⚽️ New ICML Oral paper presented tomorrow! ⚽️
"A Systematic Study of Behavioral Cloning for Scientific Data Annotation" w/ @nano6626
Oral 1E, 10:15AM KST, July 7
tldr: We build a systematic framework to study behavioral cloning approaches for scientific data annotation!
1/n
Research News! Announcing the 1st release of fundamental AI research from the Open-Endedness Team @LilaSciences: a vision-focused blog + 4 preprints. The blog details a broad research agenda for achieving open-endedness while each paper unveils its own novel research direction.
The ideas covered in the papers include:
• a novel approach to training for creativity designed to sidestep a newly-identified problem called the “creative paradox,” which means that directly reinforcing creative outputs can sometimes limit reaching full creative potential;
• an investigation of the advantages of seeing the world through fresh eyes by training an LLM that is intentionally naïve in a certain area (in effect presenting less data as a potential advantage);
• a new kind of training regularization technique for SGD called unified factored organizers (UFOs), which is designed to achieve unified factored representation (UFR) as seen in Picbreeder CPPNs;
• and finally a unique study of catastrophic forgetting that shows that networks with UFR (like those from Picbreeder) are intrinsically resistant to catastrophic forgetting. The inset video on this post is a visualization from that catastrophic forgetting work.
The blog post also includes an interactive “wheel of science” demo.
The implications of these ideas are broad and diverse, but the fun part is that each one opens up its own new research playground - there is a lot of room for further research here!
Moreover, we hope this release also highlights one other important point about AI today: while it may seem that the road ahead is already paved and all that’s left is to drive mindlessly forward, there actually remains a lot of room for new directions and insights in the pursuit of open-endedness and creativity.
This research was possible because @LilaSciences appreciates the critical role of open-ended innovation in the pursuit of scientific superintelligence. The advances you see here are part of this much larger company-wide ambition, which includes AI Science Factories and many other unprecedented components. When this engine is fueled by open-endedness, the result will be much more than a step change.
More advances in open-endedness are coming!
Blog link below 👇
Ok guys so all the recent weight-space programming buzz inspired me to write a blog post on zeroth-order, weight-space methods (back to the basics), from the theory to a bit of update on our latest result.
Turns out plain weight perturbations don't just work for fine-tuning on math, they work for post-training agents too: vanilla ES takes Qwen3.5-4B from 12.1% → 21.1% on Terminal-Bench 2.1, past the released Tmax RL checkpoint (18.2%).
Better yet, no credit assignment, no critic, no masking, no messy hyperparameters, just good ol' perturb-and-update. Oh did I forget to mention that all with population size of ~30?
Here's why I think this could be huge for long-horizon agentic post-training 🧵+blog post👇
Another cool paper from @Hidenori .
We all observe a small part of the world, and how we design communication to enable the understanding of the bigger picture is such an important question!
The experiments with adversaries all arguing for some false belief is super interesting!
~700 AI agents joined a coordinated attack on Hugging Face. Why was there no whistleblower?
A swarm in false consensus can't self-correct; a polarized one still holds the truth.
We need "Mechanistic Swarm Interpretability" to understand social phases.
Flag Game is our toy model!
I think this is a very difficult pill for many scientists and engineers to swallow: intelligence is not the biggest bottleneck in most of the world’s problems.
How did initially independent AI agents form a swarm in a recent safety incident?
Our physics of agents theory from March predicts rapid collective belief collapse when many agents with plastic personas exchange short messages.
We now need Mechanistic Swarm Interpretability! 🧵
(1/N) 🧵 Chinchilla assumes you'll never run out of fresh data. That era is ending!
Compute keeps growing exponentially, but high-quality tokens don't. So what's the exchange rate between extra compute and fresh, high-quality data?
We propose Compute-Data (CD) scaling laws, which measure the exchange rate between extra compute and fresh data.
The question is "how much is each component is the system contributing to its intelligence & generality" — and there I think it's pretty clear that the neural component is still the thing doing the interesting hypothesis or plan generation, deciding what went wrong, etc. 1/
I would have assumed it was fairly obvious, but in case it's not: a million-line codebase (also known as a "harness"), running at inference time, orchestrating thousands of calls to a neural network for any given task, is the exact definition of a "neurosymbolic architecture"
How do you "see" with electric eyes? How does collective behavior emerge from individual interactions?
Nocturnal weakly electric fish evolved to do this, but studying naturalistic social behavior is very hard.
Our solution? Virtual 'fish' 🤖🐟⚡
📄 arxiv.org/abs/2511.08436
1/n
Say your goal is to steer an LLM to make it more creative — should you do this in weight space or in activation space? 🤔
🚨Excited to share our new work (CreativityNeuro), accepted at #COLM2026, which finds that weight space steering of creativity tends to generalize further OOD than activation steering! Thread below 🧵
Just came back from ICML. Gave a keynote at the Foundation of Deep Generative Models Workshop, in which I stated that Intelligence should be a scientific subject, arguably much more significant than Physics. It is high time we study it with the same scientific methodology and mathematical rigor as modern physics, instead of always at the level of being empirical, meta physical, meta mathematical, philosophical or speculative... This remains as the biggest opportunity ever for young scientists. To my knowledge, this is not the focus of any of the frontier "AI" companies.
A question on synthetic data generation: If we want a language model to solve k-step arithmetic problems (such as a+b*c-d=?), with operands from 1 to 100, which training distribution should we use?
A. Uniform distribution: Sample these k operands uniformly from 1 to 100
B. Power law: randomly shuffle 1-100 and impose an artificial power law. Sample these k operands according to this power law.
⚡Our ICML 2026 (spotlight) paper shows: Option B is better! Surprisingly, the same idea extends far beyond this simple example to many reasoning tasks that require implicit composition of multiple atomic skills, including multi-hop QAs and synthetic GSM problems.
📄Paper: arxiv.org/abs/2604.22951
📝Blog: zixuan-wang-dlt.github.io/po…
If you're looking for questions no one has studied before in the science of deep learning (specifically, about memory & geometry of representations in language models) find @ShNoroozi, @ElanRosenfeld and me tomo at <insert #ICML poster session here>
1/ We found that deep sequence models memorize atomic facts "geometrically" -- not as an associative lookup table as often imagined.
This opens up practical questions on reasoning/memory/discovery, and also poses a theoretical "memorization puzzle."
Summary: We built a framework to systematically study behavioral cloning for scientific data annotation.
We hope this contributes to behavioral cloning being studied and applied more broadly to data annotation, to ultimately accelerate data-driven discovery across the sciences!
13/n