Lun Wang
I am a staff research scientist at Google DeepMind, working on Gemini post-training (Memory, Tool Use, and Audio). I received my PhD in computer science from UC Berkeley, advised by Prof. Dawn Song in 2022. Previously, I earned my Bachelor's degree with honors in computer science from Peking University in 2018.
Selected Publications
- Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process.
- MUSIC: MUlti-Step Instruction Contrast for Multi-Turn Reward Models.
- Can DeepFake Speech be Reliably Detected?
- Attacker's Noise Can Manipulate Your Audio-based LLM in the Real World.
EACL 2026: The 19th Conference of the European Chapter of the Association for Computational Linguistics. - Revisit Micro-batch Clipping: Adaptive Data Pruning via Gradient Manipulation
ICLR 2025: The Thirteenth International Conference on Learning Representations. - Efficiently Train ASR Models that Memorize Less and Perform Better with Per-core Clipping.
Interspeech 2024: 25th Interspeech Conference. - Unintended Memorization in Large ASR Models, and How to Mitigate It.
- Why Is Public Pretraining Necessary for Private Model Training?
- Secure Federated Correlation Test and Entropy Estimation.
ICML 2023: Fortieth International Conference on Machine Learning. - Byzantine-Robust Federated Learning with Optimal Rates and Privacy Guarantee.
AISTATS 2023: Artificial Intelligence and Statistics 2023. - Differentially Private Fractional Frequency Moments Estimation with Polylogarithmic Space.
ICLR 2022: the 10th International Conference on Learning Representations. - PRIVGUARD: Privacy Regulation Compliance Made Easier.
Usenix Security'22: the 31th Usenix Security Symposium. - BACKDOORL: Backdoor Attack against Competitive Reinforcement Learning.
IJCAI'21: the 30th International Joint Conference on Artificial Intelligence. - Towards practical differentially private causal graph discovery.
NeurIPS'20: 34th Annual Conference on Neural Information Processing Systems. - Towards Inspecting and Eliminating Trojan Backdoors in Deep Neural Networks.
ICDM'20: 20th IEEE International Conference on Data Mining. - DUET: An expressive higher-order language and linear type system for statically enforcing differential privacy.
- Towards practical differentially private convex optimization.
Oakland'19: 2019 IEEE Symposium on Security and Privacy.
Talks
- AAAI Spring Series Symposium 2024 @ Stanford: Reconcile Byzantine-robustness & Privacy in Federated Learning
- Federated Learning & Analytics Workshop 2021 @ Google: Federated Frequency Moments Estimation & Its Application
- 2020 IC3 Blockchain Camp: Aegis: Privacy-Preserving Data Analysis Framework
- 2019 San Francisco Blockchain Week: Dynamic-Committee Proactive Secret Sharing
- 2019 Large-Scale Consensus and Blockchains @ Simons Institute: CHURP: Dynamic-Committee Proactive Secret Sharing
Service
- Area Chair: ACL ARR 2025
- Program Committee: Usenix Security 2026, CCS 2025, CCS 2023
- Workshop Committee: WMARK'25, FL-AsiaCCS’25, FedKDD 2024, TPDP 2024, FL @ ICML 2023, SPAI 2020, WNA 2020
- Reviewer: Neurips, ICML, ICLR, AISTATS, Interspeech, ITCS'23, RANDOM'23, Journal of Causal Inference, IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Dependable and Secure Computing, Transactions on Machine Learning Research, Frontiers in Computer Science