Steven Wu
Associate Professor
Studies trustworthy machine learning, including privacy, AI safety, interactive learning and reinforcement learning.
Steven Wu's lab at CMU works on trustworthy & safe ML, reinforcement learning and interpretability, evaluation & safety of LMs. 71 papers since 2023, 52 at top venues, mostly ICML and NeurIPS. 8 PhD students. 16 alumni, most often to Amazon and at J.P. Morgan AI Research. Applications close .
At a glance
- papers, 2023–2026
- 19/ year
- at T1 venues
- 73%
- PhD students
- 8
- MS students
- 0
- alumni listed
- 16
Papers by yearfrom the lab homepage, 2023–2026
- 14 papers2026 · T1 14 · collecting
- 22 papers2025 · T1 16
- 16 papers2024 · T1 6
- 19 papers2023 · T1 16
T1 venuesOtherVenues: ICML, NeurIPS, ICLR, COLT, EMNLP
Research focuspaper titles that use each topic's words
- Reinforcement learning17 papers
- Trustworthy & safe ML1 paper
- Interpretability, evaluation & safety of LMs1 paper
- interactive learning0 papers
- studies trustworthy0 papers
Paper list
From the lab homepage.
202614 papers
- ICLR (T1)All Roads Lead to Likelihood: The Value of Reinforcement Learning in Fine-Tuning (opens in a new tab)
- NeurIPS (T1)Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning (opens in a new tab)
- ICLR (T1)Dimension-Free Decision Calibration for Nonlinear Loss Functions (opens in a new tab)
- ICLR (T1)Doubly-Robust LLM-as-a-Judge: Externally Valid Estimation with Imperfect Personas (opens in a new tab)
- ICLR (T1)From Curiosity to Caution: Mitigating Reward Hacking for Best-of-N with Pessimism (opens in a new tab)
- ACL (T1)Gained in Translation: Privileged Pairwise Judges Enhance Multilingual Reasoning (opens in a new tab)
- S&P (T1)Generate-then-Verify: Reconstructing Data from Limited Published Statistics (opens in a new tab)
- UAI (T1)Improving the Convergence of Private Shuffled Gradient Methods with Public Data (opens in a new tab)
- PNAS (T1)Many AI Analysts, One Dataset: Navigating the Agentic Data Science Multiverse (opens in a new tab)
- ICML (T1)Membership Inference Attacks for Unseen Classes (opens in a new tab)
- ICLR (T1)Nearly-Optimal Bandit Learning in Stackelberg Games with Side Information (opens in a new tab)
- ICLR (T1)Optimizing Canaries for Privacy Auditing with Metagradient Descent (opens in a new tab)
202522 papers
- KDD (T1)Benchmarking Fraud Detectors on Private Graph Data (opens in a new tab)
- The IEEE Secure and Trustworthy Machine Learning ConferenceChoosing Public Datasets for Private Machine Learning via Gradient Subspace Distance (opens in a new tab)
- ACM Transactions on Economics and ComputationCompeting Bandits: The Perils of Exploration Under Competition (opens in a new tab)
- WWW (T1)Differentially Private Bayesian Persuasion (opens in a new tab)
- NeurIPS (T1)Discretization-free Multicalibration through Loss Minimization over Tree Ensembles (opens in a new tab)
- WWW (T1)Inferentially-Private Private Information (opens in a new tab)
- ICML (T1)Kandinsky Conformal Prediction: Beyond Class- and Covariate-Conditional Coverage (opens in a new tab)
- ICML (T1)Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models (opens in a new tab)
- The ACM Conference on Computer-Supported Cooperative Work and Social ComputingMeasurement as Bricolage: Examining How Data Scientists Construct Target Variables for Predictive Modeling Tasks (opens in a new tab)
- UAI (T1)Multi-group Uncertainty Quantification for Long-form Text Generation (opens in a new tab)
- COLT (T1)Orthogonal Causal Calibration (opens in a new tab)
- EMNLP (T1)Persona-Augmented Benchmarking: Evaluating LLMs Across Diverse Writing Styles (opens in a new tab)
202416 papers
- ICML (T1)A Minimaximalist Approach to Reinforcement Learning from Human Feedback (opens in a new tab)
- The Thiry-Eighth Annual Conference on Neural Information Processing SystemsBridging Multicalibration and Out-of-distribution Generalization Beyond Covariate Shift (opens in a new tab)
- ICLR (T1)Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach (opens in a new tab)
- The International Conference on Web and Internet EconomicsEx-post Individually Rational Bayesian Persuasion (opens in a new tab)
- The IEEE Conference on Secure and Trustworthy Machine LearningFair Federated Learning via Bounded Group Loss (opens in a new tab)
- ICML (T1)Hybrid Inverse Reinforcement Learning (opens in a new tab)
- The IEEE Conference on Secure and Trustworthy Machine LearningImproved Differentially Private Regression via Gradient Boosting (opens in a new tab)
- ICML (T1)Membership Inference Attacks on Diffusion Models via Quantile Regression (opens in a new tab)
- The Thiry-Eighth Annual Conference on Neural Information Processing SystemsMulti-Agent Imitation Learning: Value is Easy, Regret is Hard (opens in a new tab)
- The Thiry-Eighth Annual Conference on Neural Information Processing SystemsOn the Benefits of Public Representations for Private Transfer Learning under Distribution Shift (opens in a new tab)
- The Thiry-Eighth Annual Conference on Neural Information Processing SystemsOracle-Efficient Differentially Private Learning with Public Data (opens in a new tab)
- ICML (T1)Predictive Performance Comparison of Decision Policies Under Confounding (opens in a new tab)
202319 papers
- NeurIPS (T1)Adaptive Principal Component Regression with Applications to Panel Data (opens in a new tab)
- NeurIPS (T1)Adaptive Privacy Composition for Accuracy-first Mechanisms (opens in a new tab)
- PNAS (T1)Confidence-ranked reconstruction of census microdata from published statistics (opens in a new tab)
- FAccTCounterfactual Prediction Under Outcome Measurement Error (opens in a new tab)
- ICML (T1)Fully-Adaptive Composition in Differential Privacy (opens in a new tab)
- ICML (T1)Generating Private Synthetic Data with Genetic Algorithms (opens in a new tab)
- SIAM Journal on Computing (T1)Greedy Algorithm Almost Dominates in Smoothed Contextual Bandits (opens in a new tab)
- FAccTGround(less) Truth: A Causal Framework for Proxy Labels in Human-Algorithm Decision-Making (opens in a new tab)
- ICML (T1)Inverse Reinforcement Learning without Reinforcement Learning (opens in a new tab)
- NeurIPS (T1)Learning Shared Safety Constraints from Multi-task Demonstrations (opens in a new tab)
- NeurIPS (T1)Meta-Learning Adversarial Bandit Algorithms (opens in a new tab)
- ICLR (T1)Meta-Learning in Games (opens in a new tab)
Where alumni went16 listed, 23% to industry
Industry 3Academia 10
Most often to
- Amazon
- at J.P. Morgan AI Research
Also to
- Stanford University
- Cornell University
- on the Vassar College CS Faculty
Apply to CMUFall 2027
CMU SCS Computer Science PhD
- Round
- Single round
- Opens
- GRE
- See program page
- English
- TOEFL / IELTS / Duolingo (F-1/J-1 non-native speakers)
- Fee
- USD 100 (USD 80 by early deadline)
Confirmed on the program page.
Similar labsnearest by research summary
- USUMD
Furong Huang
- Generative & foundation models
- HKCUHK
Weiyang Liu
- Generative & foundation models
- CNPeking
Tongyang Li
- Learning theory & optimization
- USUIUC
Richard Y. Zhang
- Learning theory & optimization
- SGNUS
Jiancong Xiao
- Trustworthy & safe ML
- USStanford
Ellen Vitercik
- Learning theory & optimization
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.
CMUcloses