Behrooz Tahmasebi

Research

I study how symmetry and geometric structure shape learning: how much they reduce the amount of data a model needs, when they can be exploited efficiently, and how they can be discovered from data. My work combines tools from learning theory, harmonic analysis, group representation theory, and differential geometry, with applications to deep learning and the foundations of large language models.

Tutorials

• Recent Developments in Geometric Machine Learning: Foundations, Models, and More, NeurIPS 2025.
• Model-Agnostic Machine Learning with Structure: Principles and Methods, SIAM Conference on Mathematics of Data Science (MDS26), 2026.

Media Coverage

• New algorithms enable efficient machine learning with symmetric data, MIT News, 2025.
• How symmetry can come to the aid of machine learning, MIT News, 2024.

Service

• NeurIPS: Area Chair (2026–); Reviewer (2021–2025)
• ICML: Area Chair (2026–); Reviewer (2022–2025)
• ICLR: Area Chair (2027–); Reviewer (2024–2026)
• AISTATS: Reviewer (2022–)
• AAAI: Senior Program Committee (2027–); Program Committee (2025–2026)

Publications

* denotes equal contribution.

2026

• Test-Time Symmetrization: Scaling Laws for Model-Agnostic Learning under Structure
  Behrooz Tahmasebi and Melanie Weber Preprint, 2026.
• Majority Is Not Enough: Symmetry and Structured Errors in Test-Time Voting
  Behrooz Tahmasebi and Melanie Weber Preprint, 2026.
• Label-Free Estimation of Structured Errors in Language Models
  Behrooz Tahmasebi and Melanie Weber Preprint, 2026.
• Valid Inference and Optimal Design for LLM-as-a-Judge under Rater Heterogeneity
  Behrooz Tahmasebi and Melanie Weber Preprint, 2026.
• Data Attribution under Redundant Training Data
  Behrooz Tahmasebi and Melanie Weber Preprint, 2026.
• Parameter Symmetries Are Not Task Symmetries: Which Invariances Can Weight-Space Learning Afford?
  Fynn Kiwit*, Behrooz Tahmasebi*, and Stefanie Jegelka Preprint, 2026.
• Dimension-Free Scaling Laws for Size Generalization: An Exponential Separation
  Behrooz Tahmasebi and Melanie Weber Preprint, 2026.
• Spontaneous Symmetry Breaking via Regularization: Hardness and Approximation
  Ashkan Soleymani*, Behrooz Tahmasebi*, Reyhaneh Hosseinpour, Tess Smidt, Patrick Jaillet, and Stefanie Jegelka Preprint, 2026.
• Sparse Data Augmentation for Optimization with Provable Guarantees [arXiv]
  Behrooz Tahmasebi and Melanie Weber IEEE CDC 2026.
• Data Augmentation: A Fourier Analysis Perspective [arXiv]
  Behrooz Tahmasebi, Melanie Weber, and Stefanie Jegelka COLT 2026. Oral presentation at the DeepMath Conference 2025.
• Efficient Learning and Symmetry Discovery under Exact Invariances [arXiv]
  Ashkan Soleymani*, Behrooz Tahmasebi*, Patrick Jaillet, and Stefanie Jegelka COLT 2026. Oral presentation at the NeurReps Workshop (NeurIPS 2025).
• Adaptive Symmetry Discovery for Dynamical System Identification [arXiv]
  Behrooz Tahmasebi and Melanie Weber ICML 2026.
• Achieving Approximate Symmetry Is Exponentially Easier than Exact Symmetry [arXiv]
  Behrooz Tahmasebi and Melanie Weber ICLR 2026. Oral presentation at the TAG-DS Workshop 2025.

2025

• Geometric Algorithms for Neural Combinatorial Optimization with Constraints [arXiv]
  Nikolaos Karalias, Akbar Rafiey, Yifei Xu, Zhishang Luo, Behrooz Tahmasebi, Connie Jiang, and Stefanie Jegelka NeurIPS 2025.
• Learning with Exact Invariances in Polynomial Time [arXiv]
  Ashkan Soleymani*, Behrooz Tahmasebi*, Stefanie Jegelka, and Patrick Jaillet ICML 2025. Spotlight paper (top 2.6% of submissions).
• Generalization Bounds for Canonicalization: A Comparative Study with Group Averaging [conference] [pdf]
  Behrooz Tahmasebi and Stefanie Jegelka ICLR 2025.
• A Robust Kernel Statistical Test of Invariance: Detecting Subtle Asymmetries [pdf]
  Ashkan Soleymani*, Behrooz Tahmasebi*, Stefanie Jegelka, and Patrick Jaillet AISTATS 2025. Oral presentation (top 2% of submissions).
• Regularity in Canonicalized Models: A Theoretical Perspective [pdf]
  Behrooz Tahmasebi and Stefanie Jegelka AISTATS 2025.

2024

• Coded Computing for Resilient Distributed Computing: A Learning-Theoretic Framework [arXiv]
  Parsa Moradi, Behrooz Tahmasebi, and Mohammad Ali Maddah-Ali NeurIPS 2024.
• A Universal Class of Sharpness-Aware Minimization Algorithms [arXiv] [conference] [workshop]
  Behrooz Tahmasebi, Ashkan Soleymani, Dara Bahri, Stefanie Jegelka, and Patrick Jaillet ICML 2024. Best Paper Award, HiLD Workshop at ICML 2024.
• Sample Complexity Bounds for Estimating Probability Divergences under Invariances [arXiv]
  Behrooz Tahmasebi and Stefanie Jegelka ICML 2024.

2023

• The Exact Sample Complexity Gain from Invariances for Kernel Regression [arXiv]
  Behrooz Tahmasebi and Stefanie Jegelka NeurIPS 2023. Spotlight paper (top 3.6% of submissions).
• The Power of Recursion in Graph Neural Networks for Counting Substructures [conference]
  Behrooz Tahmasebi, Derek Lim, and Stefanie Jegelka AISTATS 2023. Oral presentation (top 1.9% of submissions).

2021

• The Capacity of Associated Subsequence Retrieval [journal]
  Behrooz Tahmasebi, Mohammad Ali Maddah-Ali, and Seyed Abolfazl Motahari IEEE Transactions on Information Theory, 2021.