8 papers
Dynamic Optimization and Safety Indicator Injection for Jailbreaking Text-to-Image Models with Multimodal Safety Filters
Zixuan Chen, Hao Lin, Ke Xu +2
Text-to-image (T2I) models can generate not-safe-for-work (NSFW) content, motivating multi-stage safety pipelines with both text and image filters. Newer LLM-based filters detect l…
Stabilizing Reinforcement Learning with LLMs: Formulation and Practices
Chujie Zheng, Kai Dang, Bowen Yu +10
This paper proposes a novel formulation for reinforcement learning (RL) with large language models, explaining why and under what conditions the true sequence-level reward can be o…
Domain-Aware RAG: MoL-Enhanced RL for Efficient Training and Scalable Retrieval
Hao Lin, Peitong Xie, Jingxue Chen +3
Retrieval-Augmented Generation (RAG) systems rely heavily on the retrieval stage, particularly the coarse-ranking process. Existing coarse-ranking optimization approaches often str…
MEGA: Second-Order Gradient Alignment for Catastrophic Forgetting Mitigation in GFSCIL
Jinhui Pang, Changqing Lin, Hao Lin +4
Graph Few-Shot Class-Incremental Learning (GFSCIL) enables models to continually learn from limited samples of novel tasks after initial training on a large base dataset. Existing…
Multi-task Adversarial Attacks against Black-box Model with Few-shot Queries
Wenqiang Wang, Yan Xiao, Hao Lin +2
Current multi-task adversarial text attacks rely on abundant access to shared internal features and numerous queries, often limited to a single task type. As a result, these attack…
No Query, No Access
Wenqiang Wang, Siyuan Liang, Yangshijie Zhang +3
Textual adversarial attacks mislead NLP models, including Large Language Models (LLMs), by subtly modifying text. While effective, existing attacks often require knowledge of the v…