Samuel Vaiter

Samuel Vaiter

@samuelvaiter.com

CNRS Researcher website: https://samuelvaiter.com

3K followers274 following455 posts

Samuel Vaiter reposted

Rémi Flamary
@rflamary.bsky.social

POT Python Optimal Transport is 10 years old this week 🎂. Did not think when I pushed the first commit 10 years ago that it would still be here or have this impact (~11M dowloads on PyPI). Open Source Software is a marathon not a sprint. Thanks to all contributors and users (who cite it) 🥳.

Samuel Vaiter reposted

G'MIC
@gmic-eu.bsky.social

🎉 G'MIC 4.0.0 is here — a major release focused on stability. It also opens a new chapter: fewer releases ahead, with more focus on framework robustness & performance rather than big changes. The core is mature — now we polish it. Developed @ GREYC lab (CNRS/ENSICAEN/UniCaen) ▶️ gmic.eu @cnrs.fr

Samuel Vaiter reposted

LJAD-Off
@ljad-off.bsky.social

Congratulations to Sophie Jaffard laureate of the prize Jacques Neveu from SMAI, for her phd on mathematical models of neuronal learning

Samuel Vaiter
@samuelvaiter.com

Bienvenue à Nice, ville avec une des plus basses températures de France ! (ps: je recrute potentiellement un.e postdoc sur 24 mois d'ici fin 2026 sur des questions, plutôt théoriques, liées aux LLMs en post-training / alignment, je dis ça, je dis rien)

Samuel Vaiter
@samuelvaiter.com

Now accepted at ICML 2026! Congrats to Magamed :)

Samuel Vaiter@samuelvaiter.com

🧵 New preprint! Towards Understanding Steering Strength w/ M. Taimeskhanov 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. 📄 arxiv.org/abs/2602.02712 1/7

View full thread
View full thread
Samuel Vaiter
@samuelvaiter.com

And before our conference dinner, we had the pleasure to hear Yann Brenier with an outreach talk on the role of optimal transport in applied mathematics.

View full thread
Samuel Vaiter
@samuelvaiter.com

🧵 New preprint! Towards Understanding Steering Strength w/ M. Taimeskhanov 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. 📄 arxiv.org/abs/2602.02712 1/7

View full thread
Samuel Vaiter
@samuelvaiter.com

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

Samuel Vaiter
@samuelvaiter.com

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/steering 7/7

arxiv.orgTowards Understanding Steering StrengthA popular approach to post-training control of large language models (LLMs) is the steering of intermediate latent representations. Namely, identify a well-chosen direction depending on the task at ha...
Samuel Vaiter
@samuelvaiter.com

🧵 New preprint! Towards Understanding Steering Strength w/ M. Taimeskhanov 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. 📄 arxiv.org/abs/2602.02712 1/7

View full thread
Samuel Vaiter
@samuelvaiter.com

Result 2: Concept probability follows a sigmoidal curve. The probability that the target concept appears in the output increases smoothly with α, following a tanh shape. Off-target concepts decrease or vanish. 4/7

Samuel Vaiter
@samuelvaiter.com

Result 3: Steering always hurts global performance. Cross-entropy increases quadratically around α = 0, i.e., there's no free lunch. For large α, performance plateaus because the output becomes input-independent. 5/7

Samuel Vaiter
@samuelvaiter.com

🧵 New preprint! Towards Understanding Steering Strength w/ M. Taimeskhanov 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. 📄 arxiv.org/abs/2602.02712 1/7

Samuel Vaiter
@samuelvaiter.com

Take contrastive prompts (for ex, safe vs malicious), evaluate them through the model, compute the mean difference of representations  at layer ℓ. That's your steering vector v. Then shift activations by αv at layer ℓ. Question: what does α actually do? 2/7

Samuel Vaiter
@samuelvaiter.com

Result 1: Next-token probabilities have a bump shape. As α increases, most token probabilities rise, peak for α>0, then fall. In particular, off-target tokens peak _before_ target tokens. 3/7

Samuel Vaiter reposted

Yann Traonmilin
@ytraonmilin.bsky.social

Save the date : on organise un workshop sur les maths de l'IA à Bordeaux les 4-6 novembre. N'hésitez pas ! plus on est de fous plus on rit ! (dans les limites des capacités d'accueil haha) Dates & infos : wmathsia2026.sciencesconf.org

wmathsia2026.sciencesconf.orgWorkshop sur les mathématiques de l'IA - Sciencesconf.orgLe workshop se déroulera sous forme de présentations sur appel à contribution (résumé d'une page). Une session poster sera organisée au besoin selon le nombre de contributions.
Samuel Vaiter
@samuelvaiter.com

I tested Pocket TTS from Kyutai. It's fun to use for generating French sentences: the accent is strong, but it's still perfectly understandable, even though French doesn't seem to be in the training dataset! Beyond that, the model is quite impressive especially knowing it has only 100M parameters

github.comGitHub - kyutai-labs/pocket-tts: A TTS that fits in your CPU (and pocket)A TTS that fits in your CPU (and pocket). Contribute to kyutai-labs/pocket-tts development by creating an account on GitHub.
Samuel Vaiter
@samuelvaiter.com

Extension pour soumettre une communication : deadline au 21 janvier 23:59 !

Samuel Vaiter@samuelvaiter.com

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**.

Samuel Vaiter
@samuelvaiter.com

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**.

Samuel Vaiter
@samuelvaiter.com

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 !

mode2026.sciencesconf.orgJournées SMAI-MODE 2026 - Sciencesconf.orgLes journées SMAI-MODE, la conférence biennale du groupe MODE de la Société de Mathématiques Appliquées et Industrielles (SMAI), se tiendront du 18 au 20 mars 2026 à Nice.

Samuel Vaiter reposted

Transactions on Machine Learning Research
@tmlrorg.bsky.social

Replying to @tmlrorg.bsky.social

🏅Finalist: Gradient Scarcity in Graph Learning with Bilevel Optimization, by Hashem GHANEM, Samuel Vaiter, Nicolas Keriven 5/n openreview.net/forum?id=10Y...

openreview.netGradient Scarcity in Graph Learning with Bilevel OptimizationGradient scarcity emerges when learning graphs by minimizing a loss on a subset of nodes under the semi-supervised setting. It consists in edges between unlabeled nodes that are far from the...
Samuel Vaiter
@samuelvaiter.com

Est-ce que certain d'entre vous #ESR sont dans un laboratoire qui gère les missions / conférences avec un système de ticket genre GLPI ou Jira ? Si oui, est-ce que c'est l'apocalypse ou plutôt mieux qu'une gestion traditionnel ? Si non, quel niveau d'enthousiasme ou de dégout cela vous inspire ?