Caio

Je sais pas si ça fonctionne encore : en.wikipedia.org/wiki/RapidQ

📢 Statement from ACL and EACL 2026 Organizers

On Nov 27, OpenReview was notified of a software bug that allowed unauthorized access to authors, reviewers, and area chairs. We are grateful to the OpenReview team for fixing the issue quickly. (🧵 1/3)

Statement by OpenReview on X

ndicateur 1.5 : Admission dans l'enseignement supérieur (du point de vue de l'usager) Unité 2022 2023 2024 (Cible PAP 2024) 2025 (Cible) 2026 (Cible) 2027 (Cible) Part des néo-bacheliers ayant obtenu au moins une proposition à la fermeture de Parcoursup % 94,8 95 94,5 93,5 93,5 93,5

Dans ce contexte, la part de néo-bacheliers recevant au moins une proposition diminue légèrement (93,5 %, -1,0 point par rapport à 2024).

[ #VeilleESR #Parcoursup ] Il vient de se passer un truc extraordinaire : le gouvernement a atteint un de ses objectifs.

Celui de baisser la part des néobachliers recevant une proposition dans Parcoursup.

C'est sans doute une date historique, le début de démassification éducative, .
🧵

Cool seminar coming up at @inriaparisnlp.bsky.social. If you guys can't make it on site, a visio link will be provided (check out the seminar webpage 30mn before the talk)

Pourquoi pas Instagram ?

You need this fixed logsumexp function, otherwise you will have NaN gradients for the neural network. I personnaly came across this bug when building a CRF with the following transition structure for discontinuous named entity recognition, see here: aclanthology.org/2024.emnlp-m...

The use case: conditional random fields with forbidden transitions between tags. Altough implementing the forward function is trivial with Pytorch, see for example here: github.com/FilippoC/lig...

But now the backward pass works as expected, and you get null gradients for w1.

Yes: implement your own logsumexp function that fixes this bug. I found the workaround by coming across this github issue: github.com/pytorch/pyto...
The forward pass is basically the same, but using the custom logsumexp function.

But this will give you NaN gradient for w1 ! If you look at the gradient of w2, masked values have a null gradient, as expected. But for w1, instead of having a vector of null gradients, we have a vector of NaNs. This completly breaks gradient backprop and grad descent. So, is there a nice solution?

The input of the first logsumexp is completly masked, the second one is partially masked, the last one has no masked applied on it. Obviously, the first logit is equal to -inf. This means a "masked output probability" after softmax. We can then just compute a loss and backpropagate gradient.

One of the hardest Pytorch bug I had to debug is due to how the logsumexp behave with -inf masked inputs. Consider the following example. I build a vector of 3 logits, and each logit is the result of a logsumexp.

Hallucinant. C'est inacceptable. J'ai une ERC, que rétrospectivement j'ai eu la chance de rédiger à l'étranger avec un dispositif bien meilleur que ceux qui sont proposés en France et SURTOUT avec de quoi produire des résultats préliminaires que ce soit en terme de mentiring ou de moyens #esr

Universities across the world seeing this:

"its only wrong 45% of the time!!
Lets buy free licenses for our students, staff and faculty!!
Lets lock into contracts with rapacious predatory AI companies with shitty technofascist politics, sucking up water and jacking up electricity prices!!"

Qu'est ce que ça va être quand ils vont découvrir l'existence du manifeste ou de l'état et la revolution. 🙃

Not all scaling laws are nice power laws. This month’s blog post: Zipf’s law in next-token prediction and why Adam (ok, sign descent) scales better to large vocab sizes than gradient descent: francisbach.com/scaling-laws...

~9 months ago I spent some time with Luca Soldaini making a list of models and resources for language models that were more than just open weights (data, code, logs, etc included). It's getting out of date, could use some community contributions :)

J'ai pas encore testé, mais le dernier Mafia à l'air ultra stylé

For updates on AI, I increasingly just advise people to pick a discord they like and stick with it. Twitter stopped having interesting science chat ages ago, it’s just companies announcing their products and getting a bunch of meme QTs.

This year, EMNLP ended up desk rejecting ~100 papers. For more insight into the process, and potential future changes, please see this blog post from the PCs: 2025.emnlp.org/desk-rejecti...

@christos-c.bsky.social @carolynrose.bsky.social @tanmoy-chak.bsky.social @violetpeng.bsky.social

Merge transducers for a BPE tokenizer.

My paper "Tokenization as Finite-State Transduction" was accepted to Computational Linguistics.

This was my final PhD degree requirement :)

The goal was to unify the major tokenization algorithms under a finite-state automaton framework. For example, by encoding a BPE tokenizer as a transducer.

#ClubContexte L'École Normale Supérieure de Paris-Saclay (à ne pas confondre avec l'École Normale Supérieure, ou «Ulm», qui est à Paris et pas à Ulm) est l'ancienne École Normale Supérieure de Cachan qui a déménagé: elle n'est située ni à Paris, ni à Saclay, mais à Gif-sur-Yvette …

2 weeks is always unreasonable. 🙃

Titre d'un article : ChatGPT-5: «C'est comme parler à un expert de niveau doctorat»

Un like = une phrase réaliste dite par CHAT GPT 5 s'il s'exprimait vraiment comme quelqu'un avec un doctorat.

At 17, Hannah Cairo Solved a Major Math Mystery | Quanta Magazine After finding the homeschooling life confining, the teen petitioned her way into a graduate class at Berkeley, where she ended up disproving a 40-year-old conjecture.

What a fantastic accomplishment -- and what a fantastic story! www.quantamagazine.org/at-17-hannah...

ICML found hidden prompts in accepted papers. They have released a statement icml.cc/Conferences/...

Yes, it’s unacceptable. So is using an LLM to review a paper. Peer review is so broken.

ACL paper alert! What structure is lost when using linearizing interp methods like Shapley? We show the nonlinear interactions between features reflect structures described by the sciences of syntax, semantics, and phonology.

I think this will be my last time being an AC for an @aclmeeting.bsky.social / @emnlpmeeting.bsky.social related conference. Authors rebuttals are mostly very long LLM generated answers, and spending time reading them is not an interesting activity.

Rappelons à toutes fins utiles à l'ESR français (qui parfois utilise WeTransfer pour du professionnel) qu'il a filesender.renater.fr francetransfert.numerique.gouv.fr/upload et sans doute d'autres services dans les organismes.