Songtao Lu
Assistant Professor
Studies optimization for machine learning, including bilevel and distributed optimization.
Songtao Lu's lab at CUHK works on learning theory & optimization and distributed, cloud & edge systems. 40 papers since 2023, 29 at top venues, mostly ICML and NeurIPS. 5 PhD students. Applications close .
At a glance
- papers, 2023–2026
- 10/ year
- at T1 venues
- 73%
- PhD students
- 5
- MS students
- 0
- alumni listed
- —
Papers by yearfrom the lab homepage, 2023–2026
- 5 papers2026 · T1 3 · collecting
- 11 papers2025 · T1 8
- 13 papers2024 · T1 9
- 11 papers2023 · T1 9
T1 venuesOtherVenues: ICML, NeurIPS, ICLR, AISTATS, CCS
Research focuspaper titles that use each topic's words
- Learning theory & optimization30 papers
- optimization17 papers
- Distributed, cloud & edge systems3 papers
- distributed optimization0 papers
- studies optimization0 papers
Paper list
From the lab homepage.
20265 papers
- ICML (T1)Bilevel Optimization over Saddle Points of Zero-Sum Markov Games (opens in a new tab)
- 39-56, Academic PressGenerative Learning for Wireless Communications Fundamentals and Applications (opens in a new tab)
- ICML (T1)How Can Mamba Learn In Context with Outliers and Generalize Provably? (opens in a new tab)
- NeurIPS (T1)ParetoM3: Learning on the Pareto Set under Preference Guidance via Min-Max-Min Optimization
- SIAM/ASA Journal on Uncertainty QuantificationPCE-Net: High Dimensional Surrogate Modeling for Learning Uncertainty (opens in a new tab)
202511 papers
- JMLR (T1)Decentralized Bilevel Optimization: A Perspective from Transient Iteration Complexity (opens in a new tab)
- ICLR (T1)DUET: Decentralized Bilevel Optimization without Lower-Level Strong Convexity (opens in a new tab)
- IEEE International Conference on Acoustics, Speech, and Signal ProcessingEpigraph Based Multilevel Optimization (EMO) for Enhancing Chain-of-Thought Reasoning Capabilities (opens in a new tab)
- IEEE Journal of Selected Topics in Signal ProcessingGraph Neural Networks with Adaptive Structures (opens in a new tab)
- NeurIPS (T1)Meta-D2AG: Causal Graph Learning with Interventional Dynamic Data (opens in a new tab)
- NeurIPS (T1)Objective Soups: Multilingual Multi-Task Modeling for Speech Processing (opens in a new tab)
- NeurIPS (T1)Optimality and NP-Hardness of Transformers in Learning Markovian Dynamical Functions (opens in a new tab)
- AISTATS (T1)Q-function Decomposition with Intervention Semantics for Factored Action Spaces (opens in a new tab)
- ICLR (T1)Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis (opens in a new tab)
- ICML (T1)TSP: A Two-Sided Smoothed Primal-Dual Method for Nonconvex Bilevel Optimization (opens in a new tab)
- HiLDUnderstanding Mamba in In-Context Learning with Outliers: A Theoretical Generalization Analysis (opens in a new tab)
202413 papers
- IEEE Transactions on Big DataBiG-Fed: Bilevel Optimization Enhanced Graph-Aided Federated Learning (opens in a new tab)
- TASLP (T1)Bilevel Joint Unsupervised and Supervised Training for Automatic Speech Recognition (opens in a new tab)
- ICML (T1)Distributed Bilevel Optimization with Communication Compression (opens in a new tab)
- ICML (T1)FADAS: Towards Federated Adaptive Asynchronous Optimization (opens in a new tab)
- ICML (T1)Federated Neuro-Symbolic Learning (opens in a new tab)
- 181-197, ElsevierGraph-Aware Federated Learning (opens in a new tab)
- TF2MHow Do Nonlinear Transformers Acquire Generalization-Guaranteed CoT Ability? (opens in a new tab)
- ICML (T1)How Do Nonlinear Transformers Learn and Generalize in In-Context Learning? (opens in a new tab)
- ICLR (T1)PILOT: An O(1/K)-Convergent Approach for Policy Evaluation with Nonlinear Function Approximation (opens in a new tab)
- ICML (T1)SF-DQN: Provable Knowledge Transfer using Successor Feature for Deep Reinforcement Learning (opens in a new tab)
- NeurIPS (T1)SPARKLE: A Unified Single-Loop Primal-Dual Framework for Decentralized Bilevel Optimization (opens in a new tab)
- Computational Optimization and ApplicationsStochastic Inexact Augmented Lagrangian Method for Nonconvex Expectation Constrained Optimization (opens in a new tab)
202311 papers
- NeurIPS (T1)A Generalized Alternating Method for Bilevel Optimization under the Polyak-Łojasiewicz Condition (opens in a new tab)
- ICML (T1)Bilevel Optimization with Coupled Decision-Dependent Distributions (opens in a new tab)
- ICML (T1)Compressed Decentralized Proximal Stochastic Gradient Method for Nonconvex Composite Problems with Heterogeneous Data (opens in a new tab)
- AISTATS (T1)Distributed Offline Policy Optimization Over Batch Data (opens in a new tab)
- ICLR (T1)Joint Edge-Model Sparse Learning is Provably Efficient for Graph Neural Networks (opens in a new tab)
- IEEE International Conference on Acoustics, Speech and Signal ProcessingMeta-DAG: Meta Causal Discovery via Bilevel Optimization (opens in a new tab)
- ICLR (T1)Min-Max Multi-Objective Bilevel Optimization with Applications in Robust Machine Learning (opens in a new tab)
- NeurIPS (T1)On the Convergence and Sample Complexity Analysis of Deep Q-Networks with ε-Greedy Exploration (opens in a new tab)
- ICML (T1)Prometheus: Taming Sample and Communication Complexities in Constrained Decentralized Stochastic Bilevel Learning (opens in a new tab)
- NeurIPS (T1)SLM: A Smoothed First-Order Lagrangian Method for Structured Constrained Nonconvex Optimization (opens in a new tab)
- M3LTransformers as Multi-Task Feature Selectors: Generalization Analysis of In-Context Learning (opens in a new tab)
Where alumni went
Alumni are not listed on the lab homepage.
Apply to CUHKFall 2027
Computer Science and Engineering MPhil/PhD
- Round
- HKPFS / main round
- Opens
- —
- GRE
- See program page
- English
- Interview in English
- Fee
- See program page
Not yet confirmed for this intake.
Similar labsnearest by research summary
- USUSC
Matteo Sesia
- Learning theory & optimization
- GBImperial
Matthew Wicker
- Trustworthy & safe ML
- USMIT
Dylan Hadfield-Menell
- Trustworthy & safe ML
- USMichigan
Ji Zhu
- Learning theory & optimization
- USUT Austin
Inderjit Dhillon
- Recognition, detection & segmentation
- USNYU
Jaume Vives-i-Bastida
- Reasoning & neuro-symbolic AI
Your notes
Save this lab to keep notes.
Data
Some numbers could not be verified:
- The alumni list could not be found on the lab site.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Not collected
Tiers are venue classifications, not lab rankings.
CUHKcloses