. Would ride a rollercoaster that had never actually been tested before, only "proven" safe? I wouldn't!
So this stuff is always going to be additive and concerned with edge cases, as almost by definition, stuff that isn't edge cases will be found by comprehensive enough testing procedures.
And yet most software doesn't really need to be correct under edge cases, outside of security and data loss issues. People can tolerate a lot of incorrectness in other areas because it's just annoying, not critical.
Security is a case where formal methods could help, but I don't think LLMs will change the industries lack of interest. If anything it'll reduce it even further. Historically security took place in a fog of war. You don't know your enemies capabilities and may not be able to easily match them. But now LLMs are better at finding security bugs than most (all?) humans and ~everyone has access to them, so, from a liability perspective, all you have to do is point a frontier LLM at your codebase and let it fix as many bugs as it can find. Your enemies don't have access to anything better, so once it's done you can tick the box and say security is good enough. Meaning, nobody will fire you if there are still attacks possible.
So in the end I don't see LLMs changing the adoption formal methods.
Lean has dependant types. Wouldn't something like Haskell or Idris, that are trying to be general purpose dependantly typed languages--wouldn't they be a better start than versus?
Versus appears to just be a formal verification tool. Perhaps I misunderstand?
You want the formal verification built into the language because the tooling can start to get really crazy good. Agda is the dependantly typed language I've used the most (long ago), and the tooling was interactive in a helpful way I've never experienced with other languages.
You don't want a separate language used to verify a base language, because then everyone ends up having to know two languages. Looking at the history of computing though, I wouldn't be surprised if this happens.
The actual programming language and the verification language can be the same language though, if we want.
Lean4 is a general purpose programming language.
Haskell does not have dependent types.
in case anyone's interested i have a vibe coded fork of verus that replaces the verus-the-language side of verus with plain old Lean 4. It's still two languages, but now at least the second language is as mainstream as it gets in the field and has good automation. i haven't finished wiring up the Lean 4 infoview and vs code extensions and LLM skills into it yet, which makes it not as easy to write yet as lean 4 with the IDE bells and whistles.
I'm also playing around with using the lean's compile-to-C tooling to instead compile to rust instead and it's getting more of my focus than the lean-via-verus route right now.
if people are interested, ping me and i can put them up on gh.
Please post this!
As a meta-comment on the topic, something I have noticed is that there still exists confusion what it means to use theorem provers for projects -- the other day I read a tweet from Paradigm, a crypto-VC now seemingly AI-pilled. Some LP of theirs had made a Lean 4 formalization of the Ethereum's virtual machine. The tweet said this would have cost like $150k in API tokens ("would have", as in, I guess they get theirs for free), and took a week of inference time for an LLM to produce. I somehow got distracted to actually take a look at the code, which I found rather light on theorems. Nor did the project make use of Batteries or Mathlib which are arguably the one of the strongest motivation for me personally to use Lean4. That is, I generally rather rely on someone else getting the category theory and algebraic structures right, which then leaves me the proof obligation to show the correspondence with whatever toy I'm working on. Here I'm fine to use LLMs for proof search, very similar to how would I use a SMT solver. But what I have found is that the language models have to be really coerced into using these libraries, because otherwise the models much rather overfit and overclaim a solution with a 3 minute inference task rather than attempt to fulfill the proof obligations over 3 hours. And I feel nauseated when I need to convince the LLM (I use Claude) that filling the proof obligation is for "academic exercise" or because I'm coerced into doing so, because otherwise it will come up with reasons of its own why it does not want to do it. Now, this happens under the mental model in which I'm interested in finding equivalences with prior work. Many LLM generated Lean code reads more as if someone was interested whether X can be turned into a Lean 4 program, which is mostly yes, and that in general is a positive thing. But, if you are not interested in refinement types and theorems, why not just choose Haskell? The point is, I strongly sense that unless you have good questions to ask, then that's very evident in these languages. And, this is something the LLM won't help you -- if you don't impose a proof obligation for it, it certainly will not try to go the extra mile to conjure one for you.
Crypto guys are in the business of grifting magic beans to rubes. They are mostly interested in the aesthetics of these technologies. Using an academic programming language gives these scientific magical vibes and such that make people believe these guys are high tech and know their shit and so on. This is the reason many crypto projects do use Haskell too, it's known as an academic research language that is difficult to use and "if it compiles it works." They come at it from the perspective of "What would sound the best when I try to sell the space money of the future?"
Blockchain protocols are also faced with the most active, persistent, and well-funded state sponsored attackers of any industry, so naturally formal verification should be of great interest.
I have already written it, and I will write it again: dependent types and total functions do not scale. Maintenance is terrible.
Suppose that you have managed to write a non-trivial piece of software with dependent types encoding all sorts of properties everywhere. The actual computation and proof are intermingled. Think of the author's innocent bound-check proof in the zstd decoder, but across the whole program, with more elaborate properties and longer proofs.
Suddenly, you realize that you need to prove a new property of your program. Can you keep your existing work and build on top of it? In general, no, you have to refine every dependent type everywhere by adding a new conjunct expressing a new invariant, and adapt every proof, as the new property is threaded in the existing program.
That is because dependent types (and other staples of naive approaches to proving program properties, like a unique invariant per loop) structure the program along the wrong dimension: they encourage grouping everything that concerns a value ("put this value in a dependent type that encodes everything known about it") or a program point ("write the precondition for this function as a big conjunction mixing all the concerns"), where it works much better, for long-term maintenance, to structure the development along concerns: computational parts of the program, basic functional properties and absence of UB, termination, other functional properties, security, etc., where each layer builds on top of the previous ones without requiring them to change.
That is not to say that dependent types do not have their use. Where they shine is a module boundaries. Consider a library that exposes an opaque type and operations on it. Users of the library can only build and modify values of that type through the library's API. This type should be a dependent type. If it needs to be refined at any point to encode a new invariant, this will have no impact on the library's users, since all they do is pass around values without interpreting them. However, inside the library, I would recommend unpacking/repacking the dependent type at the library's entry points and handling the concerns separately.
I'm convinced Math is the canary for what's going to happen to knowledge work. Coding was ahead in harnessing early LLM capability, but Math, given it's pure form, has aleady racing ahead.
The dimensions to notice are: Research speedup, practitioner expertize, labour dynamics, junior entrants, world impact.
In that sense it's a canary, whatever happens to Math, will in order flow to other sciences in-order of purity: Computer science, physics, chem, bio, social sciences.
Math is the opposite of a canary because it's possible to generate infinite amounts of synthetic training data for it, unlike almost every other kind of knowledge work where correctness depends on external input.
This is really cool stuff, and I agree the future is likely to look more like this. I was surprised by the last two paragraphs ("Aside: verified assembly") -- my understanding was that this future is basically already here. I believe agl's colleagues at Google have already deployed some auto-mutated verified assembly versions of some crypto routines, based on the Fiat Crypto + CryptOpt work (https://arxiv.org/pdf/2211.10665), both involving Andres Erbsen who I think is currently at Google before starting a professorship soon. I dunno if this work would count as "cheap" (it looks like Fig. 10 unfolds over the course of a day) but spending a day auto-exploring many verified-correct machine code implementations of the same routine to find the fastest one (which you then ship forever) doesn't seem impractically expensive either.
I'm very bullish on proof automation as well. I'm currently researching AI for algorithm design and using automated theorem provers to get formal guarantees for generated algorithms.
To make a shameless plug, I'm working on a Python package called OpenATP [1] to make it easy to benchmark different models/harnesses for automated theorem proving. It supports running agents in Docker containers or Modal out of the box. If you try it out, I'd love to get your feedback!
I recently wrote about the surprisingly good performance I saw from Grok [2]. On more challenging proofs, Grok doesn't keep up with Opus/Fable and GPT 5.6. I was recently blown away by GPT 5.6 Sol. It's persistence in closing out proofs is unparalleled from what I've seen so far. OpenATP also supports Kimi and Leanstral [3], among others.
[1] https://github.com/henryrobbins/open-atp
[2] https://news.ycombinator.com/item?id=49010310
[3] https://news.ycombinator.com/item?id=48780801
> We now have LLMs which, combined with proof irrelevance, promise to be an extremely capable form of proof automation. With sufficient amounts of automation perhaps you don't need to worry about proof engineering nearly so much. You still need to avoid blowing up the type checker but, in my limited tests, LLMs can avoid that. Potentially, LLMs suddenly make dependent-type systems dramatically more practical.
When I've used interactive proof systems like Roq, I'd often kinda code myself into a hole by cutting along the wrong axes and not specifying my problem in a way that's easy to prove. After all, this is just like in math: you really want to cut at a problem the right way to get to the easy proof.
I think people are discounting how important that decision making is. It's not just about whether an LLM can churn through specific proof strategies on a problem, but also about how to pose the problem etc.
I'm not saying LLMs can't help, but I think it's less that "proof engineering is not needed" and more that "when these tools are used in the right way, proof engineering is easier". Because at the end of the day these tools work well when they have the right kind of foundations in the first place
> AWS made LNSym: a semantics and simulator for AArch64. That's cool. Perhaps we could use it to show equivalence between an optimised assembly implementation of some functions, and their Lean counterparts, and then use the assembly code at run-time? Then we could let LLMs rip at optimisation and they couldn't introduce any functional bugs. Verified assembly is well-trodden in crypto implementations, but perhaps now it could be cheap?
Trying to one-shot compcert might be hard! Thinking about it and planning it out might make it easier though...
Yeah computer proof writing involves choosing good abstractions at every turn. LLMs aren't great at that yet
That’s true, they’re not great at it. Just better than 99.99% of humans.
Source?
The fact that 99.99% of humans have never used a formal theorem prover?
Worse than 0.01% of humans means that there are 8,000,000 people better than it. I know that's being pedantic I understand what you're saying.
But every time I use Codex unless I specifically give it the abstractions it writes code that is way too specific.
How do we know if they’re better or not if they haven’t used one?
It would be even better with a few real-world implementation examples.
Cool. Now we can write bugs in our theorem descriptions instead of source code.
Seriously, please review Curry Howard Isomorphism if you’re getting pulled down this rabbit hole.
Programs are proofs. Proofs are programs.
So if you can formally describe the correct output for every input, you can have an LLM loop automatically fill in the gaps of how to get there. Congrats, that sounds at least as hard as writing the correct program in most cases.
Don’t get me wrong, I do think there are useful tools combining formal methods and llms. Let’s just not get carried away.
It is true that finding the correct specification is a formidable task; knowing what correctness even means is arguably most of the difficulty of programming. However! "Moving bugs up from programs to types" isn't how this shakes out in practice, at all. Another commenter already noted that it's often much easier to communicate your intent through specifications, because you can essentially always say what a computation should do much more simply than you can say exactly how to do it.
I think it's also important not to miss the forest for the trees: even relatively simple specifications like "the compress and decompress functions must be inverses for all inputs" rules out vast classes of bugs in a compression library. This is not a complete specification; for instance, it does not speak about how the decompressor behaves on malicious input. But in my experience, even partial specifications carry the promise of hitting warp speed with LLMs in a way that I haven't seen anywhere else. After a certain level of specification, you have decent guarantees of being able to whole-heartedly forget about the implementation details of the synthesized program. And you get a better-built, more robust program out of it at the end!
The comment at the end of the article about having LLMs directly generate assembly against specifications and letting them rip with finding custom optimizations is the sort of crazy stuff this enables. I really think we're only seeing the tip of the iceberg here. People keep asking what we can do with LLMs that we couldn't before; this is the answer.
> Congrats, that sounds at least as hard as writing the correct program in most cases.
That's not remotely true. Or, more formally speaking since we're in a thread about proof assistants, it's not remotely true, up to extensional equality, plus some choices about which axioms you use.
I can write a formal description of what it means to have property in a way that does have computational content that is equivalent to an algorithm[0], but often the clearest way to express the property is equivalent to an algorithm that literally brute forces the problem, like sorting a thing by checking every permutation until you find one that's sorted.
The magic is that you can write a spec that's clear, then have the LLM write the code and prove the spec, so you know that given the right inputs/state, it will return the right outputs/state. Then the gap is performance-like characteristics, which is a pretty great starting point and a lot easier to be just empirical about than correctness.
[0] in Lean you can also use classical logic or add your own axioms, where it's not even comutational.
> That's not remotely true. Or, more formally speaking since we're in a thread about proof assistants, it's not remotely true, up to extensional equality, plus some choices about which axioms you use.
Really unnecessary levels of snark here.
> often the clearest way to express the property is equivalent to an algorithm that literally brute forces the problem, like sorting a thing by checking every permutation until you find one that's sorted
Have you ever heard of prolog? I’m sure a prolog program can express whatever property you’re attempting to write just as tersely (if that’s your metric for hard). Almost copy/pastable to and from a theorem proofer.
Or are you saying that the program has to be the efficient implementation? Because that’s a different ball game. I’m not even going to get into how you could provably transform brute force propositional logic into efficient algorithms. (At that point we’ll have finally created the fabled “sufficiently smart compiler” and probably solved p=np).
A huge majority of software is simple business rules + CRUD that is trivially verifiable. The entire problem is showing that an efficient/reliable program actually implements those rules.
> The magic is that you can write a spec that's clear
Maybe you can. But I did spend a grad class with rocq (coq at the time) and a decade working with “systems engineers” and am not convinced that this is a realistic expectation.
> I’m not even going to get into how you could provably transform brute force propositional logic into efficient algorithms.
> The entire problem is showing that an efficient/reliable program actually implements those rules.
Reliability is a standard matter of correctness and captured (partly) by specifications. Efficiency tends to be easy to empirically test, but it is also possible to capture at the specification level [1]. Mind that specifications need not be all-consuming.
> But I did spend a grad class with rocq (coq at the time) and a decade working with “systems engineers” and am not convinced that this is a realistic expectation.
Agree! But this stuff just got massively more accessible, and the tooling around it is growing quickly. I think we'll end up growing specification systems specific to various domains which will be palatable to those "systems engineers", but I err on the side of optimism here. There's definitely a lot left to do for practicality.
[1] See the work of https://cs.nyu.edu/~shw8119 for the case of provably-efficient parallelism and garbage collectors
> Really unnecessary levels of snark here. [..] Have you ever heard of prolog?
Why do you look at the speck of sawdust in your brother's eye and pay no attention to the plank in your own?
> I’m sure a prolog program can express whatever property you’re attempting to write just as tersely
No, you won't be able to express the most basic properties in Prolog at all, let alone as tersely as in a proper specification language.
E.g. if you have a language interpreter and a bytecode interpreter, pretty much every specification language will let you express the correctness of a compiler `compile(x)` as
``` for all scripts x and inputs i,
bytecodeInterpreter(compile(x),i) == languageInterpreter(x,i) ```
Good luck expressing this in Prolog, let alone equally tersely.
Huh, I wasn't actually going for snark. I was trying to preemptively be over-specific about what I meant. As in, you said "... writing the correct program..." and I meant that I'm talking in terms where functions are equal or distinct only based on extensional equality and with some flexibility on computational interpretation and axioms of the logic.
Like, if you meant "the correct program" in a sense where two pure total functions can be different despite both having the same outputs on the same inputs, then that's not what I'm responding to.
And yeah I know prolog, but proof assistants and logic programming are totally different beasts. Definitely not copy/pastable to/from lean, at least as I've seen and used each.
> Or are you saying that the program has to be the efficient implementation? Because that’s a different ball game. I’m not even going to get into how you could provably transform brute force propositional logic into efficient algorithms. (At that point we’ll have finally created the fabled “sufficiently smart compiler” and probably solved p=np).
I think you're totally misunderstanding. What I'm saying is that in something like Lean (just because I know it best) I can say, "this function takes inputs satisfying Prop1 and returns outputs satisfying Prop2," in ways where I write some brute-force equivalent formalization of Prop1 and Prop2 in the most straightforward way, and then go on to prove that they are true of my program that is not the brute force implementation. Like the wacky famous magical inverse square root implementation from Quake III. You could write a spec that "output = 1/sqrt(input) up to float properties" and the implementation in the famous brainfuckery way. To your comment about how the spec is as hard as the implementation, "output = 1/sqrt(input)" is way easier than the weird efficient implementation, and that class of distinction is super common.
And as you said, the annoying part is showing that the efficient implementation satisfies the spec, but what's magic today is that we have great tools and LLMs can and do fill in the blanks. In practice, I write the spec by hand for the stuff I care about and then prompt the rest and know that my spec is what the LLM implemented.
> > The magic is that you can write a spec that's clear
> Maybe you can. But I did spend a grad class with rocq (coq at the time) and a decade working with “systems engineers” and am not convinced that this is a realistic expectation.
I tried rocq back when it was coq too, and now do most of my work in lean and rust and python with totally normal folks and I'm convinced that the tooling and languages are finally just about Good Enough. If you have any interest in the field, which it sounds like you do, and if you haven't checked out the ecosystem in the last couple years, I'd recommend you check it out again.
Pretty much this. I designed my own research harness and infra for theorem proving and conjecturing; however, you still need to review formulations!
I agree with the core thesis that LLMs + theorem provers might make formal methods cheap enough to be practical in software development.
The biggest issue was always cost. But there's still an alignment problem. Without human supervision, things might drift away from the original specification and intent.
From my own experience, what works best is some kind of Hoare/separation logic (contracts), as these are quite easy to follow and decompose.
Even something as simple as a minimal Haskell subset, plus a bit of LiquidHaskell, can get you really far if you are pragmatic.
Types, preconditions and theorems are still a guardrail
Agents perform better when programming in Rust than <insert dynamic language here> (at least in my experience), exactly because types matter. Those theorem provers just have a poweful and pedantic type system that insists in being always correct, in detriment of everything else
LiquidHaskell (or flux or creusot in Rust) would be an improvement over Haskell or Rust too.
In either case (full blown theorem provers, or theorem provers laid on top of regular programming languages) the models would need to either focus in posttraining on coming up with good invariants to establish, or you need to prompt agents in a very specific way to get desired results (OP complains that sometimes Claude doesn't want to do this stuff, this happens when it's in out of distribution territory)
IMO the biggest issue was always not knowing what correct is in the first place. The vast majority of software we use, the stuff that's riddled with errors, has those errors largely because what it's supposed to do is vague and never, ever deals with edge cases. You can't formally verify your application works correctly under transient network error conditions if you never thought about what your application should do under those conditions.
Perhaps that's the same thing as what you're saying, though: we don't specify these things in detail because it's expensive to spend that much time thinking through it all, when users are largely trained to just accept crashes, glitches, inconsistencies, and the occasional sprinkle of data loss.
When I did it the biggest problem is writing the spec. It was longer and more complicated than the code itself.
If you have bugs in the spec you have bugs in the code.
OK so this article is sort of about formal proof automation, but it seems more practically about Zstandard, which I very much enjoyed reading.
This can work for core algorithms for sure, but wondering if this will work for production use cases, production apps come with a lot of edge cases - which are more often than not not logical as well to the point it becomes very hard to document them all in the first place.
The flip side is that we currently put programs into production without understanding how they will behave in those edge cases. If you're lucky, they crash and then restart cleanly. If you're unlucky, they silently corrupt data or violate mission-critical invariants.
I have had production edge cases anticipated by AI before that I hadn't accounted for.
as the proof of verification decreases, the value of credentials that act as shortcut proofs of human competence will decrease, too.
Theorem provers have always made extensive use of AI and automation. Formal logic is insanely laborious, and it took Russell a monumental effort to not get very far with his manual verifications in Principia Mathematica, working out all the details by hand. In 1956, Newell came up with the Logic Theorist which was able to prove a decent chunk of the Principia automatically. When Newell informed him, Russell conceded that his manual efforts had turned out to be wasted effort.
I used HOL Light about 15 years ago. The vast majority of the proof details are done by a machine, generally split between term rewriting and then offloads to a generic automated prover for first-order logic. Your job then is formalising the theorem statements and orchestrating the automation, and the latter still takes an enormous amount of labour.
Around this time, we were getting excited about recommender systems for lemma selection. The idea was that you turn every theorem into a bunch of features and train a recommender system against the lemma needed to prove it. Then when you face a new problem, you use its features to recommend which lemmas are likely to be needed. You then throw those lemmas and your conjecture at a bunch of industrial strength automated provers, get the provers to come back with a minimal set of necessary lemmas, and then use your verified automated prover to do the real proof with a tractable set of inputs. It first got implemented in Isabelle/HOL as Sledgehammer, and was a massive improvement to tooling.
Now LLMs are here and the game looks completely different. I went back over some verification I spent a week on about 5 years ago, in a pretty obscure formal verifier called HOL Light, on a problem of my own making. I'd seen how terrible ChatGPT was at propositional logic a few years back, so this is the sort of thing I'd had in my back-pocket as an "impossible benchmark" for LLMs. So as a half-joke, I gave Claude the main theorem I wanted proving, hoping to watch it embarrass itself.
In a minute or two, it has come up with the same proof strategy I had used, proving four lemmas to get to the main theorem. This was impressive, but that's not the laborious part, and I was pretty confident it would die trying to prove just the first lemma. It has its four lemmas enumerated and goes to thinking, while I go off for a cup of tea.
I come back five minutes later and it has scratched off the first goal. I'm pretty shaken, and go off again trying to process that. Come back, the second goal has gone. And then the third. And then the fourth. And then, after just fifteen minutes, it's pulled off a week of my work, done it more efficiently, and near enough one-shotted each proof. That took me a while to process. I realised that if I had this 15 years ago, it would have done 95% of my PhD, which is to say that 15 years ago, I would have done a 20x more ambitious thesis.
I contacted my old supervisor, who said the theorem proving community are all on top of this, including the creator of HOL Light, now at AWS. HOL Light, incidentally, was used on what I still believe is the most ambitious mathematical theorem proving project to date, the verification of the Kepler Conjecture. The lead on that project recruited a team to get the proof through over about 5 years. Today, I wouldn't be surprised if he could have solo'd it in 6 months. The same goes for another extremely ambitious project, the verification of the seL4 microkernel. And for another open problem, I know there are people seriously wanting to get a verification of Fermat's Last Theorem, which sounded delusional 15 years ago, but sounds pretty plausible now.
Exactly where this goes, I am not sure. I suspect we can now start on verification projects that would have been insane to contemplate. But a few things concern me. One is that Lean seems to have all the mind-share. Maybe it deserves it, but there are very major and mature theorem proving technologies such as Coq (now Rocq), Isabelle, HOL Light, ACL2 and Mizar that I believe still win in terms of having the largest verification libraries and the biggest verified projects. These should not be forgotten about, since they will also probably have the most training data, having been going for many decades now.
The second is that verifying our crappy human specifications probably isn't going to fly. As the creator of HOL Light says, representation (how you formalise your problem domain) is still where it all matters, and the LLMs aren't very good at this. And neither are the writers of our current specs. Verifying that a piece of software meets the HTTP spec will be considered an achievement, but it's the wrong goal.
We'll need specifications that are modular and composable, that mesh together so that each piece is sanity checking the others. This is how we build mathematics, and it's how we'd need to build software. Specs need to be short and comprehensible, so that a human can verify them. If your spec is as long and complex as the implementation, it's worthless.
But I believe it is possible to design software from the ground up where the motivation is specification engineering rather than code engineering, and thereby the specifications become the only human-facing understandable part of software. The implementation is just some artifact that an LLM generates that nobody looks at unless they are curious. And I think LLMS today mean that "build the world over again" isn't as mad a thought as it used to be.
All this proof automation.
But still no P=NP.
For P ≠ 0, we get N = 1. ■