Lightnews — Scholar-powered news

Luigi Acerbi @lacerbi.bsky.social · 7d

THATS TEH SPIRIT!!!

2

Luigi Acerbi @lacerbi.bsky.social · 7d

Kudos to fellow co-organizers! @desirivanova.bsky.social
@conorhassan.bsky.social Cen-You (Scott) Li

Luigi Acerbi @lacerbi.bsky.social · 7d

Short abstract submission is open (DL Oct 16, 2025 AoE)
Accepted abstracts will be presented as posters.

🌍 Location: Copenhagen (co-located with #EurIPS)
📅 Date: 2 Dec, 2025
🔗 Website: sites.google.com/view/amortiz...
📝Call for abstracts: sites.google.com/view/amortiz...

Amortized ProbML 2025 - Call for Papers

Accepted papers will be presented as posters during the workshop poster session on 2nd December, 2025. Please be aware that this is not the same as the ELLIS UnConference poster session.

sites.google.com

1 3 1

Luigi Acerbi @lacerbi.bsky.social · 7d

We'll discuss tricky questions such as:
• When does amortization fail badly even though it's rarely talked about?
• Which applications can these neat methods actually work for?
• Can we really deploy these methods in the real world?
• ... and more!

1 1 3

Luigi Acerbi @lacerbi.bsky.social · 7d

Do you like to train neural networks to solve all your nasty probabilistic inference and sequential design problems?
Do you love letter salads such as NPs, PFNs, NPE, SBI, BED?

Then no place is better than the Amortized ProbML workshop we are organizing at #ELLIS UnConference.

Amortized ProbML:
Unlocking Probabilistic Machine Learning for Real-world Inference & Design (🇩🇰)
ELLIS UnConference Workshop
Copenhagen | 2 December, 2025

1 6 17

Reposted by Luigi Acerbi

Alexander Doria @dorialexander.bsky.social · 11d

And new paper out: Pleias 1.0: the First Family of Language Models Trained on Fully Open Data

How we train an open everything model on a new pretraining environment with releasable data (Common Corpus) with an open source framework (Nanotron from HuggingFace).

www.sciencedirect.com/science/arti...

8 51 170

Luigi Acerbi @lacerbi.bsky.social · 22d

Very interesting, looking forward to reading it in detail! For broader context, I recommend looking into the woefully underappreciated literature on (autoregressive) (transformer) neural processes. (Conditional) NPs are probabilistic models predicting data based on arbitrary sets of data.

2

Reposted by Luigi Acerbi

Liu, Shuze @liushuze.bsky.social · Aug 19

@gershbrain.bsky.social and I have a new paper in PLOS Comp Bio!

We study how two cognitive constraints—action consideration set size & policy complexity—interact in context-dependent decision making, and how humans exploit their synergy to reduce behavioral suboptimality.

osf.io/preprints/ps...

OSF

osf.io

1 11 24

Reposted by Luigi Acerbi

Aki Vehtari @avehtari.bsky.social · Aug 13

Posterior predictive checking of binary, categorical and many ordinal models with bar graphs is useless. Even the simplest models without covariates usually have such intercept terms that category specific probabilities are learned perfectly. Can you guess which model, 1 or 2, is misspecifed? 1/4

Useless posterior predictive checking bar graphs for Models 1 and 2

1 6 34

Luigi Acerbi @lacerbi.bsky.social · Jul 27

Remember to follow the official rebuttal guide.

A 26-panel infographic titled “A to Z of How to Write a NeurIPS Rebuttal,” with each panel representing a letter of the alphabet and a corresponding step or emotion. The top-left panels (A to D) show calm, professional cartoon characters with labels like “Acknowledge,” “Be concise,” “Clarify,” and “Dispute.” As the panels progress, the characters grow increasingly stressed, disheveled, and wild-eyed. Midway through, with panels like “Justify choices,” “Keep calm,” and “Obsess,” the illustrations become more chaotic and colorful. Toward the end (letters U to Z), the images turn black and white, depicting characters with frantic, exaggerated expressions, culminating in “Zombie.” The visual style shifts from clean and organized to frenetic and distorted, representing the mental unraveling of a researcher under rebuttal pressure.

1 1 9

Luigi Acerbi @lacerbi.bsky.social · Jul 20

For example for surrogate-based Bayesian inference, our VBMC works with GPs in the low-likelihood evaluation regime:
github.com/acerbilab/py...

Then for many likelihood evaluations we have other methods, like this one with normalizing flows:
arxiv.org/abs/2504.11554

Normalizing Flow Regression for Bayesian Inference with Offline Likelihood Evaluations

Bayesian inference with computationally expensive likelihood evaluations remains a significant challenge in many scientific domains. We propose normalizing flow regression (NFR), a novel offline infer...

arxiv.org

1 6

Luigi Acerbi @lacerbi.bsky.social · Jul 20

Yep!

We are not quite doing it yet in a single existing method -- mostly because it'd be hard to publish --, but definitely thought about it many times.

1 2

Luigi Acerbi @lacerbi.bsky.social · Jul 17

Superlative initiative! Massive kudos to the organisation team for getting this going in such a short time!

EurIPS Conference @euripsconf.bsky.social · Jul 16

EurIPS is coming! 📣 Mark your calendar for Dec. 2-7, 2025 in Copenhagen 📅

EurIPS is a community-organized conference where you can present accepted NeurIPS 2025 papers, endorsed by @neuripsconf.bsky.social and @nordicair.bsky.social and is co-developed by @ellis.eu

eurips.cc

1 5

Luigi Acerbi @lacerbi.bsky.social · Jul 17

I assume it was referring to different authors of multi-author papers. I also read it as boilerplate “we are not competing with NeurIPS, also go to NeurIPS” by making it clear that this is just a satellite and not a full-blown schism (plus there likely isn’t enough space for everyone).

1 3

Luigi Acerbi @lacerbi.bsky.social · Jul 5

But everything transforming into a police state is a trend I guess.

Luigi Acerbi @lacerbi.bsky.social · Jul 5

Definitely not what we should want. I found them in terrible taste and unacceptable. (I received one due to a coauthor who wasn’t even at fault, but the AC had no power to stop these automated threatening emails.)

1 4

Reposted by Luigi Acerbi

Grace Lindsay @neurograce.bsky.social · Jun 30

Holy shit

Robert Rohde @rarohde.bsky.social · Jun 30

Proposed NOAA budget zeros out ALL climate laboratories and cooperative institutions.

GFDL, NSSL, GML, etc.

This appears to also end the US greenhouse gas sampling network, including at Mauna Loa, the oldest continuous carbon dioxide monitoring site on Earth.

www.commerce.gov/sites/defaul...

Excerpt from proposed NOAA budget showing zero funding for Climate Laboratories and Cooperative Institutes

1 5 9

Reposted by Luigi Acerbi

K. Thor Jensen @kthorjensen.bsky.social · Jun 28

The singularity is awesome

Google search for “all you can eat buffet near me”
Al Overview
To find an all-you-can-eat buffet near you, you can use online search engines like Google or Yelp, specifying "all you can eat buffet near me". These searches will typically provide a list of nearby restaurants offering buffet-style dining, including their addresses, contact information, and customer reviews. You can also refine your

280 4.4K 21K

Luigi Acerbi @lacerbi.bsky.social · Jun 27

(Thanks, there are multiple issues with dark mode, fixing it...)

That's a very reasonable concern! It's the old issue of model misspecification/covariate shift/etc. -- still an open problem. One way is to make the training data *bigger* so that it contains "everything", but we are not there yet.

1

Luigi Acerbi @lacerbi.bsky.social · Jun 26

Yep, I remember reading this at some point in the past.

The concept of "amortized X" is pretty old and takes many names and many hats, on top of having being rediscovered by multiple people and multiple fields.

Incidentally, the codename for this project in my lab was "amortize everything"...

1

Luigi Acerbi @lacerbi.bsky.social · Jun 26

Blog post: lacerbi.github.io/blog/2025/ju...

You can just predict the optimum | Luigi Acerbi

instant Bayesian optimization! (kind of)

lacerbi.github.io

1 5

Luigi Acerbi @lacerbi.bsky.social · Jun 26

New blog post!

You can train a neural network to *just do things* -- such as *predict the optimum of a function*. But how do you get a big training dataset of "functions with known optima"?

Read the blog post to find out! (link 👇)

Screenshot of blog post titled "You can just predict the optimum"

3 4 18

Reposted by Luigi Acerbi

Jim Acosta @jimacosta.bsky.social · May 28

If you’re not attending Trump’s military exploitation birthday party in DC you might want to consider joining one of the “No Kings” events happening across the country on June 14. My guest today @ezralevin.bsky.social from Indivisible explains how you can get involved.
youtu.be/iJC455bmkbo

THAT PARADE COSTS HOW MUCH?

YouTube video by Jim Acosta

youtu.be

69 590 2.1K

Reposted by Luigi Acerbi

Max Berger @maxberger.bsky.social · May 28

In anti-authoritarian struggles, criticizing the regime for their blatant corruption is often one of the most important ways to mobilize the public and break through to regime supporters.

The Trump regime is the most corrupt in US history—it’s worth repeating ad nauseam

Catherine Rampell @crampell.bsky.social · May 28

More Than a Dozen U.S. Officials Sold Stocks Before Trump’s Tariffs Sent the Market Plunging www.propublica.org/article/us-o...

More Than a Dozen U.S. Officials Sold Stocks Before Trump’s Tariffs Sent the Market Plunging

Records show well-timed trades by executive branch employees and congressional aides. Even if they had no insider information, ethics experts say such trading undermines faith in government and the ma...

www.propublica.org

18 500 1.3K

Luigi Acerbi @lacerbi.bsky.social · May 26

For fairness -- it's not a complete lack, a few people are trying (and I scarcely contribute myself), but unfortunately most of the discourse is carried out elsewhere.

2