🔎 Journée internationale des droits des femmes : portrait de Simona Rota Nodari, chercheuse en mathématiques au LJAD, à découvrir dans « Ces femmes qui font bouger les lignes »
👉 univ-cotedazur.shorthandstories.com/ces-femmes-q...
#8mars #FemmesEnScience #UniversiteCoted'Azur #VilledeNice #CNRS
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
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
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
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
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
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
🧵 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
Je pensais que c'était une façon de parler le 30, du coup j'ai été voir. watzeactualfuk?
Bullshit is like entropy nowadays it can only increase.
My god I never thought about it in this way. This is just the second law of thermodynamics 😍!
The issue is that fighting bullshit (LLMs or not) requires more effort than generating it (LLMs or not). We are doomed to lose.
It’s incredible that the growth rate of administrative bullshit outpaces LLM progress, but here we are.
but what about all the meme then?
Vu que le chauffage est visiblement un détail, on trouve des alternatives. Startup-nation 🤘
À deux doigts de me traiter de jeune 🥹
S-I-X postes ?!
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
Tu sais te mettre de bonne humeur le samedi matin toi 😐
After the 2168 series, I read @fabinou.bsky.social series on building a Quake PC fabiensanglard.net/quake_pc/ (late 90s-style), it brings back so much memories 😍!
So I would say that much of it is probably the reference voice, and some "higher order" features (to speak in a old-fashioned way) are manipulated around this reference.
So I tested with my voice (speaking only French except maybe two words).
- When asking pocket-tts to generate English, this is really my voice with my horrible accent.
- When asking to generate French, it starts to be weird 😅
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
Extension pour soumettre une communication : deadline au 21 janvier 23:59 !
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**.
Mail annuel du CNRS pour savoir si tu as obtenu le prix Nobel ✅
From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers. R. Kawata · Y. Song · A. Bietti · N. Nishikawa · T. Suzuki · SV · D. Wu. Wed, Dec 3, 2025 • 4:30 PM – 7:30 PM PST, #4001 (spotlight) Differentiable Generalized Sliced Wasserstein Plans. L. Chapel · R. Tavenard · SV. Fri, Dec 5, 2025 • 11:00 AM – 2:00 PM PST, #1000 Learning Theory for Kernel Bilevel Optimization. F. El Khoury · E. Pauwels · SV · M. Arbel. Fri, Dec 5, 2025 • 4:30 PM – 7:30 PM PST, #3005
I am at #NeurIPS2025, reach out if you want to chat!
🏅Finalist: Gradient Scarcity in Graph Learning with Bilevel Optimization, by Hashem GHANEM, Samuel Vaiter, Nicolas Keriven 5/n
openreview.net/forum?id=10Y...
Happy to be at #NeurIPS 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, @samuelvaiter.com, @michael-arbel.bsky.social
