Professor @EcoleDesPonts IP Paris | AI for math | Control & PDEs | Member @ CIRCLES consortium | Scholar @ Korea IAS | Chair DESCARTES @ Hi! Paris
- Tweets 357
- Following 235
- Followers 1,051
- Likes 1,169
Pinned Tweet
For the last few months, we’ve been building a proving agent with the Numina Project. Recently, it found a proof of a conjecture in control theory that I had attempted to prove while a PhD student, and that I had not managed to solve. It had remained open since then (1/15)
3
36
162
24,471
For the aspiring PhD students among you :) I cannot emphasise enough what exceptional conditions these are for a PhD!
2
2
30
7,727
Somehow I am part of Forbes 30 under 30! Thanks @ForbesUnder30 for welcoming me in the community. I am glad that mathematics is honored this year through my works. #ForbesUnder30
@EcoledesPonts @Sorbonne_Univ_ @Rutgers_Camden @Cambridge_Uni forbes.com/30-under-30/2021/…
1
6
31
I'm very happy to have been appointed full professor by @EcoledesPonts! Thank you @EcoledesPonts for your confidence in a young faculty like me, it's a great honor!
1
17
1,260
Thrilled and honored to receive the 2019 European PhD Award on Systems & Control ! Thank you so much to the EECI and to my advisors, Jean-Michel Coron and Sebastien Boyaval. Congrats also to @YangZhe46859983. #PhD @Sorbonne_Univ_
1
2
16
We should probably create an equivalent of the Nobel Prize in computer science, something that would have the name of a famous historic figure of the early ages of computer science. Wait
1
1
11
1,252
Can language models be trained to find solutions to as yet unsolved mathematical problems? The answer is yes! Check our new article 🙂 1/n
Transformers can be trained to solve a 132-years old open problem: discovering global Lyapunov functions. New paper on Arxiv (accepted in NeurIPS 2024), with @albe_alfa and @Amaury_Hayat
arxiv.org/abs/2410.08304 1/8
1
1
11
694
Yann LeCun spreads the word about our paper and repo! Link to the repo (source code, datasets, and pre-trained model available):
github.com/facebookresearch/…
10
If true, it would not be the first time that OpenAI communicates in a way that implicitly suggests an overclaim. This would be unfortunate because 1) it diminishes people's trust in announcements about AI for science, and 2) GPT-5 is really impressive and doesn't need this.
At first, I thought GPT-5 had cracked those math problems on its own.
Turns out (as Demis pointed out) GPT-5 just looked up the answers via web search.
We really need better peer review for these “AI discovers science/math” claims.
10
2,292
I've been lucky to work 2,5 years on this paper with amazing researchers from FAIR (+Gab Ebner) :) From an era without ChatGPT and (most) LLMs, this paper was still state of the art one year after ChatGPT and the first Llama
1
9
485
My talk at @Institut_IHES ! How can Machine Learning Help Mathematicians ? I'm grateful for the chance to talk in this legendary place ! piped.video/watch?v=dlgolR2M…
Full workshop featuring @f_charton @KempeLab @syhw @andrewdudzik and Yiannis Vlassopoulos carmin.tv/en/collections/mat…
1
1
8
464
Thank you, @Polytechnique for this interview, I am deeply honored ! The selected quote is the occasion to give credit to the amazing ML experts @GuillaumeLample, @f_charton and other brilliant researchers/engineers from @facebookai I am thrilled to work with. 1/3
Congratulations to @Amaury_Hayat (X2011) for being featured on the 2021 @Forbes #30Under30 #Europe list (Science & Healthcare category)! 👏🎉
More information in this exclusive interview: bit.ly/3vJDIG3
1
2
8
Our work on language models to find Lyapunov functions is featured in @newscientist! A nice paper by @stokel
An AI system has helped tackle a longstanding tough mathematical problem involving tools called Lyapunov functions. My latest for @newscientist newscientist.com/article/245…
1
7
180
A pleasure and an honor!
🎙️Featuring speakers: Hannaneh Hajishirzi, Swarat Chaudhuri, Jeremy Avigad, @wellecks, @Amaury_Hayat, @mrinmayasachan, Moa Johansson, Leonardo de Moura, @huajian_xin
8
298
Très fier de @NathanLichtle et de son podcast très intéressant ! ingenius.ecoledesponts.fr/ar…
2
3
202
Predicting solutions to advanced math problems with deep learning models! The latest version of our ICLR 2021 paper with @f_charton and @GuillaumeLample. arxiv.org/abs/2006.06462
Deep language models can predict mathematical properties of differential systems. The final version of Learning advanced mathematical computations from examples, our ICLR 2021 paper, with @Amaury_Hayat and @GuillaumeLample, is on Arxiv arxiv.org/abs/2006.06462
1
7
Life achievement (that I'd never thought of)! a youtuber made a popularization video anime style about our paper! piped.video/watch?v=s3pxEgHk… I don't speak Japanese but I'm already a huge fan!
1
7
188
So cool! I've been looking forward to it! It's also very cool and beneficial to the research community to have a document that explains everything in detail. Hats off to Mistral!
Very excited to release our first reasoning model, Magistral. We released the weights of Magistral Small alongside a paper that presents our approach, online RL infrastructure, and findings.
7
280
The proof is expressed in a (human made) formal language (Lean for instance). So it would be readable for a mathematician familiar with Lean and definitely translatable to any human mathematician :)
6
At last open sourced :)
github.com/facebookresearch/… by @GuillaumeLample @tlacroix6 @MaLachaux @AurR0d @Amaury_Hayat @LavrilThibaut Gabriel Ebner @javier_m and others!
6
498
A good opportunity to (re)-read this great paper: arxiv.org/abs/2404.19737
6
199
Congratulations Jeremy! It's very good news to see this new institute, much needed in tomorrow's world.
“The institute will focus on the mathematical components of these tasks and use the technologies to support mathematical reasoning and computation in all its applications,” said Jeremy Avigad, director of ICARM.
1
5
359
Merci @_Mobility_TV pour cette occasion de discuter de la décarbonation des autoroutes !
Innovations et #décarbonation des autoroutes, comment les autoroutes se décarbonent ?
@PatriceGeoffron / @Paris_Dauphine, @Amaury_Hayat / @EcoledesPonts, @louis_dupasqu / @VINCIAutoroutes
piped.video/pvFSEwZE8BE
1
4
217
Nice results and glad to see progress on ATP in Lean ! You might still want to have a look at HTPS (41% on MiniF2F with search, 35% supervised only, May 2022) arxiv.org/abs/2205.11491
4
294
Well, well well... while I'm happy to see high school students being enthusiastic about research, there used to be a time when you could be relatively carefree when young and it was fine.
The NeurIPS high school track has made its way to Chinese social media.
The collaborator who sent me this said there is at least one example of a professor asking their PhD student to help write a paper for one of their kids so that it can help their college admissions abroad.
5
288
Code and datasets are available ! It's quite modular, so you can easily change it to study another open math problem 🙂 Train time is quite affordable. Feel free if you have any questions
The code for our paper: Global Lyapunov functions: a long-standing open problem in mathematics, with symbolic transformers, with @albe_alfa and @Amaury_Hayat is available at github.com/facebookresearch/…
We will be in NeurIPS: come see us at the poster session next Thursday at 5PM
1
5
376
Really cool paper by @FabianGloeckle @byoubii, @b_roziere, David Lopez-Paz and @syhw ! I've heard about it several times over the last few months, and I'm glad to see it out :)
Meta presents Better & Faster Large Language Models via Multi-token Prediction
- training language models to predict multiple future tokens at once results in higher sample efficiency
- up to 3x faster at inference
arxiv.org/abs/2404.19737
5
196
The question of LLM helping to discover new mathematics is even broader and supervised language model seems to work well, even to the point where it can help mathematicians (a few examples are given in my talk in the link given by François)
2
45
I had a similar experience; I felt like I crossed that threshold a few months ago. Also, Gemini-2.5-pro's "deep think" seems to regularly exceed the threshold of usefulness for proving (at least for me), while GPT-5's is less so (but I assume GPT-5-Pro would exceed it too).
I crossed an interesting threshold yesterday, which I think many other mathematicians have been crossing recently as well. In the middle of trying to prove a result, I identified a statement that looked true and that would, if true, be useful to me. 1/3
4
784
Grande classe @MistralAI !
4
343
Very clear and interesting presentation of our work by @ykilcher! It is well explained and has insightful comments at the end, I definitely recommend watching!
This LANGUAGE MODEL determines stability properties of differential systems, a task that usually requires multiple steps of high-level math and at least three grad students! 😮 watch the video here
piped.video/l12GXD0t_RE
@f_charton @Amaury_Hayat @GuillaumeLample @facebookai
1
4
This "small group at FAIR Paris" was amazing 😉
You misread.
There had been multiple LLM projects within FAIR for years. Some were open sourced as research prototypes (e.g. OPT175B, Galactica, BlenderBot...).
In mid-2022, FAIR started a large LLM project called Zetta, which was still going in late 2022 when ChatGPT came out.
A small group at FAIR-Paris was working on theorem proving. They needed an LLM for their own purpose and thought Zetta was too big and not ready.
They developed their own model, which eventually became Llama-1.
What happened internally between Zetta and Llama is somewhat similar to what just happened between DeepSeek and the big US players: a small team of talented folks innovated and beat the large teams.
4
575
Merci @_Mobility_TV pour cette invitation à parler de véhicules autonomes !
piped.video/watch?v=4NTpNvIN…
3
93
When we started several years ago, the young and naive mathematician I was thought that, should it work, it would qualify as black magic. Several years later and so many successes of AI, I am more rational about it, but still... n=8
3
124
With the amazing @f_charton and @albe_alfa :)
Transformers solve an open problem in symbolic mathematics: discovering Lyapunov functions, joint work with Alberto Alfarano and @Amaury_Hayat.
My talk in IAIFI today (starts at 5:00) piped.video/watch?v=yCzV97QN…
3
208
Very interesting paper! I think this method has great potential for far more problems than what they studied, worth the reading!
New preprint up! "PatternBoost: Constructions in Mathematics with a Little Help from AI," with F. Charton, A.Z. Wagner, and G. Williamson: arxiv.org/abs/2411.00566
3
186
What an amazing paper ;) Congrats to the impressive team of @syhw (I have myself nothing to do with this work, I am just lucky enough to be the PhD advisor of a great student: @FabianGloeckle).
3
256
There's a lot to be said about the US voting system for presidential elections. One of the good things is that the result is so non-robust with respect to the input data that there is suspense right to the end
1
2
267
For example, finding the largest eigenvalue in absolute value of a symmetric real matrix can be done without solving the eigenvalues equation but using instead power iteration of the matrix. The model might be discovering and leverage similar shortcuts. 2/2
3
Congrats ! This is very nice :) Just a quick note, you also mention Putnam Bench in the paper, but I think that the current state of the art is in fact 7/640 held by this paper (probably too recent for you to have seen it) leanprover.zulipchat.com/use… trishullab.github.io/PutnamB…
2
3
76
Un travail Magistral :)
Glad to share what we've been working for the past 4 months. We've got some sweet RL stack and two nice reasoning models. Try it out at chat.mistral.ai, select think/pure thinking.
2
252
Haha, I received the exact same ! Strangely enough, I also often get random invitations in biology and earth sciences
1
37
Congratulations @danij_markov !!
@danij_markov is too modest to announce the big news here, so here it goes:
immense congratulations to you Danijela for getting this permanent position at CNRS @INP_CNRS, and gigantic happiness to have the chance to be able to continue working with you in the same team!!!!!
1
3
Three years of work on algorithms to reduce traffic jams, just became reality! Thanks @Cal_Engineer @Rutgers_Camden @VanderbiltU @TempleUniv @uarizona @EcoledesPonts @UGustaveEiffel @myTDOT @ENERGY, @NissanUSA @Toyota @GM, @INSMI_CNRS and many others!
apnews.com/article/technolog…
1
3
Don't miss the release of multi-token prediction (among many others) ;)
A bunch of releases from Meta FAIR today: ai.meta.com/blog/meta-fair-r…
3
233
It's our latest paper 🎉 arxiv.org/abs/2205.11491 with @GuillaumeLample @MaLachaux @tlacroix6 @LavrilThibaut @javier_m @AurR0d and Gabriel Ebner
1
1
3
Can AI solve math problems? And give the full proof on its own? Apparently, yes! This is a step towards helping mathematicians prove theorems with AI. Let's see what the future of math will look like ✍️ arxiv.org/abs/2205.11491
Excited to release our latest work: arxiv.org/abs/2205.11491 We present a new algorithm, HyperTree Proof Search (HTPS) inspired by the recent success of AlphaZero. Our model is able to prove mathematical theorems in a fully automated way and significantly outperforms the SOTA. 1/n
1
3
Et 72% pensent qu'il y a un danger pour la santé humaine...
Dans le Baromètre de l'@IRSN sur la perception des risques par les Français, plus de 40% des Français croient que la "fumée blanche", en réalité de la vapeur d'eau, des centrales nucléaires représente un danger sérieux pour la santé humaine. Scrongneugneu. lemonde.fr/blog/huet/2024/11…
1
3
302
I'm impressed, as always with these guys!
We just released two small models, with 3B and 8B parameters. Ministral 3B is exceptionally strong, outperforming Llama 3 8B and our previous Mistral 7B on instruction following benchmarks.
mistral.ai/news/ministraux/
3
318
We have not studied the attention mechanism: over such complex calculations it would be very difficult to interpret. An intuition could be: in numerical analysis, "shortcut methods" have been discovered. 1/2
1
3
Congrats to all this team !
I’m pretty excited about our new paper, which is a follow up to our last paper using AI to help solve a problem in theoretical particle physics. (With Lance, @f_charton, Matthias, Tianji, and @merz_garrett
2
127
@Polytechnique is definitely everywhere ;)
1
159
Saint-Venant is famous, among others, for the Saint-Venant equations. He was 74 when he first introduced them. en.wikipedia.org/wiki/Adh%C3…
2
Good question ;) I won't comment on FAIR but I can said we won't claim it reason either: I would argue that it essentially captures a structure in the problem that mathematicians have not yet "mapped" with a theorem.
1
2
69
Just read it, very interesting! It's striking how the way the transformer learns is interpretable in this example
Transformers can learn to compute the greatest common divisor of two positive integers. They make deterministic predictions that can be fully explained. Training from a log-uniform distribution of operands achieves best results.
My new paper is on arXiv: arxiv.org/abs/2308.15594
2
170
Congrats @alex_conneau and Coralie! Happy to see two of the brightest people I know launch such a company!
Excited to announce the creation of WaveForms AI (waveforms.ai) – an Audio LLM company aiming to solve the Speech Turing Test and bring Emotional Intelligence to AI @WaveFormsAI
2
176
Is it new SOTA on Putnam or ex-aequo ABEL from
@FabianGloeckle
? Congrats for managing this with a 7b SFT model, that is truly impressive ! Looking forward to reading the paper !
2
77
Congrats' @f_charton and @KempeLab !
Congratulation to the winners of our debunking challenge @f_charton and @KempeLab for "Emergent properties with repeated examples" 🥳🎉
2
136
It's more a personal collaboration than an institutional one :) But I think it's not the first collaboration between researchers from ENPC and Facebook.
2
That's the concern, a model could become much better at fooling humans, even accomplished mathematicians, than at proving math.
Hence the interest in having models that are expressed in formal languages whose proofs can be verified automatically.
Most of the time, we use clarity as a proxy for truth, because we believe that we only express clearly what we understand well.
Unfortunately, the self-supervised techniques used to train language models seem to do a much better job making them clear, than making them true.
2
174
Congrats @byoubii !
Today is a good day for open science.
As part of our continued commitment to the growth and development of an open ecosystem, today at Meta FAIR we’re announcing four new publicly available AI models and additional research artifacts to inspire innovation in the community and help advance AI in a responsible way. More in the video from @jpineau1.
What we’re releasing:
🦎 Meta Chameleon
7B & 34B language models that support mixed-modal input and text-only outputs.
🪙 Meta Multi-Token Prediction
Pretrained Language Models for code completion using Multi-Token Prediction.
🎼 Meta JASCO
Generative text-to-music models capable of accepting various conditioning inputs for greater controllability. Paper available today with a pretrained model coming soon.
🗣️ Meta AudioSeal
An audio watermarking model that we believe is the first designed specifically for the localized detection of AI-generated speech, available under a commercial license.
📝 Additional RAI artifacts
Including research, data and code to measure and improve the representation of geographical and cultural preferences and diversity in AI systems.
We believe that access to state-of-the-art AI creates opportunities for everyone – not just a small handful of Big Tech companies. We’re excited to share this work and to see how the community learns, iterates and builds using this technology.
Details and access to everything released by FAIR today ➡️ go.fb.me/tzzvfg
2
135
This is only a first work, many things are left to be done to totally solve the problem (classes of systems that not defined on the whole space, singular systems, etc.) but for the first time in this area, neural networks are actually better than humans at finding a solution 7/n
1
2
141
Looking forward to trying !
Mistral fine-tuning API is out !
You can now fine-tune your own Mistral models and deploy them efficiently on La Plateforme : mistral.ai/news/customizatio…
In many cases, fine-tuning allows small models to match (and sometimes surpass) the performance of much larger models, but with a significantly lower cost and improved generation speed.
2
128
Hey @OpenAI ! how can I get access API access to o3-mini ? I'm a Tier 5 user and it would be very helpful for my projects (in particular some research projects) !
1
2
309
I'd say it depends... if it zero-shot with respect to the current LLM, it's unlikely, even with fine-tuned models that would be a challenge. However, I believe it should be possible to develop models that could and that are relying on LLM (but not only) :)
1
2
46
Le service client par téléphone de @SNCFVoyageurs c'est particulier: à 17h35 la messagerie annoncent qu'ils sont fermés parce qu'ils n'ouvrent que de 8h à 18h.
1
81
An exciting opportunity if you're interested in traffic control ;)
We are pleased to announce the CIRCLES workshop "Traffic and Autonomy", June 21-23, 2023, in Maiori (Amalfi Coast), Italy. For more information check the website: maioritrafficandautonomy.git…
1
87
Classic Mistral AI :) Can't wait to try !
magnet:?xt=urn:btih:5546272da9065eddeb6fcd7ffddeef5b75be79a7&dn=mixtral-8x7b-32kseqlen&tr=udp%3A%2F%2Fopentracker.i2p.rocks%3A6969%2Fannounce&tr=http%3A%2F%https://nitter.app/t.co/g0m9cEUz0T%3A80%2Fannounce
RELEASE a6bbd9affe0c2725c1b7410d66833e24
1
133
The crucial point is the generation method: how do you generate enough examples of problems and solutions for a problem that, precisely, you don't know how to solve? 6/n
1
1
47
Bravo @ntaconet ;)
1
Showing that a dynamical system is stable (essentially, the solutions do not explode and are arbitrarily bounded if the initial conditions are) is a difficult problem. 2/n
1
1
53
This is essentially what motivated the paper HyperTree proof search @AlbertQJiang was mentioning in the message above: arxiv.org/pdf/2205.11491.pdf. It's worth looking at if you're interested by this problematic :) (Disclaimer: I am obviously very biased)
1
58
If you're interested in computer-assisted proof and haven't (yet) heard about the @XenaProject, you should probably check this out :)
xenaproject.wordpress.com/
If you're not interested but you like maths, you should probably look at this :)
nature.com/articles/d41586-0…
1
Congratulations @ZuazuaEnrique !
#FAUCongrats Mathematician #FAUProf Enrique Zuazua , Chair in Dynamics, Control and Numerics, has received an ERC Advanced Grant. The highly prestigious awards include up to 2.5 million euros per research project for a period of five years. @ZuazuaEnrique @UniFAU_DCN
1
104
It's still true :) If you like AI and math feel free to fill in and give your opinon !
We are recruiting emergency reviewers for the ICML 2024 AI for Math Workshop (icml.cc/virtual/2024/worksho…). You are reviewing 1-2 papers and submitting them by June 11th, AoE. If you are interested, please use this form: docs.google.com/forms/d/e/1F…
Many thanks for your help and time!
1
104
Cela étant dit le point du traitement médiatique me semblait être: même une année où l'activité humaine est très réduite, on n'a pas atteint le niveau d'activité très faible tel que la concentration de CO2 diminue, ce qui n'est pas contradictoire avec votre point.
1
Some more interesting thoughts on the debate
This doesn't convince me. If web search was on, surely GPT-5 would find the new version of the paper (arXiv shows it by default). Upon studying this, it would at least know that a better bound is possible, and perhaps a rough sense of what sorts of techniques are relevant.
1
451
