Marvin Schmitt

😀 Enter the amortized Bayesian workflow from here 👉 pipme.github.io/amortized-Ba... Huge thanks to my great collaborators: @avehtari.bsky.social, @paulbuerkner.com, @stefanradev.bsky.social, @lacerbi.bsky.social, @marvin-schmitt.com!

1/ Wouldn't it be nice if you could perform Bayesian inference *efficiently* but also *reliably*? Amortized inference offers the former, while MCMC is often presented as the "gold standard" for accuracy and reliability. Enter the Amortized Bayesian Workflow...

Posterior-SBC now also with peer-review stamp in Statistics and Computing doi.org/10.1007/s112... (update your bib files)

I took a Waymo today. I had some initial safety concerns, but I felt reassured when I saw the driver had their seat belt fastened.

Sunrise in San Francisco ☀️

ICLR decisions are out! #ICLR #ICLR2026 @iclr-conf.bsky.social

Autumn in Helsinki‘s urban woodlands 🇫🇮❤️

Is this the real life? Is this just fantasy?

Many treat uncertainty = a number. At Apple, we're rethinking this: LLMs should output strings that reveal all information of their internal distributions. We find that Reasoning, SFT, CoT can't do it - yet. To get there, we introduce the SelfReflect benchmark. arxiv.org/pdf/2505.20295

Natural idea. Looks like a nice paper too. arxiv.org/abs/2508.21184

BayesFlow released version 2.0.4, presented numerous findings at the MathPsych/ICCM 2025 conference at Ohio State University, and expanded its contributor list to 25 active members! Congrats to BayesFlow on all these new huge accomplishments!

This is a fantastic European initiative to offer an alternative attendance option to the largest AI conference 🇪🇺✨

The S in Sunday stands for Sinnamon buns; sourdough-based of course 🧡

Extending my artisanal repertoire: Homemade pickled red onions 💜

My 5-year-old asked me whether there are "things that only have one side". Me: *takes a deep breath* Actually …

This is a fantastic read about Cameron‘s work in developer advocacy. The post is not only valuable for dedicated devrel folks, but also has many insights for all open source developers. Check it out 👇

We’re going Italian! 🇮🇹 Sourdough focaccia with rosemary and cherry tomatoes 🌿🍅

I'm putting together a visualization workshop for PhD students 🧪📊 Looking for examples of the good, the bad, and the ugly. Do you have examples for a great (or awful) figure? Plots and overview/explainer figures are welcome. Thanks 🧡

Back to a basic wheat bread. I‘ve been told that it pairs well with chocolate spread 👀

Today we’re serving a traditional Finnish rye bread with crispy crust and fine crumb 🇫🇮

For my next trick… Traditional rye/wheat mix bread ✨

My first attempt at sourdough bread went surprisingly well. Any tips or advice from the bread baking pros on here?

🧠 Check out the classic examples from Bayesian Cognitive Modeling: A Practical Course (Lee & Wagenmakers, 2013), translated into step-by-step tutorials with BayesFlow! Interactive version: kucharssim.github.io/bayesflow-co... PDF: osf.io/preprints/ps...

Because it wasn't annoying enough that my domain was down for 35 days... Now a bot network farm automatically bought my old personal domain the second it got released. Odds are that they will (1) host explicit+spam content on the website and/or (2) try to sell it back to me for an ungodly price.

Clouded mind. (created with #midjourney)

If you’re into bluesy tunes and parody lyrics, check out our cover song! 🎶 There ain’t no cure for the PhD blues 🎶

My domain is down. In the meantime, you can find my website at: marvinschmitt.github.io

With his research on Bayesian Deep Learning, Vincent has his finger on the pulse of the time in modern AI 🤖 I’m already looking forward to tune in to this conversation between @alex-andorra.bsky.social and @vincefort.bsky.social. … and it also comes with a video version! 👇 youtu.be/3hYYGiucS0U

Any advice on how to design a pretty cover for a PhD thesis? 📔 #AcademicSky

Fantastic initiative by @serge.belongie.com and Søren Hauberg 🇪🇺🇩🇰 Please take a moment to answer through the poll below and share in your networks! 👇

Official press release from the website of the Finnish Research Council: www.aka.fi/en/about-us/...

Making some final adjustments to my strategy for tomorrow's PhD defense. As always, @xkcd.com is very helpful as inspiration ⚔️

Introducing our Gemma 3 open models, the most capable models that you can run on a single GPU or TPU. Multimodal, multilingual, 128k context length, and exceeds quality of other open models that are an order of magnitude larger in terms of hardware footprint. 🎉 blog.google/technology/d...

This International Women's Day, we're challenging you to #AccelerateAction to get more women & girls in AI and machine learning. It's your turn! Reply with a suggestion for action or tag an inspiring female scientist. #IWD2025 #AccelerateAction #WomenInELLIS

New paper Säilynoja, Johnson, Martin, and Vehtari, "Recommendations for visual predictive checks in Bayesian workflow" teemusailynoja.github.io/visual-predi... (also arxiv.org/abs/2503.01509)

I can highly recommend this CodeRefinery workshop on reproducible research tooling 🔨 Last year, I took one of their 3-day workshop about Slurm HPC clusters. It was well-planned, beautifully executed, and I learned *a lot*. Check out the details 👇

It’s great to see that European AI is getting such tailwind on a political level 🇪🇺 InvestAI is a huge endorsement of the ELLIS mission to position Europe at the frontier of AI research which is sustainable on many levels.

Amortized inference for finite mixture models ✨ The amortized approximator from BayesFlow closely matches the results of expensive-but-trustworthy HMC with Stan. Check out the preprint and code by @kucharssim.bsky.social and @paulbuerkner.com👇

1/3 IMO, this method gets back to a more common way of testing parameter recovery. Back in the day, I would typically fit my model to data, generate posterior predictions using the fitted posterior means, then re-fit the model to the posterior predictions to see if I can recover those means well.

Join us this Thursday for a talk on efficient mixture and multilevel models with neural networks by @paulbuerkner.com at the new @approxbayesseminar.bsky.social!

A reminder of our talk this Thursday (30th Jan), at 11am GMT. Paul Bürkner (TU Dortmund University), will talk about "Amortized Mixture and Multilevel Models". Sign up at listserv.csv.warwick... to receive the link.

Bridging LLMs and uncertainty quantification is one of the most interesting open problems in modern deep learning. Follow Michael‘s work for some bleeding-edge research on this topic 👀👇

This upcoming ICLR paper solves inverse problems in imaging with a creative combination of flow matching and the plug-and-play framework. Check out Paul‘s explainer thread 👇🧵

I’ve been using Scholar Inbox for a while now and I absolutely love it. If you don’t use it yet, check it out now! 👉 www.scholar-inbox.com

Exciting new partnership in AI & Life Sciences announced today in Brussels

Paul Bürkner (@paulbuerkner.com) will talk about amortized Bayesian multilevel models in the next Approximate Bayes Seminar on January 30 ⭐️ Sign up to the seminar’s mailing list below to get the meeting link 👇