Emile van Krieken

LLMs are nothing more than models of the distribution of the word forms in their training data, with weights modified by post-training to produce somewhat different distributions.

The AI discourse sometimes seems to center on "Is AI good or is it bad?"

I find this framing unproductive. AI is not a fixed thing.

I would prefer to ask "How might we use this technology for good, and mitigate the bad?"

What a shame if the best use we can come up with is no use at all.

Graph Homomorphism Distortion: A Metric to Distinguish Them All and in the Latent Space Bind Them A large driver of the complexity of graph learning is the interplay between structure and features. When analyzing the expressivity of graph neural networks, however, existing approaches ignore featur...

To kick off the PhD journey with @pseudomanifold.topology.rocks:

What are the limitations of the WL metric, and what is an 𝘪𝘯𝘧𝘰𝘳𝘮𝘢𝘵𝘪𝘷𝘦 𝘮𝘦𝘵𝘳𝘪𝘤?

We answer these questions with our 𝗚𝗿𝗮𝗽𝗵 𝗛𝗼𝗺𝗼𝗺𝗼𝗿𝗽𝗵𝗶𝘀𝗺 𝗗𝗶𝘀𝘁𝗼𝗿𝘁𝗶𝗼𝗻

arxiv.org/abs/2511.03068

@olgatticus.bsky.social, Kavir and @erikjbekkers.bsky.social

"He is from [MASK] [MASK]" → "San York"? dLLMs fail because they ignore token dependencies. This Factorization Barrier arises from a structural misspecification: models are restricted to fully factorized outputs. We break this barrier with CoDD, enabling coherent parallel generation. 🚀

Sam is a snake

time traveler from 12 months from now just sent me this

In light of the current funding situation (worldwide), a modest proposal: instead of pouring billions of dollars into GenAI claiming "it *could* accelerate science and research," consider putting 1% of that amount in what *will* accelerate science and research. Namely, funding science and research.

why do science? it won,t make the model Bigger

Storchastic: A Framework for General Stochastic Automatic Differentiation Modelers use automatic differentiation (AD) of computation graphs to implement complex Deep Learning models without defining gradient computations. Stochastic AD extends AD to stochastic computation g...

I spent way too long trying to understand stop gradients lol arxiv.org/abs/2104.00428 (see the first appendix).

I'd argue it is about the loss, but rather you're defining a surrogate loss that should optimise the true loss you're interested in.

X is hiring a creative writing specialist at $40 an hour to make Grok better at writing and a true LOL at the qualifications

New open source: cuthbert 🐛

State space models with all the hotness: (temporally) parallelisable, JAX, Kalman, SMC

Best conference with the best people and in the best place 😎 😜

Also the submission deadline is conveniently one month later than #ICML2026, just in case you needed it 😅

🦕The 20th conference on Neurosymbolic AI will be in Lisbon, Portugal, September 1-4, 2026!

The CFP is out: 2026.nesyconf.org/call-for-pap... with two phases:
🚨 Deadline 1: Feb 24 (abstract), Mar 3 (full)
🚨 Deadline 2: Jun 9 (abstract), Jun 16 (full)

#neurosymbolic #NeSy2026

We introduce epiplexity, a new measure of information that provides a foundation for how to select, generate, or transform data for learning systems. We have been working on this for almost 2 years, and I cannot contain my excitement! arxiv.org/abs/2601.03220 1/7

Good call! I maintain a list of Neurosymbolic folks on Bsky, see here 🦕
go.bsky.app/RMJ8q3i

I am recruiting 1 PhD student (4-year position) and 2 postdocs (3-year positions) to work on logic and machine learning at the University of Helsinki:
- PhD 1: jobs.helsinki.fi/job/Helsinki...
- Postdoc 1: jobs.helsinki.fi/job/Helsinki...
-Postdoc 2: jobs.helsinki.fi/job/Helsinki...

#XAI, #neurosymbolic methods #nesy and #causal #representation #learning #CRL all care about learning #interpretable #concepts, but in different ways.

We are organizing this #ICLR2026 workshop to bring these three communities together and learn from each other 🦾🔥💥

Submission deadline: 30 Jan 2026

Thanks for the fantastic talk, and totally agree! (Writing this in the train from Copenhagen :-))

Emile will present our work on Knowledge Graph Embeddings at Eurips' Salon des Refusés on Friday!
We show how linearity prevent KGEs from scaling to larger graphs + propose a simple solution using a Mixture of Softmaxes (see LLM literature) to break the limitations at a low parameter cost. 🔨

NeSy conference The NeSy conference studies the integration of deep learning and symbolic AI, combining neural network-based statistical machine learning with knowledge representation and reasoning from symbolic appr...

Recordings of the NeSy 2025 keynotes are now available! 🎥

Check out insightful talks from @guyvdb.bsky.social, @tkipf.bsky.social and D McGuinness on our new Youtube channel www.youtube.com/@NeSyconfere...

Topics include using symbolic reasoning for LLM, and object-centric representations!

🚨 New paper alert!
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.

Interested in meeting up in Copenhagen? Do shoot a message!

And finally #3

🔨 Rank bottlenecks in KGEs:

At Friday's "Salon des Refuses" I will present @sbadredd.bsky.social 's new work on how rank bottlenecks limit knowledge graph embeddings
arxiv.org/abs/2506.22271

#2
🍇 GRAPES: At Tuesday's ELLIS Unconference poster session.
We study adaptive graph sampling for scaling GNNs!

Work with Taraneh Younesian, Daniel Daza, @thiviyan.bsky.social, @pbloem.sigmoid.social.ap.brid.gy

arxiv.org/abs/2310.03399

Almost off to @euripsconf.bsky.social in Copenhagen 🇩🇰 🇪🇺! I'll present 3 posters:

🧠 Neurosymbolic Diffusion Models: Thursday's poster session.

Going to NeurIPS? @edoardo-ponti.bsky.social and @nolovedeeplearning.bsky.social will present the paper in San Diego Thu 13:00
arxiv.org/abs/2505.13138

Beyond Smoothed Analysis: Analyzing the Simplex Method by the Book Narrowing the gap between theory and practice is a longstanding goal of the algorithm analysis community. To further progress our understanding of how algorithms work in practice, we propose a new alg...

The simplex algorithm is super efficient. 80 years of experience says it runs in linear time. Nobody can explain _why_ it is so fast.

We invented a new algorithm analysis framework to find out.

Precies hetzelfde hier...

Want to use your favourite #NeSy model but afraid of the reasoning shortcuts?🫣

Fear not💪🏻In our #NeurIPS2025 paper we show that you just need to equip your favourite NeSy model with prototypical networks and the reasoning shortcuts will be a problem of the past!

I'm in Suzhou to present our work on MultiBLiMP, Friday @ 11:45 in the Multilinguality session (A301)!

Come check it out if your interested in multilingual linguistic evaluation of LLMs (there will be parse trees on the slides! There's still use for syntactic structure!)

arxiv.org/abs/2504.02768

🌍Introducing BabyBabelLM: A Multilingual Benchmark of Developmentally Plausible Training Data!

LLMs learn from vastly more data than humans ever experience. BabyLM challenges this paradigm by focusing on developmentally plausible data

We extend this effort to 45 new languages!