
Postdoc at @UCPH_Research. Interested in the cocktail-party problem, causal inference and identifiability. Previously @MPI_IS.
Joined March 2019
- Tweets140
- Following467
- Followers712
- Likes507
Luigi Gresele retweeted
🎥 The recording of the third ELLISxUniReps Speaker Series session with @victorveitch and @luigigres
is now available at: youtu.be/8mM_B0VKj-8.
Next appointment: 31st July 2025 – 16:00 CEST on Zoom with 🔵Keynote: @Pseudomanifold
(University of Fribourg) 🔴@FlorentinGuth (NYU & Flatiron)
🎉 Get ready for our 3rd @ELLISforEurope × UniReps Speaker Series session! 🔥
🗓️ When: 8th July, 2025 – 16:00 CEST
📍 Where: ethz.zoom.us/j/66426188160
🎙️ Speakers: Keynote by @victorveitch & Flash Talk by @luigigres
👉 Stay updated by joining our Google group: groups.google.com/u/2/g/elli…
Luigi Gresele retweeted
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)
Luigi Gresele retweeted
My thesis is now online!
umontreal.scholaris.ca/items…
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.
Luigi Gresele retweeted
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
Luigi Gresele retweeted
🧵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
Luigi Gresele retweeted
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 👇
arxiv.org/abs/2410.23501v1
This was fantastic news. Thanks @ELLISforEurope!
My thesis, "Learning Identifiable Representations: Independent Influences and Multiple Views", can be found here: tobias-lib.ub.uni-tuebingen.…
Congrats to the recipients of the 2024 ELLIS PhD Award!
Co-winners: @Ana_koloskova (efficiency in decentralized learning) & @luigigres (identifiable representation learning)
Runner up: @schwarzjn_ (sparse parameterizations)
Read more about them: bit.ly/4fg4jkg #AI #ML
Really grateful to my supervisor, @bschoelkopf, and to all the co-authors of the papers on which my thesis is based. It was great to work with you! Still processing all the things I learned
1
147
Luigi Gresele retweeted
Why do NNs often learn similar representations? Existing identifiability results offer theoretical insights, but applying them in practice poses challenges.
We’ll present our new work exploring these challenges next week at @unireps #NeurIPS2024 🇨🇦🎉
openreview.net/pdf?id=SQKUZS…
1/
Luigi Gresele retweeted
📢 here's a new 𝚐𝚊𝚍𝚓𝚒𝚍 for Causal Structure Learning:
𝐆raph 𝐀𝐃𝐉ustment 𝐈dentification 𝐃istances
to compare DAGs or CPDAGs (!)
in terms of the identifying formulas they suggest for causal effects.
💻 Try it: pip install gadjid
📚 Read it: doi.org/10.48550/arXiv.2402.…
1/
Luigi Gresele retweeted
Formally speaking, it shouldn't be possible. Using a benchmark based on @yudapearl ladder of causation (association < intervention < counterfactual), @zhijingjin et al. show that GPT-4 struggles with counterfactual reasoning, obtaining only 60% accuracy.
arxiv.org/abs/2312.04350
#827 is the correct number! (Sorry, jet lag) See you there!
If you are at #NeurIPS2023 and you want to hear about our work on Causal Component Analysis, come to our poster at 10:45! #413, Hall B1+B2
@liang_wendong1, @armin_kekic, @JKugelgen, Simon Buchholz, @MichelBesserve, @bschoelkopf
If you are at #NeurIPS2023 and you want to hear about our work on Causal Component Analysis, come to our poster at 10:45! #413, Hall B1+B2
@liang_wendong1, @armin_kekic, @JKugelgen, Simon Buchholz, @MichelBesserve, @bschoelkopf
Check out this thread by @liang_wendong1 for a summary:
teapawt.net/liang_wendong1/status/…
Introducing our #NeurIPS2023 paper: "Causal Component Analysis" (CauCA) – a causal generalization of Independent Component Analysis (ICA), and a special case of Causal Representation Learning (CRL). 1/10
📝: arxiv.org/abs/2305.17225
💻: github.com/akekic/causal-com…
1
3
405
I ended up in a part of NeurIPS that looks like one of the Twitter posts by @tetraduzione
1
17
1,506
Luigi Gresele retweeted
If you wonder whether #LLMs can solve causal reasoning, talk to us at our Poster “CLadder: Assessing Causal Reasoning in LLMs”
⭐️Time & Location: Today 10:45-12:45 (Poster #413)
⭐️6 authors in-person: @ZhijingJin @chenyuen0103 @felixludos @luigigres @kamalojasv181 @maxhkw
Can LLMs pass 𝐂𝐚𝐮𝐬𝐚𝐥 𝐓𝐮𝐫𝐢𝐧𝐠 𝐓𝐞𝐬𝐭? In our #NeurIPS2023 paper “CLadder: A Benchmark to Assess Causal Reasoning Capabilities of LMs”, we test LLMs on the Pearl's 3 Rungs of Causality, and found that LLMs are *far* from solving them.➡️arxiv.org/abs/2312.04350
Luigi Gresele retweeted
4/5 Wed 6 pm (Poster Session 4)
"Nonparametric Identifiability of Causal Representations from Unknown Interventions"
(joint work with @JKugelgen @MichelBesserve, Liang Wendong, @luigigres, Armin Kekić, David Blei, and @bschoelkopf )
We study the task of learning a causal representation in non-parametric settings, which includes both the latent variables and the corresponding causal graph. The paper introduces a new equivalence class over the learned representation, capable of encoding the causal constraints of the true underlying latent variables. There are some interesting impossibilities (minimality of the targeted equivalence class) that are addressed in this definition: its ambiguities cannot be solved with data coming from unknown interventions. We show that identifiability is possible from single-node, perfect interventions. Interestingly, we further show that certain measures of influence between variables are preserved within the equivalence class, which is a significant advancement in the context of inferences from the learned object. The empirical results corroborate the theory.
Link: arxiv.org/pdf/2306.00542.pdf
Luigi Gresele retweeted
Interested in identifiable causal representation learning with flexible nonlinear modules?
👇 A summary of our #NeurIPS paper “Nonparametric Identifiability of Causal Representations from Unknown Interventions”.
📜 arxiv.org/abs/2306.00542
@MPI_IS @CambridgeMLG @DataSciColumbia































