6 papers
Logic Jailbreak: Efficiently Unlocking LLM Safety Restrictions Through Formal Logical Expression
Jingyu Peng, Maolin Wang, Nan Wang +7
Despite substantial advancements in aligning large language models (LLMs) with human values, current safety mechanisms remain susceptible to jailbreak attacks. We hypothesize that…
AdaSwitch: Balancing Exploration and Guidance in Knowledge Distillation via Adaptive Switching
Jingyu Peng, Maolin Wang, Hengyi Cai +5
Small language models (SLMs) are crucial for applications with strict latency and computational constraints, yet achieving high performance remains challenging. Knowledge distillat…
How to Utilize Complementary Vision-Text Information for 2D Structure Understanding
Jiancheng Dong, Pengyue Jia, Derong Xu +9
LLMs typically linearize 2D tables into 1D sequences to fit their autoregressive architecture, which weakens row-column adjacency and other layout cues. In contrast, purely visual…
MOSAIC: Composable Safety Alignment with Modular Control Tokens
Jingyu Peng, Hongyu Chen, Jiancheng Dong +5
Safety alignment in large language models (LLMs) is commonly implemented as a single static policy embedded in model parameters. However, real-world deployments often require conte…
Behavior-Equivalent Token: Single-Token Replacement for Long Prompts in LLMs
Jiancheng Dong, Pengyue Jia, Jingyu Peng +7
Carefully engineered system prompts play a critical role in guiding the behavior of LLM agents, but their considerable length introduces significant drawbacks, including increased…
Stepwise Reasoning Error Disruption Attack of LLMs
Jingyu Peng, Maolin Wang, Xiangyu Zhao +6
Large language models (LLMs) have made remarkable strides in complex reasoning tasks, but their safety and robustness in reasoning processes remain underexplored. Existing attacks…