I feel LLMs are indeed getting better at writing models. But, in my experience, they struggle to come up with correct safety and liveness properties unless you closely work with them. And of these two, they struggle the most with correct liveness properties.

Also for some problems I observe that models produced by LLMs often cause state space explosion. For simpler models they can fix this when you guide them though.

I’m sure LLMs will get even better.

That said, I take slightly different approach. Lamport said “If you're thinking without writing, you only think you're thinking.” So taking that advice I always try to write the first draft with hand and once I have the final shape in place I then turn to an LLM for further exploration and experimentation if I have to.

I remember NVIDA sponsored a TLA+ challenge last year https://foundation.tlapl.us/challenge/index.html

I don't use tla+ to model real-world systems anymore, Claude is able to model systems in Lean 4 and the binary executable can handle real input or I can directly generate c / rust on proofs with numeric types that have ring structure (integers, rationals, bits).

https://github.com/lambdaclass/truth_research_zk

This post reads like an accidental advertisement for approaches like Verus [1], which couple the implementation and verification so you can't end up with a model that diverges from the actual implementation. I'm personally much more optimistic about the verus approach, but I freely admit that's my builder bias speaking.

[1] https://github.com/verus-lang/verus

Claude has certainly been getting better with TLA+. It's not perfect yet but for laughs I got it to model the rules of Monopoly last night [1]. I haven't done any exhaustive checking on it yet, but it certainly looks passable.

It is pretty impressive at how good it's gotten at this, in a relatively short amount of time no less. I still usually write my specs by hand, but who knows how much longer I'll be doing that.

[1] https://pdfhost.io/v/KU2j37YKrP_Monopoly

Just a question to people who may know better than me about this.

I thought the whole point of trying to write out TLA+ is so that you get a better idea of what you want and put it into formal language?

I get that an LLM can assist/help with expressing what we want in formal language a bit, but if one automates all this there is no human intent/design anymore.

If the LLM generates both the design (TLA+) and writes an arbitrary program that satisfies said design -- what exactly have we proved?

What assurance do humans get since human doesn't know or cannot specify what they want.

I've been building a TLA style Temporal Logic library for Verus (using LLMs). My experience so far is that LLMs are surprisingly useful at generating the mechanical proof scaffolding (when they're not occasionally trying to cheat with `assume(false)` statements), but they are not a substitute for knowing what property you actually want.

>... we asked Claude to write a TLA+ specification (spec) for Etcd’s Raft implementation. It passed syntax checks, ran through the TLC model checker, and at first glance looked like a polished formal model.

This is a mistaken use of TLA+!

Leslie Lamport insists that he invented it to be a way of creating "blueprints" for systems.

- You are supposed to go from TLA+ Spec to System (codebase or hardware).

- Not codebase to TLA+ like the author has done.

Otherwise, you may simply model an existing bug properly and the pass all the checks based of its implementation.

He (Leslie Lamport) insists that the value AI can provide is in compiling TLA+ specs to a code base.

Sorry, must be a very naive question, but what if you give LLM just a source code (maybe even obfuscate the names like Raft and Etcd) and ask it to create a TLA+ spec of that?

Hopefully some people find this interesting too:

TLAiBench[0]: A dataset and benchmark suite for evaluating Large Language Models (LLMs) on TLA+ formal specification tasks, featuring logic puzzles and real-world scenarios.

[0]: https://github.com/tlaplus/TLAiBench

is the training data for these testcases in benchmark not already there ? how do llms perform in novel complex systems spec design ?

hmm

It looks quite complicated and I have no idea what it is doing. Obviously, since I don't know about TLA+. But what about someone who knows TLA+? It still seems hard to make sure it is valid. And it's just for a relatively simple game.

What's the advantage of provable correctness if it's apparently not easy to prove even for people who understand TLA+? I'm not trying to be a party pooper, just curious.

Isn't logical incorrectness less of a problem in software than failures of imagination or conscientiousness in modeling the domain?

> I haven't done any exhaustive checking on it yet, but it certainly looks passable.

isn't that exactly the kind of fails LLMs do the most? first-glance-passable nonsense?