Why does L1 regularization induce sparse models? Many illustrate this using the least squares problem with a norm constraint. The least squares level sets are drawn next to the different unit "circles". I prepared a cool animation which I believe makes it even clearer 🙂

Last-Iterate Convergence of Randomized Kaczmarz and SGD with Greedy Step Size Michał Dereziński, Xiaoyu Dong arxiv.org/abs/2604.09909 [𝚌𝚜.𝙻𝙶 𝚖𝚊𝚝𝚑.𝙽𝙰 𝚖𝚊𝚝𝚑.𝙾𝙲 𝚜𝚝𝚊𝚝.𝙼𝙻]

If the NeurIPS paper checklist isn't a good motivation to do entirely theoretical work, I don't know what is.

Will be speaking today at the GRaM workshop at ICLR @iclr_conf (gram-workshop.Github.io/ ) about how we are building efficient and effective billion-scale Graph Foundation Models at Meta!!! 🤩 (GraphBFF 👉 lnkd.in/dB67TZN6). Anddddd we also have an awesome paper at the workshop! "Improving LLM Predictions via Inter-Layer Structural Encoders", poster session A :) (lnkd.in/dZCnztY5) #iclr2026

New blogpost out 📃 "Detecting LLM Misbehaviors from the Inside Out with Deep Learning on Structured Data" (ffabffrasca.substack.com/p/d…) [1/8]

How does an LLM’s past influence its future?🤔 In our new paper with @FazlBarez,@mtutek,@boknilev, Shay Cohen, we show that conversational history creates a "geometric trap" in the latent space, confining the model’s trajectory➡️making old habits e.g. hallucinations hard to break

🚨 2026 @Princeton ML Theory Summer School Mini-courses by: - Subhabrata Sen @subhabratasen90 - Lenaic Chizat @LenaicChizat - Sinho Chewi - Elliot Paquette @poseypaquet - Elad Hazan @HazanPrinceton - Surya Ganguli @SuryaGanguli August 3 - 14, 2026 Apply by March 31. Link 👇 Sponsors: @NSF, @PrincetonAInews, @EPrinceton @JaneStreetGroup, @DARPA, @PrincetonPLI, Princeton NAM, Princeton AI2, Princeton PACM Some amazing speakers from previous years: @Andrea__M, @TheodorMisiakie, @KrzakalaF, @_brloureiro, @rakhlin, @DimaKrotov, @CPehlevan, @SoledadVillar5, @SebastienBubeck, @tengyuma

Please rewrite your proof, fully rigorously. - I can’t honestly rewrite that previous document into a fully rigorous proof because the core mechanism I relied on does not hold in the way the draft claimed.

1/6 🧵 Calibration is hard. Multicalibration—fixing errors across every possible subgroup—is usually impossible at scale. Until now. Introducing MCGrad: A production-ready multicalibration library from Meta, accepted at KDD 2026. 🚀 github.com/facebookincubator…

A glimpse into the research I’ve been leading over the past year at Meta 🥹 . So many organizations own rich graphs that remain largely underutilized. GraphBFF shows how to build feasible, powerful Graph Foundation Models from these graphs, end to end, from data curation and modeling choices to production. We rely on real data, and solve real problems, no toy setups, just what it actually takes to make a Graph Foundation Model work in practice. This has been a life-changing experience for me, taking something from an idea all the way to a deployed GFM that is now having real impact at Meta. The preprint is now available on arxiv.

An LLM hallucinating a citation with your name on it is just the universe telling you to write that paper

📌 [1/4] A Graph Meta-Network for Learning on Kolmogorov-Arnold Networks We introduce a weight-space model for KANs, where learning happens directly over the KANs' 1D functions. This work was done during my Meta internship. openreview.net/pdf?id=ONpyYa…

We are back. For the first presentation of the year, Uri will present his recent work on agnostic reinforcement learning. See you next Tuesday!

In memory of my father, an educator who loved the Mishna. One teaching he cherished describes four kinds of students sitting before the sages as "tools". Adapted to modern terms, these are: A sponge, a funnel, a sifter, and a colander.

Remember our ICML25 "Graph Learning Will Lose Relevance Due To Poor Benchmarks"? Fear no more! GraphBench is here! 🤩 We give you: The next generation of Graph Benchmarking! Including: -New shiny high-quality datasets from diverse domains spanning seven domains, including chip design, algorithmic reasoning, and weather forecasting. -Standardized hyperparameter tuning procedures, enabling fair and principled model comparison - Strong, transparent baselines that accurately reflect algorithmic progress - Comprehensive coverage of graph learning tasks, datasets, and modern GNN architectures - Reproducibility-focused design, minimizing variance and evaluation artifacts - Forward-looking benchmark designed for next-generation graph learning research A huge collab with: @chrsmrrs, @mmbronstein, @michael_galkin, @HolgerHoo, Timo Stoll, @ChendiQian, @benfinkelshtein, Ali Parvis, Darius Weber, @ffabffrasca, @HadarShavit, @antoinesrdin, Arman Mielke, Marie Anastacio, Erik Müller,

Since linear probes are popular again, maybe it’s a good time to point to the many issues with them, which were examined in detail in the NLP Interpretability community. The “mechanistic?” piece by @sarahwiegreffe and @nsaphra has many useful pointers. aclanthology.org/2024.blackb…

Accelerate your transformer model with the new Block-Sparse-Flash-Attention! github.com/Danielohayon/Bloc… This training-free, drop-in replacement extends FlashAttention-2 with minimal code changes (CUDA Kernels Included). Paper: arxiv.org/abs/2512.07011

NeurIPS 2025 papers per 1 Million People 1. Singapore – 64.51 2. Switzerland – 22.13 3. Israel – 11.17 4. UAE – 9.47 5. UK – 7.50 6. US – 7.44 7. Denmark – 7.37 8. Australia – 7.31 9. Canada – 6.93 10. South Korea – 5.78

Sure, they overlap, but continual learning (a.k.a. lifelong learning) has been a distinct research area for about a decade now (and even earlier). One key difference is the focus metric: OL: regret, i.e., performance on the *next* task CL: forgetting of *previous* expertise