@misc{WRZP2026KVSurgeon,
author = {Wagner, Moritz and Roux, Christophe and Zimmer, Max and Pokutta, Sebastian},
title = {{The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction}},
year = {2026},
month = oct,
doi = {10.48550/arXiv.2610.25933},
eprint = {2610.25933},
archivePrefix = {arXiv},
}
~/pokutta $ whoami
Sebastian Pokutta
Professor @ TU Berlin · Vice President @ Zuse Institute Berlin · Chair @ MATH+ · Chair @ MODAL
# artificial intelligence, optimization, and machine learning
# tl;dr: we use computers to learn from data and make better decisions.
~/pokutta $ cd lab
→ Interactive Optimization and Learning (IOL) iol.zib.de ↗ research ↗ openings ↗
research
My group works on Artificial Intelligence, Optimization, and Machine Learning; from foundational algorithms to real-world systems. We develop new optimization and learning methods with a focus on efficiency and sustainability, combine learning with decision-making, and build agentic AI systems for complex tasks. A particular focus is on AI4Science and AI4Math, where we use AI to drive mathematical discovery and scientific insight. Recent examples include global forest monitoring from satellite data, entanglement certification via optimization, new constructions in extremal combinatorics, AI-assisted mathematical proofs, and AI for creativity and education. group homepage more about research and projects
selected recent work
all publications2026 The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction preprint arXiv llmmlpruning2026 The exact worst-case asymptotic rate of Nesterov acceleration for strongly convex optimization preprint arXiv visuals complexitylowerboundsopt@misc{PS2026NesterovExactRate, author = {Pokutta, Sebastian and Stellato, Bartolomeo}, title = {{The exact worst-case asymptotic rate of Nesterov acceleration for strongly convex optimization}}, year = {2026}, month = oct, doi = {10.48550/arXiv.2610.15125}, eprint = {2610.15125}, archivePrefix = {arXiv}, }2026 CryptHarness: A Harness for Agent-Assisted Cryptanalytic Research Proceedings of the NeurIPS 2026 Workshop on Secure and Trustworthy Quantum Machine Learning (SaTQuML) agenticcryptanalysisllmml (Oral Presentation + Workshop)@inproceedings{PKMP2026CryptHarnessSaTQuML, author = {Pelleriti, Nico and Kraus, Nicolai and May, Alexander and Pokutta, Sebastian}, title = {{CryptHarness: A Harness for Agent-Assisted Cryptanalytic Research}}, booktitle = {{Proceedings of the NeurIPS 2026 Workshop on Secure and Trustworthy Quantum Machine Learning (SaTQuML)}}, year = {2026}, month = sep, }2026 Zero: A Meta-Agent for Mathematical Research Proceedings of the Sixth NeurIPS Workshop on Mathematical Reasoning and AI (MATH-AI) agenticai4mathml@inproceedings{PMZP2026ZeroMathAI, author = {Pelleriti, Nico and Plato, Konrad and Zimmer, Max and Pokutta, Sebastian}, title = {{Zero: A Meta-Agent for Mathematical Research}}, booktitle = {{Proceedings of the Sixth NeurIPS Workshop on Mathematical Reasoning and AI (MATH-AI)}}, year = {2026}, month = sep, }2026 Curvature-Dependent Lower Bounds for Frank-Wolfe to appear in Proceedings of NeurIPS arXiv slides complexityfwlowerboundsopt@inproceedings{HRP2026CurvatureFWLowerBounds, author = {Halbey, Jannis and Roux, Christophe and Pokutta, Sebastian}, title = {{Curvature-Dependent Lower Bounds for Frank-Wolfe}}, booktitle = {{to appear in Proceedings of NeurIPS}}, year = {2026}, month = sep, doi = {10.48550/arXiv.2605.10595}, eprint = {2605.10595}, archivePrefix = {arXiv}, }2026 Frank-Wolfe Beyond 1/t Convergence to appear in Proceedings of NeurIPS arXiv slides complexityfwopt@inproceedings{P2026FrankWolfe, author = {Pokutta, Sebastian}, title = {{Frank-Wolfe Beyond 1/t Convergence}}, booktitle = {{to appear in Proceedings of NeurIPS}}, year = {2026}, month = sep, doi = {10.48550/arXiv.2604.28006}, eprint = {2604.28006}, archivePrefix = {arXiv}, }2026 A Free Lunch in LLM Compression: Revisiting Retraining after Pruning to appear in Proceedings of NeurIPS arXiv llmmlpruningsparsity@inproceedings{WZRP2025, author = {Wagner, Moritz and Roux, Christophe and Zimmer, Max and Pokutta, Sebastian}, title = {{A Free Lunch in LLM Compression: Revisiting Retraining after Pruning}}, booktitle = {{to appear in Proceedings of NeurIPS}}, year = {2026}, month = sep, doi = {10.48550/arXiv.2510.14444}, eprint = {2510.14444}, archivePrefix = {arXiv}, }2026 Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks to appear in Proceedings of NeurIPS arXiv creativityhaiillmml@inproceedings{HGHLKMP2026LLMCreativityVariance, author = {Haase, Jennifer and Gonnermann-M{\"u}ller, Jana and Hanel, Paul H. P. and Leins, Nicolas and Kosch, Thomas and Mendling, Jan and Pokutta, Sebastian}, title = {{Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks}}, booktitle = {{to appear in Proceedings of NeurIPS}}, year = {2026}, month = sep, doi = {10.48550/arXiv.2601.21339}, eprint = {2601.21339}, archivePrefix = {arXiv}, }2026 The Agentic Researcher: A Practical Guide to AI-Assisted Research in Mathematics and Machine Learning Proceedings of the ICML 2026 Workshop on AI as a Tool for Mathematics, Computer Science, and Machine Learning (AI4Research) pdf arXiv summary code agenticai4mathml (Oral Presentation + Workshop)@inproceedings{ZPRP2026AgenticResearcherAI4Research, author = {Zimmer, Max and Pelleriti, Nico and Roux, Christophe and Pokutta, Sebastian}, title = {{The Agentic Researcher: A Practical Guide to AI-Assisted Research in Mathematics and Machine Learning}}, booktitle = {{Proceedings of the ICML 2026 Workshop on AI as a Tool for Mathematics, Computer Science, and Machine Learning (AI4Research)}}, year = {2026}, month = jun, doi = {10.48550/arXiv.2603.15914}, eprint = {2603.15914}, archivePrefix = {arXiv}, }2026 Symmetric Extension Complexity of the Spanning Tree Polytope preprint arXiv extendedformulationipopt@misc{P2026SymmetricExtensionComplexity, author = {Pokutta, Sebastian}, title = {{Symmetric Extension Complexity of the Spanning Tree Polytope}}, year = {2026}, month = jun, doi = {10.48550/arXiv.2606.17017}, eprint = {2606.17017}, archivePrefix = {arXiv}, }2026 A Counterexample to Ziegler’s Cross-Polytope Conjecture for Simplicial 0/1-Polytopes preprint arXiv combinatoricsopt@misc{KP2026CrossPolytope, author = {Kaibel, Volker and Pokutta, Sebastian}, title = {{A Counterexample to Ziegler's Cross-Polytope Conjecture for Simplicial 0/1-Polytopes}}, year = {2026}, month = jun, doi = {10.48550/arXiv.2606.31640}, eprint = {2606.31640}, archivePrefix = {arXiv}, }2026 From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs Proceedings of the 43rd International Conference on Machine Learning (ICML) arXiv summary cognitivellmmlxai@inproceedings{SZRPG2026SemanticGeometry, author = {Schiekiera, Louis and Zimmer, Max and Roux, Christophe and Pokutta, Sebastian and G{\"u}nther, Fritz}, title = {{From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs}}, booktitle = {{Proceedings of the 43rd International Conference on Machine Learning (ICML)}}, year = {2026}, month = may, doi = {10.48550/arXiv.2602.00628}, eprint = {2602.00628}, archivePrefix = {arXiv}, }2026 Lower Bounds for Frank-Wolfe on Strongly Convex Sets Proceedings of the 43rd International Conference on Machine Learning (ICML) pdf arXiv summary slides complexityfwlowerboundsopt@inproceedings{HDZRSP2026FWLowerBounds, author = {Halbey, Jannis and Deza, Daniel and Zimmer, Max and Roux, Christophe and Stellato, Bartolomeo and Pokutta, Sebastian}, title = {{Lower Bounds for Frank-Wolfe on Strongly Convex Sets}}, booktitle = {{Proceedings of the 43rd International Conference on Machine Learning (ICML)}}, series = {Proceedings of Machine Learning Research}, volume = {306}, pages = {39468--39494}, publisher = {PMLR}, year = {2026}, month = may, doi = {10.48550/arXiv.2602.04378}, eprint = {2602.04378}, archivePrefix = {arXiv}, }2026 Neural Concept Verifier: Scaling Prover-Verifier Games via Concept Encodings Proceedings of the 43rd International Conference on Machine Learning (ICML) arXiv summary mlxai (Spotlight + Conference Proceedings)@inproceedings{TASSP2026NCVICML, author = {Turan, Berkant and Asadulla, Suhrab and Steinmann, David and Kersting, Kristian and Stammer, Wolfgang and Pokutta, Sebastian}, title = {{Neural Concept Verifier: Scaling Prover-Verifier Games via Concept Encodings}}, booktitle = {{Proceedings of the 43rd International Conference on Machine Learning (ICML)}}, year = {2026}, month = may, doi = {10.48550/arXiv.2507.07532}, eprint = {2507.07532}, archivePrefix = {arXiv}, }2026 Faster Algorithms for Structured Matrix Multiplication via Flip Graph Search Proceedings of the International Symposium on Symbolic and Algebraic Computation (ISSAC) pdf arXiv compalgcomputational (Distinguished Student Author Award at ISSAC 2026)@inproceedings{KGP2025FlipGraph, author = {Khoruzhii, Kirill and Gel{\ss}, Patrick and Pokutta, Sebastian}, title = {{Faster Algorithms for Structured Matrix Multiplication via Flip Graph Search}}, booktitle = {{Proceedings of the International Symposium on Symbolic and Algebraic Computation (ISSAC)}}, year = {2026}, month = apr, doi = {10.1145/3815436.3815455}, eprint = {2511.10786}, archivePrefix = {arXiv}, }2026 The Hidden Cost of Tokenization: Why (most) Non-English Speakers Pay More for Less preprint arXiv summary fairnessllmmlmultilingual@misc{HP2026TokenizationCost, author = {Haase, Jennifer and Pokutta, Sebastian}, title = {{The Hidden Cost of Tokenization: Why (most) Non-English Speakers Pay More for Less}}, year = {2026}, month = jan, doi = {10.5281/zenodo.18416604}, }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@inproceedings{Pelleriti2025NeuralSOS, author = {Pelleriti, Nico and Spiegel, Christoph and Liu, Shiwei and Mart\'inez-Rubio, David and Zimmer, Max and Pokutta, Sebastian}, title = {{Neural Sum-of-Squares: Certifying the Nonnegativity of Polynomials with Transformers}}, booktitle = {{to appear in Proceedings of the International Conference on Learning Representations (ICLR)}}, year = {2026}, month = jan, doi = {10.48550/arXiv.2510.13444}, eprint = {2510.13444}, archivePrefix = {arXiv}, }2025 Conditional Gradient Methods MOS-SIAM Series on Optimization pdf arXiv mloptsurvey@book{CGFWSurvey2022, author = {Braun, G{\'a}bor and Carderera, Alejandro and Combettes, Cyrille W. and Hassani, Hamed and Karbasi, Amin and Mokhtari, Aryan and Pokutta, Sebastian}, title = {Conditional Gradient Methods}, publisher = {MOS-SIAM Series on Optimization}, year = {2025}, month = jan, isbn = {978-1-61197-855-1}, eprint = {2211.14103}, archivePrefix = {arXiv}, }2024 Quantum Optimization: Potential, Challenges, and the Path Forward Nature Reviews Physics pdf arXiv optphysicsquantumsurvey@article{AbbasEtAl2023, author = {Abbas, Amira and Ambainis, Andris and Augustino, Brandon and B{\"a}rtschi, Andreas and Buhrman, Harry and Coffrin, Carleton and Cortiana, Giorgio and Dunjko, Vedran and Egger, Daniel J. and Elmegreen, Bruce G. and Franco, Nicola and Fratini, Filippo and Fuller, Bryce and Gacon, Julien and Gonciulea, Constantin and Gribling, Sander and Gupta, Swati and Hadfield, Stuart and Heese, Raoul and Kircher, Gerhard and Kleinert, Thomas and Koch, Thorsten and Korpas, Georgios and Lenk, Steve and Marecek, Jakub and Markov, Vanio and Mazzola, Guglielmo and Mensa, Stefano and Mohseni, Naeimeh and Nannicini, Giacomo and O'Meara, Corey and Pe{\~n}a Tapia, Elena and Pokutta, Sebastian and Proissl, Manuel and Rebentrost, Patrick and Sahin, Emre and Symons, Benjamin C. B. and Tornow, Sabine and Valls, Victor and Woerner, Stefan and Wolf-Bauwens, Mira L. and Yard, Jon and Yarkoni, Sheir and Zechiel, Dirk and Zhuk, Sergiy and Zoufal, Christa}, title = {{Quantum Optimization: Potential, Challenges, and the Path Forward}}, journal = {Nature Reviews Physics}, year = {2024}, month = sep, eprint = {2312.02279}, archivePrefix = {arXiv}, }2024 The Frank-Wolfe algorithm: a short introduction Jahresbericht der Deutschen Mathematiker-Vereinigung, 126, 3–35 pdf arXiv mlopt@article{P2023, author = {Pokutta, S.}, title = {{The Frank-Wolfe algorithm: a short introduction}}, journal = {{Jahresbericht der Deutschen Mathematiker-Vereinigung}}, volume = {126}, pages = {3--35}, year = {2024}, month = jan, eprint = {2311.05313}, archivePrefix = {arXiv}, }
talks & teaching
archive-
11/2026
Agentic AI Workflows in Mathematics
-
10/2026
How Machines Explore, Conjecture, and Discover Mathematics
-
10/2026
How Machines Explore, Conjecture, and Discover Mathematics
-
09/2026
The Agentic Researcher: Agentic AI for Analytics and Operations Research
-
09/2026
The Agentic Researcher: Engineering Autonomous Research Workflows
-
WS/
2026 Discrete Optimization and Machine Learning