Sebastian Pokutta

Artificial Intelligence, Optimization, and Machine Learning. Professor @TUBerlin, Vice President @ZuseInstitute, Chair @MATH+, Chair @MODAL

Research Lab

My group is interested in Artificial Intelligence, Optimization, and Machine Learning and its applications. We develop new methodologies (e.g., new optimization and learning algorithms), work on combining learning and decision-making, as well as design AI Systems for real-world deployment in various application contexts. Recent examples include the estimation of biomass from satellite data, entanglement and non-locality thresholds via optimization, new constructions in extremal combinatorics via AI, as well as questions around AI and creativity. group homepage more about research and projects

TL;DR. We use computers to learn from data and make better decisions.

ETC. Hashed Commitments AI News

Prospective Students. If you are interested in working in our group or writing your MS/BS thesis please check our openings.

Select Recent Papers

(see publications for a complete list)

  1. Muhtar, D., Song, X., Pokutta, S., Zimmer, M., Pelleriti, N., Hofmann, T., and Liu, S. (2026). When Does Sparsity Mitigate the Curse of Depth in LLMs. Proceedings of ICLR 2026 Workshop on Decoding LLM Training and Adaptation (DeLTa). [PDF] llmmlsparsity
  2. Halbey, J., Deza, D., Zimmer, M., Roux, C., Stellato, B., and Pokutta, S. (2026). Lower Bounds for Frank-Wolfe on Strongly Convex Sets. Preprint. [arXiv] complexityfwlowerboundsopt
  3. Leins, N., Gonnermann-Müller, J., Teichmann, M., and Pokutta, S. (2026). Investigating the Influence of Spatial Ability in Augmented Reality-assisted Robot Programming. Preprint. [arXiv] arhrimlrobotics
  4. Schiekiera, L., Zimmer, M., Roux, C., Pokutta, S., and Günther, F. (2026). From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs. Preprint. [arXiv] [summary] cognitivellmmlxai
  5. Berthold, T., Kamp, D., Mexi, G., Pokutta, S., and Pólik, I. (2026). Global Optimization for Combinatorial Geometry Problems Revisited in the Era of LLMs. To Appear in Proceedings of ISCO, Lecture Notes in Computer Science. [arXiv] computationalllmnlpopt
  6. Xiao, W., Besançon, M., Gelß, P., Hendrych, D., Klus, S., and Pokutta, S. (2026). Graph Isomorphism: Mixed-Integer Convex Optimization from First-Order Methods. To Appear in Proceedings of CPAIOR. [arXiv] computationalgraphsmipopt
  7. Pauls, J., Schrödter, K., Ligensa, S., Schwartz, M., Turan, B., Zimmer, M., Saatchi, S., Pokutta, S., Ciais, P., and Gieseke, F. (2026). ECHOSAT: Estimating Canopy Height Over Space And Time. Preprint. [arXiv] [code] [visuals] ai4sciencemlsustainability
  8. Haase, J., Gonnermann-Müller, J., Hanel, P. H. P., Leins, N., Kosch, T., Mendling, J., and Pokutta, S. (2026). It’s Not Just the Prompt: Model Choice Dominates LLM Creative Output. To Appear in Proceedings of the ACM CHI Conference on Human Factors in Computing Systems (CHI). creativityhaiillmml
  9. Gonnermann-Müller, J., Haase, J., Leins, N., Kosch, T., and Pokutta, S. (2026). Maintaining Stable Personas? Examining Temporal Stability in LLM-Based Human Simulation. To Appear in Proceedings of the ACM CHI Conference on Human Factors in Computing Systems (CHI). haiillmmlmultiagentsocial
  10. Geiselmann, Z., Joswig, M., Kastner, L., Mundinger, K., Pokutta, S., Spiegel, C., Wack, M., and Zimmer, M. (2026). Patchworked Curves of Degree Seven. Preprint. [arXiv] ai4mathalggeomcombinatoricscompalg
  11. Khoruzhii, K., Gelß, P., and Pokutta, S. (2026). Tensor Decomposition for Non-Clifford Gate Minimization. Preprint. [arXiv] compalgquantum
  12. Kerdreux, T., Scieur, D., Martinez-Rubio, D., d’Aspremont, A., and Pokutta, S. (2026). Strong Convexity of Sets in Riemannian Manifolds. To Appear in Proceedings of the International Conference on Learning Representations (ICLR). [arXiv] mlopt
  13. Gonnermann-Müller, J., Haase, J., Leins, N., Igel, M., Fackeldey, K., and Pokutta, S. (2026). FACET: Multi-Agent AI Supporting Teachers in Scaling Differentiated Learning for Diverse Students. Preprint. [arXiv] educationhaiillmmlmultiagent
  14. Gonnermann-Müller, J., Haase, J., Leins, N., Kosch, T., and Pokutta, S. (2026). Stable Personas: Dual-Assessment of Temporal Stability in LLM-Based Human Simulation. Preprint. [arXiv] haiillmmlmultiagentsocial
  15. Haase, J., and Pokutta, S. (2026). The Hidden Cost of Tokenization: Why (most) Non-English Speakers Pay More for Less. Preprint. [arXiv] fairnessllmmlmultilingual
  16. Haase, J., Gonnermann-Müller, J., and Pokutta, S. (2026). Building Socially Grounded Multi-Agent LLM Systems Requires the Transition from Static LLM Prompting to Autonomous Multi-Agent Ecosystems. Preprint. [arXiv] haiillmmlmultiagentsocial
  17. Haase, J., Gonnermann-Müller, J., Hanel, P. H. P., Leins, N., Kosch, T., Mendling, J., and Pokutta, S. (2026). Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks. Preprint. [arXiv] creativityhaiillmml
  18. Leins, N., Gonnermann-Müller, J., Teichmann, M., and Pokutta, S. (2026). Beyond Static Instruction: A Multi-Agent AI Framework for Adaptive Augmented Reality Robot Training. To Appear in Proceedings of ACM/IEEE International Conference on Human-Robot Interaction (HRI), Late-Breaking Reports. [arXiv] arhrimlrobotics
  19. Pelleriti, N., Spiegel, C., Liu, S., Martínez-Rubio, D., Zimmer, M., and Pokutta, S. (2026). Neural Sum-of-Squares: Certifying the Nonnegativity of Polynomials with Transformers. To Appear in Proceedings of the International Conference on Learning Representations (ICLR). [arXiv] ai4mathcompalgml
  20. Iommazzo, G., Martínez-Rubio, D., Criado, F., Wirth, E., and Pokutta, S. (2026). Linear Convergence of the Frank-Wolfe Algorithm over Product Polytopes. To Appear in Proceedings of AISTATS. [arXiv] mlopt
  21. Urbano, A., Romero, D. W., Zimmer, M., and Pokutta, S. (2026). RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization. To Appear in Proceedings of the International Conference on Learning Representations (ICLR). [arXiv] mlsymmetry
  22. Takahashi, S., Pokutta, S., and Takeda, A. (2026). Accelerated Convergence of Frank–Wolfe Algorithms with Adaptive Bregman Step-Size Strategy. To Appear in Proceedings of the International Conference on Learning Representations (ICLR). [arXiv] complexityfwopt
  23. Martínez-Rubio, D., and Pokutta, S. (2026). Beyond Short Steps in Frank-Wolfe Algorithms. To Appear in Proceedings of the International Conference on Learning Representations (ICLR). [arXiv] mlopt
  24. Haase, J., and Pokutta, S. (2026). Human–AI Cocreativity: Exploring synergies across levels of creative collaboration. In J. C. Kaufman & M. Worwood (Eds.), Generative Artificial Intelligence and Creativity (pp. 205–221). [PDF] [arXiv] creativityhaiimlsocial
  25. Dang, S., Deza, A., Gupta, S., McNicholas, P. D., Pokutta, S., and Sugiyama, M. (Eds.). (2026). Data Science and Optimization (Vol. 91). Springer. mlopt
  26. Zimmer, M., Roux, C., Wagner, M., Hendrych, D., and Pokutta, S. (2025). SparseSwaps: Tractable LLM Pruning Mask Refinement at Scale. Preprint. [arXiv] llmmlpruningsparsity
  27. Kuzinowicz, D., Lichocki, P., Mexi, G., Pfetsch, M. E., Pokutta, S., and Zimmer, M. (2025). Objective Coefficient Rounding and Almost Symmetries in Binary Programs. Preprint. [arXiv] computationalmipoptsymmetry
  28. Xu, L., Liu, Y.-C., and Pokutta, S. (2025). Convex semidefinite tensor optimization and quantum entanglement. Preprint. [arXiv] optquantum
  29. Khoruzhii, K., Gelß, P., and Pokutta, S. (2025). Faster Algorithms for Structured Matrix Multiplication via Flip Graph Search. Preprint. [arXiv] compalgcomputational
  30. Roux, C., Zimmer, M., d’Aspremont, A., and Pokutta, S. (2025). Don’t Be Greedy, Just Relax! Pruning LLMs via Frank-Wolfe. Preprint. [arXiv] fwllmmloptpruningsparsity
  31. Wagner, M., Roux, C., Zimmer, M., and Pokutta, S. (2025). A Free Lunch in LLM Compression: Revisiting Retraining after Pruning. Preprint. [arXiv] llmmlpruningsparsity
  32. Gonnermann-Müller, J., Haase, J., Fackeldey, K., and Pokutta, S. (2025). FACET: Teacher-Centred LLM-Based Multi-Agent Systems – Towards Personalized Educational Worksheets. Preprint. [arXiv] educationhaiillmmlmultiagent
  33. Liu, Y.-C., Halbey, J., Pokutta, S., and Designolle, S. (2025). A Unified Toolbox for Multipartite Entanglement Certification. Preprint. [arXiv] optphysicsquantum
  34. Haase, J., and Pokutta, S. (2025). Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research. Preprint. [arXiv] haiimlsocial
  35. Porto, L. E. A., Designolle, S., Pokutta, S., and Quintino, M. T. (2025). Measurement incompatibility and quantum steering via linear programming. Preprint. [arXiv] optphysicsquantum
  36. Wirth, E., Peña, J., and Pokutta, S. (2025). Adaptive Open-Loop Step-Sizes for Accelerated Convergence Rates of the Frank-Wolfe Algorithm. Preprint. [arXiv] complexityfwopt
  37. Braun, G., Carderera, A., Combettes, C. W., Hassani, H., Karbasi, A., Mokthari, A., and Pokutta, S. (2025). Conditional Gradient Methods. MOS-SIAM Series on Optimization. [PDF] [arXiv] mloptsurvey
  38. Abbas, A., Ambainis, A., Augustino, B., Bärtschi, A., Buhrman, H., Coffrin, C., Cortiana, G., Dunjko, V., Egger, D. J., Elmegreen, B. G., Franco, N., Fratini, F., Fuller, B., Gacon, J., Gonciulea, C., Gribling, S., Gupta, S., Hadfield, S., Heese, R., … Zoufal, C. (2024). Quantum Optimization: Potential, Challenges, and the Path Forward. Nature Reviews Physics. [PDF] [arXiv] optphysicsquantumsurvey
  39. Pokutta, S. (2024). The Frank-Wolfe algorithm: a short introduction. Jahresbericht Der Deutschen Mathematiker-Vereinigung, 126, 3–35. [PDF] [arXiv] mlopt

Select Recent Talks and Teaching

Recent Blog Posts

Select Outreach

News