BSO: Safety Alignment Is Density Ratio Matching
Tien-Phat Nguyen, Truong Nguyen, Thin Nguyen, Duy Minh Ho Nguyen, Ngoc-Thanh Dinh, Trung Le
arXiv preprint arXiv:2605.12339
All research directions
Human-AI Alignment
Preference optimization, safety alignment, and token-based methods for aligning foundation models with human intent.
Goals
To build principled alignment methods that steer powerful models toward human preferences and safety constraints without sacrificing capability.
Overview
This direction develops theory and algorithms for aligning large language and multimodal models with human values. I aim to develop helpful, secure, and smart foundation models aligned with multiple human preferences, values, and cultures.
Key topics
- Preference optimization and RLHF
- Safety alignment for safer models without hurting model usefulness
- Off-policy reference tuning
- Multi-cultural alignment
- Model sovereignty
- Multi-objective preference alignments
News
Paper TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching accepted at ICML 2026
Paper f-Divergence Self-Play for Tabular Anomaly Detection via Large Language Models accepted at ICML 2026
Paper Sharpness-Aware Minimization in Logit Space Efficiently Enhances Direct Preference Optimization accepted at ICLR 2026
Paper CTPD: Cross Tokenizer Preference Distillation accepted at AAAI 2025
Paper Token-Level Self-Play with Importance-Aware Guidance for Large Language Models accepted at NeurIPS 2025
Papers in this direction
f-Divergence Self-Play for Tabular Anomaly Detection via Large Language Models
Vuong Hoang Tran, Van Linh Ngo, Dang Nguyen, Thin Nguyen, Phuoc Nguyen, Mehrtash Harandi, Trung Le
International Conference on Machine Learning (ICML 2026)TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching
Truong Nguyen, Tien-Phat Nguyen, Linh Ngo Van, Duy Minh Ho Nguyen, Khoa D. Doan, Trung Le
International Conference on Machine Learning (ICML 2026)CTPD: Cross Tokenizer Preference Distillation
Truong Nguyen, Van-Phi Dat, Ngan Nguyen, Van-Linh Ngo, Trung Le, Hong-Thanh Nguyen
AAAI Conference on Artificial Intelligence (AAAI 2026)Token-Level Self-Play with Importance-Aware Guidance for Large Language Models
Tue Le, Hoang Tran Vuong, Quyen Tran, Linh Ngo Van, Mehrtash Harandi, Trung Le
Conference on Neural Information Processing Systems (NeurIPS 2025)