ICML 2026 was a blast. If you missed our spotlight on Neural Concept Verifiers, the blog post covers it in ~5 minutes. Paper: arxiv.org/abs/2507.07532 Blog: www.pokutta.com/blog/neural... Joint with Turan, Asadulla, Steinmann, Kersting, Stammer. #ICML2026 #NeuralConceptVerifier #XAI

Neural Concept Verifiers: Proving with Concepts, not Pixels

TL;DR: This is a short summary of our paper Neural Concept Verifier: Scaling Prover-Verifier Games via Concept Encodings by Berkant Turan, Suhrab Asadulla, David Steinmann, Kristian Kersting, Wolfgang Stammer, and Sebastian Pokutta. The paper was accepted as a spotlight at ICML 2026 (top ~2.2% of submissions). In a nutshell, we move Prover-Verifier Games from pixel space to concept space: a prover selects a sparse set of human-readable concepts, a verifier must classify using only those concepts, and an adversarial prover tests whether the verifier can be fooled. The result is a scalable route to verifiable, nonlinear concept-based classification on datasets…

pokutta.com

Last week in Magdeburg, Germany I demoed our agentic research framework (arxiv.org/abs/2603.15914) when visiting Volker Kaibel & Sebastian Sager. Volker suggested testing Ziegler's cross-polytope conjecture for simplicial 0/1-polytopes.

The Agentic Researcher: A Practical Guide to AI-Assisted Research...

AI tools and agents are reshaping how researchers work, from proving theorems to training neural networks. Yet for many, it remains unclear how these tools fit into everyday research practice....

arxiv.org

Thrilled that "The Agentic Researcher" (joint w/ Max Zimmer, Nico Pelleriti, & Christophe Roux) is accepted as an Oral at the ICML 2026 AI4Research workshop in Seoul! 🇰🇷 Let's move AI in science beyond writing assistants to rigorous, autonomous collaborators. 🧵👇

Great panel at @JWI_Berlin on 'The Cucumber's Dream' w/ Katrin Frisch, Tobias Wilke & Jennifer Haase. What happens to creativity when AI writes literature? We are still very much at the beginning of this journey, and where we'll end up is wide open. 🤖📚 #AI #literature

New paper (with Jennifer Haase): the hidden cost of tokenization. Tokenization is not a neutral preprocessing step. It determines how much you pay, how much context you get, how much compute burns, and how well a model reasons in your language. 🧵 thread:

A few weeks ago I talked about using Pi with Qwen3.6 27B locally - which (still) is a beast. This Pi plugin squeezes out even more: score outcomes -> use reflective prompt evaluation -> inject small priors. Improves the model w/o retraining/fine-tuning. github.com/pokutta/pi-...

GitHub - pokutta/pi-prior-plugin: Context priors plugin for pi

Context priors plugin for pi . Contribute to pokutta/pi-prior-plugin development by creating an account on GitHub.

github.com

Qwen 3.6 27B + Pi on a MacBook Pro, fully local: a beast. 27B dense model, flagship-level agentic coding, running entirely on hardware in your hands. Speed is impressive, Utility is extremely high. It punches orders of magnitude above its weight. Local AI is getting real.

Much of the current AI-for-mathematics discussion is framed as a contest between learned systems and classical methods, often missing the point. It is not AI vs classical, but where AI belongs in the stack. Three examples from our work: www.pokutta.com/blog/not-ev... #ai #math #agentic

Not every discovery needs an LLM

TL;DR: AI-driven scientific discovery with systems such as AlphaTensor, AlphaEvolve, etc is en vogue and we are also heavily invested in that space. However, these LLM- and RL-driven approaches, while impressive often in themselves, are not always the right tool to get the job done. In this post I will talk about three recent projects from our group that take different angles on LLM- and RL-driven mathematical discovery systems. Two of them revisit classical problems made famous through these AI systems and show that classical structured search matches or beats them on these specific problem classes; with a tiny fraction of the compute. The third example goes the other way around: AI not as discovery engine, but an AI-accelerated subroutine sits inside a classical structured search and lets us settle a thirty-year-old conjecture on real algebraic plane curves of degree seven. This points to an interesting shift in the recent AI vs. classical discussion: it is not so much about whether

pokutta.com

For a decade it was open whether Frank-Wolfe's O(1/√ε) rate on strongly convex sets is tight. We show it is: Ω(1/√ε), even for a simple quadratic on a unit ball.

SCIP Optimization Suite 10.0 is out 🎉 → Numerically exact solving mode for rational MILPs → 9-20% faster on hard MINLPs → IIS detection for debugging infeasible models → Better symmetry handling ... and many more For full details see: www.pokutta.com/blog/resear... #optimization

Our book on conditional gradients and Frank-Wolfe methods with Gábor Braun, @ACarderera, @CyrilleCmbt, Hamed Hassani, @aminkarbasi and Aryan Mokhtari just appeared in the MOS-SIAM Series on Optimization #ml #optimization #ai epubs.siam.org/doi/book/10...

New blog post: committing to secrets via hashing - a short, first-principles introduction #crypto #hashing => www.pokutta.com/blog/hashed...