PostDoc @EPFL. I like doing stuff with representations. 🇨🇭🚴
Switzerland
Joined February 2021
- Tweets110
- Following287
- Followers148
- Likes592
Pinned Tweet
Happy to share what we’ve been up to! 🏺
TL;DR: Some of the „representations converge as models scale“ evidence is a measurement artifact. With a simple null calibration, the global convergence trend disappears, while local neighbor relationships stay consistent.
Are neural nets across modalities really converging to the same representation as they scale, as the Platonic Representation Hypothesis suggests?
We show that common representational similarity metrics are confounded by network width & depth. We propose a permutation-based null calibration that fixes this.
Result❓
• Global convergence largely disappears.
• Local neighborhoods persist.
We propose the alternative Aristotelian Representation Hypothesis: Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships
Very proud of @FabianGroger and @ShuoWen18 for this work!
Paper: arxiv.org/abs/2602.14486
Webpage: brbiclab.epfl.ch/aristotelia…
Code: github.com/mlbio-epfl/aristo…
Fabian Gröger retweeted
#NeurIPS2026 To align or not to align?
Building on our Aristotelian Representation Hypothesis, we turn shared local neighborhoods into a practical guide for choosing compatible model pairs.
We introduce C🧭MPASS, a cheap diagnostic that estimates model-pair compatibility before alignment or merging. It helps decide what to do next: act now, invest more compute, switch training paradigms, or deprioritize a pair.
Across 72 vision-language alignment experiments and 392 model merges, COMPASS correlates downstream success:
• Alignment: 0.91 correlation
• Model merging: 0.70 correlation
Check your C🧭MPASS before you train.
Kudos to @FabianGroger and Petar Damjanovic 🙌
Paper: arxiv.org/abs/2610.18638
Webpage: brbiclab.epfl.ch/projects/co…
Code: github.com/mlbio-epfl/COMPAS…
Fabian Gröger retweeted
Another banger for people interested in vision and language alignment/PRH!
After @dominik_schnaus got people excited a few days ago that we can unpaired align vision and language embedding spaces, now this paper shows: you can also unpaired train the projector from vision *tokens* to LLM input space
#NeurIPS2026 What does projector training actually do in multimodal LLMs and can we skip it while retaining most downstream performance?
We investigate its role and find that it mainly provides coarse alignment. This insight led us to PORTAL 🌀: a training-free projector initialization requiring no paired captions and no gradient updates. Skip projector training stage entirely, then proceed with multimodal instruction tuning.
Across 16 benchmarks, PORTAL 🌀:
• Beats the standard 558k-pair pipeline in 3 of 7 model configurations.
• Reaches an average 98% of its score in the other four.
• With a little projector training, matches the baseline with 135x fewer alignment FLOPs.
Proud of my students @FabianGroger, Naël Ouerghemi and @ShuoWen18 for this work!
Paper: arxiv.org/abs/2610.18598
Webpage: brbiclab.epfl.ch/projects/po…
Code: github.com/mlbio-epfl/PORTAL
Fabian Gröger retweeted
#NeurIPS2026 What does projector training actually do in multimodal LLMs and can we skip it while retaining most downstream performance?
We investigate its role and find that it mainly provides coarse alignment. This insight led us to PORTAL 🌀: a training-free projector initialization requiring no paired captions and no gradient updates. Skip projector training stage entirely, then proceed with multimodal instruction tuning.
Across 16 benchmarks, PORTAL 🌀:
• Beats the standard 558k-pair pipeline in 3 of 7 model configurations.
• Reaches an average 98% of its score in the other four.
• With a little projector training, matches the baseline with 135x fewer alignment FLOPs.
Proud of my students @FabianGroger, Naël Ouerghemi and @ShuoWen18 for this work!
Paper: arxiv.org/abs/2610.18598
Webpage: brbiclab.epfl.ch/projects/po…
Code: github.com/mlbio-epfl/PORTAL
Fabian Gröger retweeted
In PRH, we argued that representational geometry is converging, but it's still an open question exactly in what sense.
I want to share some new evidence that might clarify the picture.
The evidence comes from our work here, led by @dominik_schnaus: dominik-schnaus.github.io/un…
1/n
Come do a postdoc with us at EPFL :)
We're looking for people excited about AI + biology, broadly defined! From foundation models and multimodal learning to spatial biology and AI agents for science.
Please spread the word.
Fabian Gröger retweeted
🔵🔴 Join us for the 4th edition of UniReps on December 12, 2026, in Paris!
Working on these or similar topics? Submit your paper by the extended deadline: October 10 (AoE)
Submit at: openreview.net/group?id=UniR…
full CFP: unireps.org/2026/call-for-pa…
Fabian Gröger retweeted
Happy to share that my first ever paper is going to NeurIPS! :)
LLMs, trained only on text, might already know more about other modalities than we realized; we just need to find ways elicit it.
project page: sophielwang.com/sensory
w/ @phillip_isola and @thisismyhat
Fabian Gröger retweeted
**new paper: Superposition in Brains**
What if Barlow was wrong? In his neuron doctrine (1972), he suggested to search for the 'trigger features' of each neuron to understand cortex. But what if that is the wrong coordinate system to understand neural representations?
biorxiv.org/content/10.64898…
Fabian Gröger retweeted
Register your interest!
🚀 UniReps 4th Edition is coming! 🌍✨
We’re aiming to host a UniReps satellite event around NeurIPS 2026 in Paris!
Join us to connect, collaborate, and shape the future of representation learning.
Your voice matters! Let us know which dates work best for you.
docs.google.com/forms/d/e/1F…
Fabian Gröger retweeted
It surprised me that the platonic representation paper got so much discussion, but this paper didn't.
It's also one of the most severe academic burns I've seen in a while!
Fabian Gröger retweeted
The Platonic Representation Hypothesis is mostly a statistical illusion.
New research shows that the apparent "global convergence" of scaled AI models is actually a mathematical artifact of model width and depth selection bias.
Once calibrated, global convergence vanishes. 🧵
Fabian Gröger retweeted
New ICML 2026 paper challenges the Platonic Representation Hypothesis: model width and depth mechanically inflate similarity scores, creating a misleading global convergence trend which disappears after that bias correction.
What survives is local neighborhood alignment across image, text, and video models (similar things stay near each other even across very different architectures). They call it the Aristotelian Representation Hypothesis :)
So while two models may not share the same representation space in any strong global sense, they can still agree on local neighborhoods (what is similar to what). That is probably the part we actually use in retrieval, transfer, and multimodal systems, and which could be transferred between learned approximations of different models.
arxiv.org/pdf/2602.14486
Fabian Gröger retweeted
Would you like to join the research effort on JEPA and World Models easily?
After a full year of hard work, we’re excited to finally release stable-worldmodel:
an open-source, scalable platform built to accelerate JEPA & World Model research!
📄: github.com/galilai-group/sta…
Fabian Gröger retweeted
Excited to share our #ICML paper introducing PACER, a scalable framework for causal discovery from large-scale interventional data.
PACER guarantees acyclicity by design, enabling direct optimization over valid causal structures and showing that scalable causal discovery is achievable through principled search-space design.
Across protein signaling networks and large-scale Perturb-seq benchmarks, PACER matches or outperforms state-of-the-art causal discovery methods while being much more efficient and scaling to thousands of variables.
Huge kudos to everyone involved in this work, especially Ramon Vinas Torne and three fantastic Master students Silvia Fabregas, Soyon Park and Ivo Ban 🚀
Paper: openreview.net/pdf?id=BS8upx…
Code: github.com/mlbio-epfl/PACER
Project page: brbiclab.epfl.ch/pacer
Fabian Gröger retweeted
Aristotelian Representation Hypothesis: "representations in neural networks are converging to shared local neighborhood relationships"
"The apparent convergence in Platonic Representation Hypothesis largely disappears after calibration, while local neighborhood similarity, but not local distances, retains significant agreement across different modalities."
Fabian Gröger retweeted
Excited to share that our work on the Aristotelian Representation Hypothesis ⬇️ has been accepted at #ICML!
Big shoutout to @FabianGroger and @ShuoWen18 👏
We'll also be presenting another #ICML paper on causal discovery from large-scale interventional data, preprint coming 🔜
Are neural nets across modalities really converging to the same representation as they scale, as the Platonic Representation Hypothesis suggests?
We show that common representational similarity metrics are confounded by network width & depth. We propose a permutation-based null calibration that fixes this.
Result❓
• Global convergence largely disappears.
• Local neighborhoods persist.
We propose the alternative Aristotelian Representation Hypothesis: Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships
Very proud of @FabianGroger and @ShuoWen18 for this work!
Paper: arxiv.org/abs/2602.14486
Webpage: brbiclab.epfl.ch/aristotelia…
Code: github.com/mlbio-epfl/aristo…
Fabian Gröger retweeted
Revisiting the Platonic Representation Hypothesis: An Aristotelian View
arxiv.org/abs/2602.14486
✍️ @FabianGroger @ShuoWen18 @mariabrbic via @arxiv, MLBio Lab at @EPFL_en
👉 "We revisit the Platonic Representation Hypothesis with our calibration framework, which reveals a nuanced picture: the apparent convergence reported by global spectral measures largely disappears after calibration, while local neighborhood similarity, but not local distances, retains significant agreement across different modalities"
👉 "Based on these findings, we propose the Aristotelian Representation Hypothesis: representations in neural networks are converging to shared local neighborhood relationships"
@Corix_JC @ahier @sim010101 @Sonu_Monika @maponi @sallyeaves @CEO_AISOMA @dinisguarda @JagersbergKnut @Shi4Tech @FernandaKellner @PVynckier @EstelaMandela @chidambara09 @sulefati7 @SusanHayes_ @theomitsa @TarakRindani @FrRonconi @Nicochan33 @JoannMoretti @pchamard @Khulood_Almani @TheAIObserverX @TysonLester @Eli_Krumova @amalmerzouk @mikeflache @Analytics_699 @MaryRich78 @NathaliaLeHen @sminaev2015 @WillyRayNick @DanielleLargier @RLDI_Lamy
ALT Figure 1: The Aristotelian Representation Hypothesis: Local relations (“who is near whom”), rather than distances between data points, are preserved across different representation spaces. Representation learning algorithms will converge to shared local neighborhood relationships. 📸 Source: @arxiv
Fabian Gröger retweeted
2-min teaser of my @ELLIS_Amsterdam #FoMo2026 talk.
341 slides. ~1 week to create from analog stage (brainstorming to sticky note storyboarding) to digital (slides).
Mix of my recent research, teaching, history, comedy, and sound effects.
It was fun making and delivering it!
Day 3 of the @ELLISforEurope FOMO School!
We start with Europe-@CSProfKGD of @VectorInst, @YorkUniversity on understanding what's inside the black box.
First: Why should we?
👉Safety. Regulation. Expand human knowledge.
PS:Yours truly reporting from a standing desk today 🫡
Fabian Gröger retweeted
Must-read AI research of the week:
▪️ The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems
▪️ Intelligent AI Delegation
▪️ Discovering Multiagent Learning Algorithms with LLMs
▪️ Towards a Science of AI Agent Reliability
▪️ Frontier AI Risk Management Framework in Practice
▪️ Experiential Reinforcement Learning
▪️ Learning Personalized Agents from Human Feedback
▪️ PANINI: Continual Learning in Token Space via Structured Memory
▪️ Recall Is the Bottleneck for Parametric Factuality
▪️ Reinforced Fast Weights with Next-Sequence Prediction
▪️ On Surprising Effectiveness of Masking Updates in Adaptive Optimizers
▪️ Revisiting the Platonic Representation Hypothesis: An Aristotelian View
▪️ Understanding vs. Generation: Navigating Optimization Dilemma in Multimodal Models
Find the full list here: turingpost.com/p/fod141
