Raphaël Millière

Transformer-based neural networks achieve impressive performance on coding, math & reasoning tasks that require keeping track of variables and their values. But how can they do that without explicit memory? 📄 Our new ICML paper investigates this in a synthetic setting! 🧵 1/13

What is the interplay between representations learned from (language) surface forms alone, and those learned from more grounded evidence (e.g.,vision)? Excited to share new work understanding “Cross-modal taxonomic generalization” in (V)LMs 1/

My thoughts on connectomics and upload: 1) there is zero question connectomes are invaluable, and we need to get them for mouse, monkey, and human 2) the human, or even monkey, connectome seems a long ways off given costs (roughly $1/neuron). The projectome (map of all the axons) seems eminently reachable and should be a top priority imho 3) but even having the full connectome would only tell you numbers of synapses, not actual synaptic weights, and the two can be hugely divergent (eg only 5% of synapses onto V1 layer 4 neurons come from thalamus, even though this is the major driving input) 4) given #2 & #3, I think we can get to upload in the sense of building a functionally equivalent organism much faster through understanding the algorithms of the primate brain than through blind copying 5) in putting together something as complex as the human brain we would definitely want to check that the various pieces work as we go, which we can only do if we understand these pieces 6) I don't think upload in the sense of blindly creating a digital copy is the path to the abundant transhumanist future--actual understanding of brain structures so we can intelligently interface with them, and emulate their function in code without copying all the details, is. All to say, we need functional understanding to go hand in hand with anatomical mapping!

Can large language models *introspect*? In a new paper, @kmahowald and I study the MECHANISM of introspection in big open-source models. tldr: Models detect internal anomalies through DIRECT ACCESS, but don't know what the anomalies are. And they love to guess “apple” 🍎

@dioscuri Surely you mean qwen 3.5 27B!

In this country, first you get the money, then you get the RAM, then you get the Qwen2.5-72B

@AndrewLampinen Good luck with the next steps!

After 5.5 years (or 7 or 9, counting internships), today was my last day at Google/DeepMind. When I was in London recently, I walked through the two floors that were (most of) DeepMind when I first joined, and thought about how much the company and field have changed since then.

1/ 👶 BabyLM is back at EMNLP 2026! We are excited to announce that the 4th BabyLM Challenge & Workshop will once again bring together researchers interested in sample-efficient, developmentally plausible language modeling. @emnlpmeeting More in🧵

fyi, @babyLMchallenge has been doing this for 4 years now. some interesting ideas from our past competitions for folks to consider: 1. mixing causal and masked LM objectives (GPT-BERT) 2. mixture of experts as a way to better model human cognition

#VideoReason We are open-sourcing the entire VBVR stack to speed-up the arrival of video reasoning as the next fundamental paradigm of intelligence - 150+ synthetic generators - 1 million training clips - Cloud-scale data factory - Unified EvalKit - 100 rule-based evaluators - Strong baseline model Checkout at video-reason.com

I speculated about this back in 2024, but it's nice to finally have some hard mechanistic evidence about video models to adjudicate bombastic claims about their ability to "simulate physics" artificialcognition.net/posts/video-ge…

Great work! "These findings suggest that modern video models do not use factorized representations of physical variables like a classical physics engine. Instead, they use a distributed representation that is nonetheless sufficient for making physical predictions."

How does the brain know which neurons to adjust during learning in order to optimize behavior? MIT researchers discovered that brains can use cell-by-cell error signals to do this — surprisingly similar to how AI systems are trained via backpropagation. mcgovern.mit.edu/2026/02/25/neu…

@dileeplearning Congrats Dileep! Looking forward to seeing what comes out of this

@benno_krojer Very cool!

You can now "pip install latentlens" 🔨 It comes with: * pre-computed embeddings for several popular LLMs and VLMs * a txt file with sentences describing WordNet concepts, which we recommend as a standard corpus to get embeddings from * ... Try it out and let us know what we can improve!

@peligrietzer @littmath my original point is that huge autoregressive transformers trained on lots of data aren't necessary and perhaps not even straightforwardly sufficient (?) for tasks like pro-level go-playing, though I'm not sure what that says about how good they can get at advanced maths

@peligrietzer @littmath Right - my mental model is that RL helps by concentrating LLMs' probability mass on good/long reasoning trajectories, but they're still doing inherently serial computation unless you graft on "Deep Think"-like parallel sampling strategies so not sure how far to take this analogy

Compared to @littmath I'm more open to the possibility of a 'glass fence' for RL-LLM performance. Could be that just like GOFAI is good for Chess-complexity tasks but not for Go-complexity tasks, big data neural nets are good for Go-complexity tasks but not for all tasks

LUNAR Lab is looking for a postdoc to work on understanding and interpreting reasoning in LLMs and humans, broadly construed. The position is funded by Schmidt Sciences and is for at least 18 months, with the likely option to extend. Apply here! forms.gle/CrmrCzun79G9Ca…

The humanities are vital for helping us navigate a world with increasingly powerful AI. A nice piece by Agnieszka Piotrowska in @timeshighered timeshighereducation.com/opinion/humani…

@apaphilosophy Here's the paper: link.springer.com/article/10.100… Make sure to check out the rest of the fantastic special issue put together by @cdkg_phil & @hendrycks

I'm honored to receive this prize from @apaphilosophy! I look forward to seeing more philosophical work on safety issues arising with frontier AI rather than hypothetical future systems. Philosophers have much to contribute to both normative & technical aspects of alignment.