Song Mei
Assistant Professor
Studies the theory of deep learning and transformers using statistics and statistical physics.
Song Mei's lab at UC Berkeley works on learning theory & optimization. 20 papers since 2023, 9 at top venues, mostly NeurIPS and ICLR. 4 PhD students. 5 alumni. Applications close .
At a glance
- papers, 2023–2026
- 5/ year
- at T1 venues
- 45%
- PhD students
- 4
- MS students
- 0
- alumni listed
- 5
Papers by yearfrom the lab homepage, 2023–2026
- 0 papers2026 · T1 0 · collecting
- 9 papers2025 · T1 4
- 6 papers2024 · T1 2
- 5 papers2023 · T1 3
T1 venuesOtherVenues: NeurIPS, ICLR
Research focuspaper titles that use each topic's words
- Learning theory & optimization13 papers
- transformers4 papers
- statistics1 paper
- statistical physics0 papers
- physics0 papers
Paper list
From the lab homepage.
20259 papers
- NeurIPS (T1)A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics
- IEEE Transactions on Information Theory, +Deep Networks as Denoising Algorithms: Sample-Efficient Learning of Diffusion Models in High-Dimensional Graphical Models
- NeurIPS (T1)Generalization or Hallucination? Understanding Out-of-Context Reasoning in Transformers
- International Conference on Machine Learning (ICML)Implicit Bias of Gradient Descent for Non-Homogeneous Deep Networks
- International Conference on Machine Learning (ICML)Improving LLM Safety Alignment with Dual-Objective Optimization
- NeurIPS Datasets & Benchmarks TrackOVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models
- NeurIPS (T1)Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning
- ICLR (T1)U-Nets as Belief Propagation: Efficient Classification, Denoising, and Diffusion in Generative Hierarchical Models
- To appear in the Annals of Statistics, +Unified Algorithms for RL with Decision-Estimation Coefficients: No-Regret, PAC, and Reward-Free Learning
20246 papers
- National Science ReviewAn Overview of Diffusion Models: Applications, Guided Generation, Statistical Rates and Optimization
- International Conference on Learning Representations (ICLR)How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations
- NeurIPS (T1)Large Stepsize Gradient Descent for Non-Homogeneous Two-Layer Networks: Margin Improvement and Fast Optimization
- Conference on Language Modeling (COLM)Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
- NeurIPS (T1)Statistical Estimation in the Spiked Tensor Model via the Quantum Approximate Optimization Algorithm
- International Conference on Learning Representations (ICLR)Transformers as Decision Makers: Provable In-Context Reinforcement Learning via Supervised Pretraining
20235 papers
- The Annals of StatisticsLocal convexity of the TAP free energy and AMP convergence for Z2-synchronization
- International Conference on Machine Learning (ICML)Lower Bounds for Learning in Revealing POMDPs
- ICLR (T1)Partially Observable RL with B-Stability: Unified Structural Condition and Sharp Sample-Efficient Algorithms
- NeurIPS (T1)Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection
- NeurIPS (T1)What can a Single Attention Layer Learn? A Study Through the Random Features Lens
Where alumni went5 listed, 0% to industry
Industry 0Academia 2
Also to
- Caltech
- Flatiron institute
Apply to UC BerkeleyFall 2027
UC Berkeley EECS Computer Science PhD
- Round
- Single round
- Opens
- —
- GRE
- not required
- English
- TOEFL/IELTS (non-English-country degrees)
- Fee
- See program page
Confirmed on the program page.
Similar labsnearest by research summary
- USNYU
Shengjie Wang
- Learning theory & optimization
- USUC Berkeley
Somayeh Sojoudi
- Trustworthy & safe ML
- USMichigan
Clayton Scott
- Learning theory & optimization
- USUT Austin
Aryan Mokhtari
- Learning theory & optimization
- USUCSD
Daniel M. Kane
- Learning theory & optimization
- USUCLA
Yuan Tian
- Trustworthy & safe ML
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.
UC Berkeleycloses