Lars Holdijk

My research focuses on the relationship between stochastic thermodynamics and generative AI: its theoretical principles, the methods it enables, and how we can potentially use this synergy to build more efficient hardware to run these models. In particular, I exploit this relationship to build new methods that push generative AI into problems in computational chemistry and biochemistry, especially the computational challenge of free-energy estimation.

Most recently I was a Research Scientist at Normal Computing in New York. Before starting my PhD I obtained a master's degree from the University of Amsterdam, supervised by Max Welling and Yarin Gal as part of the ELLIS MSc programme. Prior to that I worked as a student researcher at Porsche and obtained my bachelor's degree in computer science from the University of Groningen; at the time, my research focused primarily on robust and interpretable machine learning.

Recent / Upcoming highlights

Generative AI and Stochastic Thermodynamics: A Tale of Free Energies

My book with Max Welling and Sirui Lu, is out now with Cambridge University Press. It can be ordered on Amazon here.

Upcoming Talk: Non-equilibrium free-energy estimation, from protocols to learned transport

I'll be speaking at the Newcastle Workshop on Non-Equilibrium Sampling: Diffusion, Flows, and Particles, September 4-6 2026.

Publications

[2026] M. Welling, S. Lu, L. Holdijk. Generative AI and Stochastic Thermodynamics: A Tale of Free Energies. Cambridge University Press. (book)
[2026] L. Holdijk, D. Melanson, B. Birchall, V. Cheung, N. Lehrter, M. Aifer, S. Duffield, J.O. Ernst, R. Salegame, A.J. Martinez, G. Crooks, P.J. Coles, Z. Belateche, M. Bright. CN101: A Digital Thermodynamic Computer for Generative AI. arXiv:2608.00754. (under review)
[2026] L. Holdijk, NM. Anand, M. Bronstein, M. Welling. Learning Escorted Protocols for Multistate Free-Energy Estimation. ICLR'26.
[2026] Z. Mensch, L. Holdijk, S. Duffield, M. Aifer, P.J. Coles, M. Welling, M.C.N. Cheng. Robust Stochastic Gradient Posterior Sampling with Lattice Based Discretisation. ICML'26.
[2026] R. Salegame, J.O. Ernst, N. Lehrter, M. Komeili, D.M. Saberi, L. Holdijk, M. Khomiakov, T.D. Ahle. TitanBench: A Benchmark for Agentic RTL Design and Verification. (under review, NeurIPS'26)
[2026] D. Kim, R. Salegame, L. Holdijk, J.O. Ernst. AutoCircuit: Agentic Pareto Frontier Exploration for Analog Circuit Designs. ICML'26 workshop on AI4Research.
[2025] J. Lee, M. Plainer, Y. Du, L. Holdijk, R. Brekelmans, D. Beaini, K. Neklyudov. Scaling Deep Learning Solutions for Transition Path Sampling. ICLR'25 Frontiers in Probabilistic Inference.
[2023] L. Holdijk*, Y. Du*, F. Hooft, P. Jaini, B. Ensing, M. Welling. Stochastic Optimal Control for Collective Variable Free Sampling of Molecular Transition Paths. NeurIPS'23.
[2021] P. Jaini*, L. Holdijk*, M. Welling. Learning Equivariant Energy Based Models with Equivariant Stein Variational Gradient Descent. NeurIPS'21.
[2020] S. Saralajew*, L. Holdijk*, T. Villmann. Fast Adversarial Robustness Certification of Nearest Prototype Classifiers for Arbitrary Seminorms. NeurIPS'20.
[2019] S. Saralajew*, L. Holdijk*, E. Asan, M. Rees, T. Villmann*. Classification-by-Components: Probabilistic Modeling of Reasoning over a Set of Components. NeurIPS'19.
[2019] S. Saralajew, L. Holdijk, M. Rees, T. Villmann. Robustness of Generalized Learning Vector Quantization Against Adversarial Attacks. WSOM'19.
[2018] S. Saralajew, L. Holdijk, M. Rees, T. Villmann. Prototype-based Neural Network Layers: Incorporating Vector Quantization. NeurIPS'18 workshop on Adversarial Vision.

(*: denotes equal contribution)

Page updated

Google Sites

Report abuse