Christian Wolf

There should be more motion blur

Science in 2026

#CVPR2027 author instructions are out: cvpr.thecvf.com/Conferences/... No paper id in the pdf. Perhaps a direct consequence of the multiple (*sigh*) recent OpenReview leaks? This will avoid modifying the pdfs after submission in case a remap to new IDs is needed.

Are there other examples in history where a new technology completely and quickly upended work and procedures, in the same way as LLMs have destroyed publishing papers?

OMG Many ICLR papers have simulated placeholder values in their experimental results. researchonresearch.blog/2026/10/20/s...

Small scale research was always in danger once deep learning took off, but at least theory survived for some time. Now, the only research (in CV/ML) surviving is Very large scale Involves humans Thoroughly tested w real robots A paper with results which my Claude can give me now has less value.

🚨 New work with Adrien Ramanana, @nicolasdufour.bsky.social, Yutong Bai and Patrick Pérez collab @kyutai-labs.bsky.social CADRE: Rethinking Video Autoencoders for Interactive World Models We investigate what makes a good world model codec 📜 arxiv.org/abs/2610.30129 🖥️ kyutai.org/blog/2026-10... 🧵👇

The reason I don't like the Villeneuve Dune movies is that each and every moment of the movies carries a heavy tone indicating that it is "important" which is normally only given to outstanding moments. If everything is important, nothing is. Claude written text feels the same way.

The new ELLIS Unit in Grenoble will celebrate its kick-off on 8 December! Join them for scientific talks featuring ELLIS President Serge Belongie and Mirco Ravanelli from Concordia University. Participation is free! Programme & registration: https://bit.ly/4zC71LA

Who wins in a Tesla - Space X merger? fortune.com/2026/10/22/i...

🤖 We're hiring a Postdoc in Socially Intelligent AI & Robotics at King's College London! Join our SAIR Lab to advance human–robot interaction and AI for healthcare. 📅 Deadline: 18 Nov 2026 Full job description and application: my.corehr.com/pls/kclrecru...

This makes me far more skeptical of theory-only papers. This doesn't just apply to academic writing. If you ask someone to invest the time in reading (an email, an ad, a blog post, or a million other things) - when it's clearly AI generated you're asking me to invest more time than you may have.

I think I discovered the system prompt that LLMs use when they are asked to generate entier paper submissions.

#ICLR2027 reviews done!

🚨 New preprint: "Dino Forcing Flow Models: Do not denoise what you can predict" @arrijitghosh.bsky.social @lucasdegeorge.bsky.social P. Couairon, A.A. Efros, @vickykalogeiton.bsky.social and me 📜 arxiv.org/abs/2610.11751 🖥️ arijit-hub.github.io/projects/din... 🤖 github.com/arijit-hub/d... 🧵👇

5 journals in probability and statistics decided to close temporarily their submission system, due to the crazy increase in the number of submissions 😮 Annales de l'Institut Poincaré B, Bernouilli, Electronic Communications in Probability, Electronic Journal of Probability and ESAIM: P&S.

Claude: "This implementation will require roughly 2 days". Does it in 10min. As long as this phenomenon persists, we know that human data dominates in training. Or am I missing something here?

What is TMLR doing under a 300% increase in submissions? - Suspensions of >150 authors - Publicly naming major violations - Submission quotas distinguishing slop patterns - Desk rejections 6% → 53%; focus on clarity - AI reviews evaluating soundness blog.tmlr.org/2026/adaptin...

I was Today years old when I learned that these two characters are played by the same actress (Jenette Goldstein)

🚀 🤖 Fully Funded PhD Opportunity in Human–Robot Interaction at King's College London! ⏳ Application deadline: 11 January 2027 🎯 Start date: October 2027 🔗 Full details and application instructions: www.findaphd.com/phds/project...

Doing my first #iclr2027 review: in the age of LLM-slop papers, this is IMO probably the most important criterion which should help us to judge whether a paper should be accepted. Did we learn something new, or was something obvious tested with Claude?

Naver Labs Europe's chateau has been renovated and some of our teams are moving back there 😍 (It's one of our 4 buildings).

If Halloween was a thing where I live, I'd dress up as the looming risk of a break down of the banking system due to LLM-agent based cyber attacks.

I really enjoy finding gaps in our understanding of text-book deep learning. The RNN Homunculus: Long-range input vanishes RNN gradients. Gates (GRU, LSTM) help, but gates are themselves an RNN. Ie: gates are not sufficient. We need short-range data to learn gates first. arxiv.org/abs/2610.14279

It is the *best* time to start your own research group in AI @compscibristol.bsky.social Why? 🌟 Bristol hosts the largest supercomputer in the UK - Isambard AI, 4th greenest in the world. 🧑‍🎓 New undergrad degree CS w AI &multiple masters 6⃣ Nov DL to apply for position, Link in 🧵

Whatever you think of what happens in geopolitics lately, it has a really good side effect on Europe's efforts to create and to push its own global payment system. Which cannot be sanctioned (and whose benefits stay in the region).

Funnily (hum...) all presentations and figures now look a like and have the same style.

Recurrent Looped Transformer Zhang et al. Combine context/history transformer with recurrence: have a causal encoder with a recurrent decoder. arxiv.org/abs/2610.07591

I am registered for the Paris venue of NeurIPS 2026! See you there 😀

We welcome all authors of accepted NeurIPS papers who chose the Paris venue. We are excited and working hard to make this a great event! Registrations for authors are not yet open but be reassured that 1 place is reserved for the presenter of each paper accepted to the main conf. WS authors: 1/2

he cannot be trusted outside so he gets the harness

Le Chonk is here 🐱 Mistral's Large 4 is a 1 trillion parameter model trained in its own European data centres, with open weights coming on 27 October. Looking at the benchmarks, it's up there with the best open models. It even does a bit better than GPT-6 Astra on some aspects. Bien joué, Mistral !

Procedural Core: A Compact Recurrent Initialization for Vision Transformers By @damienteney.bsky.social 's group Idea: train a ViT with parameters tied over layers (="recurrent ViT") on abstract procedurally data. Then initialise models of ANY arch (!) with these weights. arxiv.org/abs/2609.37631

Having fun with Blender ... Distilling context transformers into recurrent transformers: arxiv.org/abs/2606.21562 @weinzaepfelp.bsky.social @mbsariyildiz.bsky.social

I am running a study on how computer vision researchers perceive dual-use risks of the field. If you work in CV, and would like to help us, the survey takes less than 3min. Reply or DM me for the link. You will be helping us a lot! Reposts are much appreciated 🙏

Our second call for NeurIPS Paris 2026 volunteers will open tomorrow TUE Oct 6th, 2pm CET. We are still seeking for PhD students and postdocs to assist during the conference in exchange for complimentary registration. Application form: forms.gle/bBFUyUNngtCc...

Just checked the 3 papers I need to review for ICLR 2027. Topics: perfect match. The paper IDs however ... we were warned but it's still a shock when you see them.

arXiv has updated our policy on rate limiting for all submitters. This update was made to fairly distribute moderator time & support the arXiv community of staff, volunteers, readers & authors. Please read our announcement to learn more: https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/

Je sens que Bluesky est pret pour des memes sur la causalité.