Samuel Vaiter

@tonysilveti I don't want to look at my tokens usage this evening 🥹

@vaiter its better to get them to yell at each other

So, yelling at a LLM to simplify a Lean proof really works!

Now accepted at ICML 2026! Congrats to @MagamedTm

TL;DR: Steering strength α has precise, predictable effects on LLM outputs. Understanding these laws may help design better steering strategies. Paper: arxiv.org/abs/2602.02712 Code: github.com/MagamedT/steer… 7/7

We validate everything empirically on 11 models (GPT-2, Gemma 3, Qwen 3, Llama 2, Mistral 7B) across 8 safety-related concepts. All our theorems are confirmed experimentally. 6/7

🧵 New preprint! "Towards Understanding Steering Strength" with @MagamedTm and D. Garreau Activation steering is a popular way to control LLM behavior at inference. But how much should you steer? We provide the first theoretical analysis of the steering strength α. 1/7

@SebAaltonen @IceSolst This is the same shit for every craft out there: Oh look 500 hundreds songs a day Oh look 1000 video game assets a day Oh look 15k LOC a day Oh look 50 videos post-processed a day

@IceSolst Who is going to review those 15k lines produced every day? Quality must be shit. I don't understand why you need to produce that much code. Less is more. Don't want to drown in boilerplate.

Btw avg software engineer ships 10-50 LoC/day. A junior with falsely inflated self-confidence might ship 200 shitty lines. You have to be an absolute dumbass disconnected from any figment of reality to claim your goal is 15k/day. It’s a misleading and toxic claim. 15k/day CVEs.

@chanwoopark20 Maybe of interest to you: we derive a *local* Lipschitz constant of the softmax in this paper #page=27" target="_blank" rel="nofollow noopener">arxiv.org/pdf/2303.07203… (Lemma H.6). It is local, but gives more information than the 1/2 that can be quite pessimistic for small perturbations.

A somewhat well-known property of the softmax function is that it is 1-Lipschitz (from Gao et al 2018). While thinking about whether this bound is tight, I came across a fun and surprisingly simple result: softmax is in fact 1/2-Lipschitz (Newhouse, Feb 2025). The argument is quite elementary—perhaps too simple for a mathematician—but I nonetheless found it very interesting. Later, I discovered that Nair (Oct 2025) independently obtained the same 1/2-Lipschitz bound. The difference lies in the choice of the ℓp​ norm, but I am fairly confident that the author was unaware of Newhouse’s earlier work as the author did not cite Newhouse's paper. Finally—and perhaps most intriguingly—I found a Math StackExchange post from nearly nine years ago showing that softmax is \sqrt{d-1}/d-Lipschitz. Since we assume d≥2, this immediately implies a 1/2-Lipschitz bound as well. Math StackExchange knows everything...

@Andrea__M Now *this* is science fiction!

Even when all math problems will be solved by AI, we will be left with the really hard questions. Like, can we build a slide projector that works 99.9% of the times?

Pour s'inscrire et pour plus d'informations : mode2026.sciencesconf.org Les orateurs et oratrices pléniers sont Pierre Ablin, Yann Brenier, Julie Delon, Stéphane Gaubert, Francisco J. Silva Álvarez (prix J.J Moreau) et Irène Waldspurger. Attention, le nombre de place est limité !

Les inscriptions aux journées MODE 2026 à Nice sont désormais ouvertes. Elles se dérouleront du 18 au 20 mars à l'Hôtel Saint-Paul. Les inscriptions sont ouvertes jusqu'au 1 mars (majoration > 9/02). La deadline pour soumettre une communication est le **15 janvier**.

I am at #NeurIPS2025, reach out if you want to chat!

🏅Finalist: Gradient Scarcity in Graph Learning with Bilevel Optimization, by Hashem Ghanem, @vaiter, @n_keriven 5/n openreview.net/forum?id=10YJT…

Happy to be at #NeurIPS2025 in San Diego to present our poster ‘Learning Theory for Kernel Bilevel Optimization’ #3005, Fri at 4:30 p.m. Stop by/ping me to chat, especially about statistics, causality, generative models! Let's connect! Joint w/ E. Pauwels, @vaiter, @MichaelArbel

Les exposés pléniers seront donnés par : Pierre Ablin, Julie Delon, Stéphane Gaubert, Irène Waldspurger, ainsi que que le/la lauréat(e) du prix J.J. Moreau. Le mini-cours sera sur les PEP (Performance Estimation Problem) par A. Dieuleveut et A. Taylor

Les journées SMAI-MODE 2026 auront lieu à Nice du 18 au 20 mars 2026, précédées par un mini-cours les 16/17. Vous pouvez déposer votre proposition de contrib sur mode2026.sciencesconf.org Inscription à partir du 1er décembre, fin le 1 mars (tarif majoré à compter du 1 février).