News 🎉 We’re thrilled to announce our final panelist: David Silver! Don’t miss David and our amazing lineup of speakers—submit your latest RL work to our NeurIPS workshop. 📅 Extended deadline: Sept 2 (AoE)

We've extended the deadline for our workshop's calls for papers/ideas! Submit your work by August 29 AoE. Instructions on the website: arlet-workshop.github.io/neurips2025/...

The OpenReview link for our calls (for papers and ideas) is available, submit here: openreview.net/group?id=Neu... We look forward to receiving your submissions!

last year's edition was so much fun I'm really looking forward to this one!! join us in San Diego :))

Join us for Nneka's presentation tomorrow! Last talk before the summer break.

Join us tomorrow for Dave's talk! He will present his recent work on randomised exploration, which received an outstanding paper award at ALT 2025 earlier this year.

new preprint with the amazing @lviano.bsky.social and @neu-rips.bsky.social on offline imitation learning! learned a lot :) when the expert is hard to represent but the environment is simple, estimating a Q-value rather than the expert directly may be beneficial. lots of open questions left though!

🚨 New paper accepted at SIMODS! 🚨 “Nonlinear Meta-learning Can Guarantee Faster Rates” arxiv.org/abs/2307.10870 When does meta learning work? Spoiler: generalise to new tasks by overfitting on your training tasks! Here is why: 🧵👇

Dhruv Rohatgi will be giving a lecture on our recent work on comp-stat tradeoffs in next-token prediction at the RL Theory virtual seminar series (rl-theory.bsky.social) tomorrow at 2pm EST! Should be a fun talk---come check it out!!

new work on computing distances between stochastic processes ***based on sample paths only***! we can now: - learn distances between Markov chains - extract "encoder-decoder" pairs for representation learning - with sample- and computational-complexity guarantees read on for some quick details.. 1/n

A new blog post with intuitions behind continuous-time Markov chains, a building block of diffusion language models, like @inceptionlabs.bsky.social's Mercury and Gemini Diffusion. This post touches on different ways of looking at Markov chains, connections to point processes, and more.

Mattes Mollenhauer, Nicole M\"ucke, Dimitri Meunier, Arthur Gretton: Regularized least squares learning with heavy-tailed noise is minimax optimal https://arxiv.org/abs/2505.14214 https://arxiv.org/pdf/2505.14214 https://arxiv.org/html/2505.14214

Later today, Sikata and Marcel will talk about their recent work on oracle-efficient RL with ensembles. Join us!

Excited to share what I've been up to: bringing text diffusion to Gemini! Diffusion models are _fast_, and hold immense promise to challenge autoregressive models as the de facto standard for language modeling.

The tutorials, workshops, and community events for #COLT2025 have been announced! Exciting topics, and impressive slate of speakers and events, on June 30! The workshops have calls for contributions (⏰ May 16, 19, and 25): check them out! learningtheory.org/colt2025/ind...

Announcing the first workshop on Foundations of Post-Training (FoPT) at COLT 2025! 📝 Soliciting abstracts/posters exploring theoretical & practical aspects of post-training and RL with language models! 🗓️ Deadline: May 19, 2025

looking forward to giving a talk at the 2025 GHOST day in the beautiful city of Poznan (Poland) --- hope to see you there this weekend!

Is Best-of-N really the best we can do for language model inference? New paper (appearing at ICML) led by the amazing Audrey Huang (ahahaudrey.bsky.social) with Adam Block, Qinghua Liu, Nan Jiang, and Akshay Krishnamurthy (akshaykr.bsky.social). 1/11

Spectral Representation for Causal Estimation with Hidden Confounders at #AISTATS2025 A spectral method for causal effect estimation with hidden confounders, for instrumental variable and proxy causal learning arxiv.org/abs/2407.10448 Haotian Sun, @antoine-mln.bsky.social, Tongzheng Ren, Bo Dai

We'll be presenting our work on Oracle-Efficient Reinforcement Learning for Max Value Ensembles at the RL theory seminar! Been following this series for a while, super excited we get to present some of our work. 🥳

new paper online: "CONFIDENCE SEQUENCES FOR GENERALIZED LINEAR MODELS VIA REGRET ANALYSIS" TL;DR: we reduce the problem of designing tight confidence sets for statistical models to proving the existence of small regret bounds in an online prediction game read on for a quick thread 👀👀👀 1/

Last seminars before the summer break: 04/29: Max Simchowitz (CMU) 05/06: Jeongyeol Kwon (Univ. of Widsconsin-Madison) 05/20: Sikata Sengupta & Marcel Hussing (Univ. of Pennsylvania) 05/27: Dhruv Rohatgi (MIT) 06/03: David Janz (Univ. of Oxford) 06/10: Nneka Okolo (MIT)

We are organising EWRL in Tübingen in September. We are working on the program and we will soon be able to make first announcements. Stay tuned, follow @ewrl18.bsky.social

Mark your calendars, EWRL is coming to Tübingen! 📅 When? September 17-19, 2025. More news to come soon, stay tuned!

Tim is amazing! go apply 🤩

Last seminar before we take a short break to prepare for the next season! You can send us paper recommendations via this form: forms.gle/Uak8LpVPynNM...

There is a new RL theory seminar tomorrow ! Join us :)

I’ve put together a short list of opportunities for early career academics willing to come to Europe: www.cvernade.com/miscellaneou... This mostly covers France and Germany for now but I’m willing to extend it. I build on @ellis.eu resources and my own knowledge of these systems.

Another week, another super nice RL theory talk! This time with David

oh cool news in the red there!

www.youtube.com/watch?v=9_Pe... An interview with Rich. The humility of Rich is truly inspiring: "There are no authorities in science". I wish people would listen and live by this.

Today at 6 PM UTC, Dhruv Rohatgi will give a presentation at the RL seminars. Hope to see many of you!

Video now public from the talk "Learning to act in noisy contexts using deep proxy learning" at the NeurIPS'24 Workshop on Causal Representation Learning! Video: neurips.cc/virtual/2024... Slides: www.gatsby.ucl.ac.uk/~gretton/cou...

another week, another LLM and, yes, another RL seminar. join us at 6 PM UTC today!

super happy about this preprint! we can *finally* perform efficient exploration and find near-optimal stationary policies in infinite-horizon linear MDPs, and even use it for imitation learning :) working with @neu-rips.bsky.social and @lviano.bsky.social on this was so much fun!!

can Grok 3 prove a lower bound on the bayesian regret though? didn't think so... but Itai can! today at 6 PM UTC

Most of the talk around AI and energy use refer to an older 2020 estimate of GPT-3 energy consumption, but a more recent paper directly measures energy use of Llama 65B as 3-4 joules per decoded token. So an hour of streaming Netflix is equivalent to 70-90,000 65B tokens. arxiv.org/pdf/2310.03003

just realized 2025 is the year when the Adam paper gets a test of time award and maybe TRPO? what else?

Hi folks! I'm excited to be on BlueSky! I'm looking forward to posting about computer science research, ML, scientific advances, tasty food, nature, and making groan-worthy puns.

We’re back! Join us for the next talks 🤓

*PLS SHARE* Open position for an *Associate Professor* in Machine Learning at our department (@enginyeria-upf.bsky.social / @upf.edu), via the Serra Hunter programme. DEADLINE: January 13th 2025 www.upf.edu/web/personal...

I'm looking for an emergency reviewer or two for an AISTATS submission in the area of Bayesian ML theory, which at present has the misfortune of three low-quality reviews. If you're able to help, please DM and let me know! Please link your Google Scholar if I don't already know you personally.

Yesterday was my last day at Google DeepMind. Interning there when the first GDM Nobel prize, AlphaProof, the Gemini releases and more all happened and having such amazing/ambitious colleagues was quite humbling and really exciting. I look forward to 2025!

Disentanglement is an intriguing phenomenon that arises in generative latent variable models for reasons that are not fully understood. If you’re interested in learning why, I highly recommend giving Carl’s blog a read!

The slides for my lectures on (Bayesian) Active Learning, Information Theory, and Uncertainty are online now 🥳 They cover quite a bit from basic information theory to some recent papers: blackhc.github.io/balitu/ and I'll try to add proper course notes over time 🤗

Found slides by Ankur Moitra (presented at a TCS For All event) on "How to do theoretical research." Full of great advice! My favourite: "Find the easiest problem you can't solve. The more embarrassing, the better!" Slides: drive.google.com/file/d/15VaT... TCS For all: sigact.org/tcsforall/

Very curious about this one! Join us tomorrow 🤓

Hello Blue Sky! Let's see what you are good for.

🚨 New Paper 🚨 Can LLMs perform latent multi-hop reasoning without exploiting shortcuts? We find the answer is yes – they can recall and compose facts not seen together in training or guessing the answer, but success greatly depends on the type of the bridge entity (80% for country, 6% for year)! 1/N

Then a consequence of Sandroni's theorem (bsky.app/profile/aaro...) is that there is no empirical test that can distinguish this scenario from the alternative scenario in which the forecaster actually knows the outcome distribution at each round and correctly forecasts the most likely outcome. 2/2