Scholar
Mickel Liu
Google Scholar ID: 2oog2ZcAAAAJ
University of Washington
Citations & Impact
Citations
2,080
H-index
6
i10-index
6
Publications
8
Co-authors
7
list available
Publications
Resume (English only)
Academic Achievements
- Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models (Preprint, equal contribution): Used self-play RL and hidden Chain-of-Thought to discover diverse adversarial attacks for safer LLM alignment
- BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset (NeurIPS 2023, equal contribution): Introduced a human-preference dataset showing that decoupling helpfulness and harmlessness improves safety without performance loss
- Safe RLHF: Safe Reinforcement Learning from Human Feedback (ICLR 2024 Spotlight): Proposed a constrained RLHF algorithm using Lagrangian methods to balance harmlessness and helpfulness, outperforming existing alignment methods
- Baichuan 2: Open large-scale language models (Technical Report, author): Contributed to open-sourcing Baichuan2 models achieving state-of-the-art results among open-source models on benchmarks like MMLU, CMMLU, GSM8K, HumanEval, and SuperCLUE-agent
- Proactive Multi-Camera Collaboration For 3D Human Pose Estimation (ICLR 2023, equal contribution): Developed a multi-agent RL framework for collaborative 3D pose estimation in dynamic crowds using Shapley-value-inspired rewards
