Formalization is in principle just a translation process and should be a much simpler problem than the actual IMO problem. Besides, they also trained a Gemini model which formalizes natural language problems, and this is how they generated training data for AlphaProof. I would therefore expect that they could have also formalized the IMO problems with that model and just did it manually because the point is not to demonstrate formalizing but instead proof capabilities.

Yet the facts at hand are the opposite of what you say. Reliable formalizer was the more difficult problem than solving formalized IMO problems, because they have not produced one.

That does not necessarily follow from the facts at hand. For example they may have prioritized work on the proof solver itself as they may feel that that is the more important result. Alternatively if their goal is to build a proof solver then building the formalizer would be useless if they could not build the actual proof solver.

A proof solver existed. They were improving the proof solver explicitly by making the formalizer a part of the training. Formalizer reliability is the key novelty. It turns out it was only reliable enough for training. So unless they made the problem statement at the outset that "we'll only make the formalizer strong enough to train but not use", I disagree with that assessment.

That's a good point. The formalizer was used to created the training data for the proof solver, so they likely worked on it more than if they just used it as a preprocessing step during inference. It is still possible that they worked on the formalizer until they got good results from it enough to create good training data, and then began training as soon as possible and did not spend too much time trying to improve the formalizer. Depending on how long the training was expected to take, perhaps that is a reasonable assumption. Although I think I agree more with your view now.

> Formalization is in principle just a translation process and should be a much simpler problem than the actual IMO problem

maybe not, because you need to connect many complicated topics/terms/definitions together, and you don't have a way to reliably verify if formalized statement is correct.

They built automatic formalization network in this case, but didn't trust it and formalized it manually.

If they could have solved it, they would have. But I agree that language models will be able to do it.

Yes and it is difficult for me to believe that there is not useful human analysis and understanding involved in this translation that the AI is helpless without. But that I suppose is a problem that could be tackled with a different model...

Even so, having a human formalize the problems and an AI to find machine checkable proofs could be very useful for mathematicians.

It is vastly easier to do the formalization than to actually solve the problem. Any undergraduate with some lean familiarity could do it in minutes.

Disagree! Some problems are much harder than others. If you don't believe me please go formalize P5 in this year imo.

I formalized it last night, to a level that an IMO trainer agreed was adequate. Took maybe 15 minutes.

Find n such that p(n) and not p(n-1).

  p(n):
    exists(f: state -> move) such that solves(f, n)
  
  state: solved | illegal | (k, is_first_move in {T,F}, px in (1,2023), py in (1,2024+1), mapping from x,y to {T,F, ?})
  
  initial_state(n): (n, T, 1, 1, {(x,y) -> ?})
  
  move: U|D|L|R|(x in (1,2023))
  
  power: ((a -> a), integer) -> (a->a)
  power(h, 0)(x) = h(x)
  power(h, k)(x) = h(power(k-1))(x)
  
  solves(f, n): exists l such that for every board, power(make_move(board, f), l)(initial_state(n)) = solved
  
  board: permutation of (1, 2, 3, ..., 2022+1)
  
  make_move: (board, (state -> move)) -> (state -> state)
  make_move(board, f)(solved): solved
  make_move(board, f)(illegal): illegal
  make_move(board, f)(s = (k, is_first_move, px, py, m))
     if is_first_move = T:
        if k = 0:
            illegal
        else if f(s) is a number:
            (k, F, f(s), 1, m)
        else:
            illegal
    else:
        if f(s) is a number:
            illegal
        else if py = 2025:
            solved
        else if board(px) = py and py != 1:
            (k - 1, T, px, py, m + {((px, py), T)})
        else:
            dy = {U:-1,D:1,L:0,R:0}(f(s))
            dx = {U:0,D:0,L:-1,R:1}(f(s))
            px' = px+dx
            py' = py+dy
            if px' < 1 or px' > 2023 or py' < 1 or py' > 2025:
                illegal
            (k, F, px', py', m + {((px, py), F)})

I find it incredibly impressive you did this in 15 minutes! You should really help out in formalizing math (completely serious).

Personally in the past I tried a few times to formalize some statements and sometimes I found that the mathlib libraries were pretty lacking in these more open-ended problems (I wanted to reason about lists and stuff). But it seems that I am just very bad at formalization lol.

Formalization is mechanical work, lets leave it for computers to do :)

I completely agree

Yeah, I was just referring to the problems that it actually did.

But formalization is the easy part for humans. I'm sure every mathematician would be be happy if the only thing required to prove a result was to formalize it in Lean and feed it to the AI to find the proof.

Not sure every mathematician would be happy to do this... it sounds much less pleasant than thinking. It's like saying mathematicians would rather be programmers lol. It's a significant difficult problem which i believe should be left completely to AI. Human formalization should become dead

IIUC, a Gemini-based system could translate the natural language questions into Lean, but in the blog post they don’t really commit to whether this was done just to generate training data or was used in the competition.

Formalizations for the competition were done by hand.

Does this formalization process include giving it the answer it should try to prove?

Nope, per Oliver Nash who worked on the thing: https://news.ycombinator.com/item?id=41070372