Luigi Acerbi

Yes, this still happens! Before doing a manual check, I deploy agents to do the first check for me (literally a /doublecheck skill in Claude Code which I force to call automatically at the end of each task). This often catches issues, including once a totally fabricated MCMC analysis...

Look who's there!

With (in pseudo-random order) @mummitrollet.bsky.social @nasrullohloka.bsky.social @huangdaolang.bsky.social @conorhassan.bsky.social @sfrancesco.bsky.social @arnosolin.bsky.social @samikaski.bsky.social and several others not on here -- check their names above!

More info soon; all papers fit our AI4science research agenda within the Finnish Center for Artificial Intelligence & @ellisinstitute.fi -- building efficient methods for inference, uncertainty quantification and decision making, leveraging powerful autoregressive transformers and diffusion models.

A bit of a delayed celebration, but happy that our three submitted papers were accepted at @iclr-conf.bsky.social 2026! This was a... complicated year for ICLR, but hopefully now we can focus on the science.

This game was 100% designed, made, and tested by Claude Code with one prompt to "make a complete Sierra-style adventure game with EGA-like graphics and text parser, with 10-15 minutes of gameplay." I gave two prompts to play test the game & deploy it.

Play: enchanted-lighthouse-game.netlify.app

GitHub - acerbilab/svbmc: Stacking Variational Bayesian Monte Carlo (S-VBMC) algorithm for combining Variational Bayesian Monte Carlo (VBMC) posteriors to boost inference performance. Stacking Variational Bayesian Monte Carlo (S-VBMC) algorithm for combining Variational Bayesian Monte Carlo (VBMC) posteriors to boost inference performance. - acerbilab/svbmc

7/ Work by @sfrancesco.bsky.social, @chengkunli.bsky.social & myself, with many thanks to the Research Council of Finland.

The S-VBMC code is available as an easy-to-use Python library: github.com/acerbilab/sv...

Check out the paper: openreview.net/forum?id=M2i...

6/ S-VBMC is an inexpensive post-processing step, so it greatly improves posterior quality at a negligible computational cost!

5/ This optimization is made possible by VBMC’s handy property of providing a closed-form solution for individual components of the ELBO (I_m,k), allowing the following formulation for M independent VBMC solutions:

4/ S-VBMC “stacks” the Gaussian mixtures posteriors output by independent VBMC runs by maximizing the “stacked” ELBO with respect to the weights of the individual Gaussian components. It doesn’t change the components, it just re-weights them!

VBMC vs. S-VBMC

3/ However, VBMC’s relatively conservative active learning strategy can lead it to miss some portions of the true posterior when this has challenging properties (multiple modes, long tails). S-VBMC fixes this!

GitHub - acerbilab/pyvbmc: PyVBMC: Variational Bayesian Monte Carlo algorithm for posterior and model inference in Python PyVBMC: Variational Bayesian Monte Carlo algorithm for posterior and model inference in Python - acerbilab/pyvbmc

2/ Bayesian inference of model parameters can be a complex problem to solve, especially with expensive likelihood functions. We addressed this in the past with Variational Bayesian Monte Carlo (VBMC repo: github.com/acerbilab/py...).

Stacking Variational Bayesian Monte Carlo paper in TMLR.

1/ Excited to share our new work published in Transactions on Machine Learning Research (TMLR), Stacking Variational Bayesian Monte Carlo (S-VBMC)!

I'd like to propose the following norm for peer review of papers. If a paper shows clear signs of LLM-generated errors that were not detected by the author, the paper should be immediately rejected. My reasoning: 1/ #ResearchIntegrity

Diffusion Models in Simulation-Based Inference: A Tutorial Review Diffusion models have recently emerged as powerful learners for simulation-based inference (SBI), enabling fast and accurate estimation of latent parameters from simulated and real data. Their score-b...

What an amazing Yule gift from @stefanradev.bsky.social & colleagues: a tour-de-force tutorial on diffusion models for simulator-based inference.

This is one of the most comprehensive and useful review/tutorials I have ever seen -- a must read! Kudos to all the authors!

arxiv.org/abs/2512.20685

EurIPS was pretty fun! Thanks to @lacerbi.bsky.social , @conorhassan.bsky.social , and all organizers of the Amortized ProbML team for the great workshop and for inviting me as a speaker! 🙌 Also great to be able to present our spotlight paper on sparsity on the main stage 😍

Tobias Niehues @tobnie.bsky.social is presenting our work with @dominikstrb.bsky.social 'Amortized Bayesian decision-making for inferring decision-making parameters from behavior' at the Amortized ProbML Workshop and the @ellis.eu UnConference. Please come by our poster!

4/ If you are at NeurIPS, go talk to @huangdaolang.bsky.social at the poster session today!

Joint work with Daolang, @wenxinyi.bsky.social , @ayushbharti.bsky.social & @samikaski.bsky.social, FCAI & @ellisinstitute.fi

Website: huangdaolang.com/aline/

3/ Information gain would normally be hard to estimate, but this real synergy makes training work like a charm and surprisingly well.

Since we like to make everything flexible, you can also pick & choose your information gain targets (data or subsets of parameters) at runtime.

2/ Why *both* inference and design? Not just because we *can*, but because we *must* (h/t @paulbuerkner.com for this catchphrase).

We warm-start the network first by training the inference head via MLE and use that to compute the information gain required to train the policy head.

1/ Very happy with this spotlight paper at
@neuripsconf.bsky.social where we continue our "Amortize Everything" agenda.

After the Amortized Conditioning Engine (Chang et al., AISTATS, 2025) which amortizes all sorts of inference tasks, here with ALINE we amortize both inference & design.

I’m super excited to present our new work in #Eurips2025 and #Neurips2025! We developed FNOPE: a new simulation-based inference (SBI) method which excels at inferring function-valued parameters!

Paper: openreview.net/forum?id=yB5...
Code: github.com/mackelab/fnope
(1/9)

Are there non-predatory for-profit publishers? This "thing" clearly never went through peer review despite what they say.

Well, maybe it did go through "peer review": LLM reviewers reviewed an LLM-written paper, fair enough.

Updated website: sites.google.com/view/amortiz...

Co-organized with Cen-You (Scott) Li, @conorhassan.bsky.social, @desirivanova.bsky.social as well as @samikaski.bsky.social & FCAI

Amortized ProbML workshop page snapshot

If you are at ELLIS UnConference on Dec 2 (pre
@euripsconf.bsky.social ), join us at our Amortized ProbML workshop!

Come for our great speakers: @paulbuerkner.com Yingzhen Li @davidruegamer.bsky.social @liza-semenova.bsky.social & poster session; stay for our hot takes on amortized probML!

I was thinking of sharing them publicly on my blog but then realized it could be problematic / misunderstood / misused, so for now I refrained from it... but yes I can send them privately. I'll email you later.

We regularly use LLMs to review our own papers before submission. I agree that in principle it's hard to "prove" that a review is entirely LLM-generated vs. LLM-written... but in practice it's quite easy especially when you see multiple "LLM-smelling" points in a review.

AISTATS adopted the iron fist against "LLM-generated" reviews and I fully agree with it.

Hopefully we can figure out a way to allow responsible, "LLM-assisted" reviews where one genuinely uses the LLM to assist (not replace) the reviewer, which can be a force multiplier.

Even just the fact that we have oral tests (throughout school and then university) shocks a lot of people.

It's easy to fall into the fallacy "I went through this system so it is obviously the best system" and surely it has downsides but tbh it does prepare you for a lot of things in life...

Does Object Binding Naturally Emerge in Large Pretrained Vision Transformers? Object binding, the brain's ability to bind the many features that collectively represent an object into a coherent whole, is central to human cognition. It groups low-level perceptual features into h...

Very interesting work on emerging object-binding representations in vision transformers by @kordinglab.bsky.social arxiv.org/abs/2510.24709

This might seem an oddly specific property but the good ol' binding problem reflects a fundamental primitive of cognition, epistemology, you name it.