Publications
No matching publications. Try another keyword.
2026

LEAF-OPD: Learning to Forget from Privileged Contexts via On-Policy Distillation
Zirui Pang, Zhijie Deng, Yuancheng Wang, Qihan Yu, Xin Li, Jiaheng Wei, Zixin Zhong
TL;DR: We propose LEAF-OPD, an LLM unlearning method that distills a teacher into a student with privileged context on forget data and without it on retain data.

Replay What Matters: Off-Policy Replay for Efficient LLM Reinforcement Unlearning
Zirui Pang, Chenlong Zhang, Haosheng Tan, Zhuoran Jin, Jiaheng Wei, Zixin Zhong
TL;DR: We propose ReRULE, which improves RULE by replaying hard cases for more efficient and effective unlearning.

CAMEO: A Conditional and Quality-Aware Multi-Agent Image Editing Orchestrator
Yuhan Pu, Hao Zheng, Ziqian Mo, Zirui Pang, Hill Zhang, Tianyi Fan, Shuhong Wu, Jiaheng Wei
TL;DR: We propose CAMEO, a structured multi-agent framework that reframes conditional image editing as a quality-aware, feedback-driven process and improves average win rate by 20% over state-of-the-art models.
2025

Label Smoothing Improves Gradient Ascent in LLM Unlearning
Zirui Pang, Hao Zheng, Zhijie Deng, Ling Li, Zixin Zhong, Jiaheng Wei
TL;DR: We propose Smoothed Gradient Ascent (SGA), a Gradient Ascent method improved with generalized label smoothing, which achieves both strong forgetting quality and high model utility in LLM unlearning tasks.

OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models
Hao Zheng, Zirui Pang, Ling Li, Zhijie Deng, Yuhan Pu, Zhaowei Zhu, Xiaobo Xia, Jiaheng Wei
TL;DR: OFFSIDE benchmarks multimodal unlearning with real-world football transfer rumors, revealing that current methods struggle to forget visual misinformation and remain vulnerable to recovery and prompt attacks.

GUARD: Generation-time LLM Unlearning via Adaptive Restriction and Detection
Zhijie Deng, Chris Yuhao Liu, Zirui Pang, Xinlei He, Lei Feng, Qi Xuan, Zhaowei Zhu, Jiaheng Wei
TL;DR: We propose Generation-time Unlearning via Adaptive Restriction and Detection (GUARD), a framework that enables dynamic unlearning during LLM generation.

When VLMs Meet Image Classification: Test Sets Renovation via Missing Label Identification
Zirui Pang, Haosheng Tan, Yuhan Pu, Zhijie Deng, Zhouan Shen, Keyu Hu, Jiaheng Wei
TL;DR: We investigate the limitations of VLM in image classification task, and propose a VLM-based test-data renovation pipeline that detects potential label errors and missing labels.