Sébastien Lachapelle
@seblachap
Followers
826
Following
300
Media
27
Statuses
119
Research Scientist at SAIL Montreal (Samsung) interested in causality and identifiable representation learning. PhD from @Mila_Quebec, @UMontrealDIRO
Montréal, CA
Joined March 2016
1/ Excited for our oral presentation at #NeurIPS2023 on "Additive Decoders for Latent Variables Identification and Cartesian-Product Extrapolation"! A theoretical paper about object-centric representation learning (OCRL), disentanglement & extrapolation https://t.co/LLtRlj7ayB
4
26
105
Congratulations to my colleague at Samsung @jm_alexia for winning the ARC Prize 2025 Paper Award! 1st place! Well deserved! This is a BIG deal :)
ARC Prize 2025 Paper Award Winners 1st / "Less is More: Recursive Reasoning with Tiny Networks" (TRM) / A. Jolicoeur-Martineau / $50k 2nd / "Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC-AGI" (SOAR) / J. Pourcel et al. / $20k 3rd /
0
0
6
ARC Prize 2025 Paper Award Winners 1st / "Less is More: Recursive Reasoning with Tiny Networks" (TRM) / A. Jolicoeur-Martineau / $50k 2nd / "Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC-AGI" (SOAR) / J. Pourcel et al. / $20k 3rd /
4
30
286
Link to official announcement, in French:
nouvelles.umontreal.ca
Les Prix de la meilleure thèse 2025 ont été remis à trois doctorants dont les travaux se sont nettement démarqués dans leur secteur d’études.
0
0
0
🚨I have received the "Best Thesis Award 2025" from the Université de Montréal in the "Natural Sciences and Engineering" category! 🚨 Thesis: https://t.co/kdraSXNoB6 Thesis defense: https://t.co/tff79g0lT8 Explainer for a wide audience (in French): https://t.co/RlmVJtPZDH
My thesis is now online! https://t.co/kdraSXMQLy This is more than just a list of publications. I invested a lot of time and passion writing this thesis in hope that it will make for an interesting read. Here's a summary of what you'll find in it.
5
4
23
Finally published: “Top-down and bottom-up neuroscience: overcoming the clash of research cultures” https://t.co/oQU40t0vTX... Looking for ways to better understand different neuroscientific perspectives and enable productive collaborations
5
47
188
If you're at ICML on Saturday, check out our workshop paper on identifiable steering from multi-concept shifts in language models!
I will be at the Actionable Interpretability Workshop (@ActInterp, #ICML) presenting *SSAEs* in the East Ballroom A from 1-2pm. Drop by (or send a DM) to chat about (actionable) interpretability, (actionable) identifiability, and everything in between!
0
0
9
Oh and in case some of you are interested, my thesis defense is available here:
1
0
9
This would not have been possible without my amazing supervisor @SimonLacosteJ and my great coauthors: @_kurowasan @TristanDeleu @divyat09 @Qu3ntinB @alexandredrouin @alex_lacoste_ @prlz77 @yash_j_sharma @_katieeverett @bouzoukipunks @Yoshua_Bengio & Rémi Le Priol
1
1
4
... It further provides four concrete problem settings where this pattern occurs. Everything is formalized in the framework of statistical decision theory.
1
0
1
The additional chapter also argues that identifiability guarantees can often be used as a mediating step toward proving generalization guarantees, providing an additional motivation to study identifiability.
1
0
0
An extra chapter that describes three different interpretations of identifiability, namely: - the realist interpretation - the independent-learners interpretation, and - the interpretability interpretation. (also see @luigigres's thesis for more on this topic)
1
1
5
A background section covering, among other topics, - Statistical Decision Theory - Causal inference/discovery - Identifiability proof for independent component analysis (ICA) - Identifiability proof for AMUSE (a cool temporal variant of ICA)
1
0
0
The five papers I've published during my PhD together with "Prologues" which - tell the stories of how each paper came about - give some of their limitations - review works that came after publication, and - identify potential future directions.
1
0
0
My thesis is now online! https://t.co/kdraSXMQLy This is more than just a list of publications. I invested a lot of time and passion writing this thesis in hope that it will make for an interesting read. Here's a summary of what you'll find in it.
7
17
116
In this one, we explore the linear properties of next-token predictors like LLMs through the lens of identifiability. I learned quite a bit while working on this project! In collaboration with folks from University of Trento and Copenhagen University!
🧵Why are linear properties so ubiquitous in LLM representations? We explore this question through the lens of 𝗶𝗱𝗲𝗻𝘁𝗶𝗳𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆: “All or None: Identifiable Linear Properties of Next-token Predictors in Language Modeling” Published at #AISTATS2025🌴 1/9
0
2
11
This was a really fun collaboration with folks from the Max Planck Institute in Tuebingen. Make sure to pass by our poster if you are at ICLR!
I'm at #ICLR2025 presenting our work on compositional generalization! (Sat. 10 AM; Hall 3 + Hall 2B, #310) We provide a general and unifying theory of compositional generalization, based on a new principle called interaction asymmetry! 📜 https://t.co/J4ZUfxrQm4 (See 🧵)
0
0
21
I'm at #ICLR2025 presenting our work on compositional generalization! (Sat. 10 AM; Hall 3 + Hall 2B, #310) We provide a general and unifying theory of compositional generalization, based on a new principle called interaction asymmetry! 📜 https://t.co/J4ZUfxrQm4 (See 🧵)
2
25
127
1\ Hi, can I get an unsupervised sparse autoencoder for steering, please? I only have unlabeled data varying across multiple unknown concepts. Oh, and make sure it learns the same features each time! Yes! A freshly brewed Sparse Shift Autoencoder (SSAE) coming right up. 🧶
1
10
46
This was a really fun collaboration :) Working on this project allowed me to gain a better understanding of the linear representation hypothesis in LLM by formalizing it using ideas from identifiable/causal representation learning. Check it out!
Yo! We have an accepted paper at #AISTATS2025!! Time to prepare for Thailand 🪷🏖️🌴🐒 Huge thanks to my coauthors @luigigres, @sweichwald, and @seblachap for all the joint effort! More details soon 👇 https://t.co/FG4jouYkcG
0
1
6