Yihan Wang

The paper introduces 'infilling extraction' to accurately model training data memorization in Diffusion Language Models (DLMs), finding that bidirectional masking significantly increases the extractability of verbatim training data compared to traditional prefix-only methods.

The paper introduces the threat model of sequential data poisoning, demonstrating that multiple, collaborating attackers can exploit compound vulnerabilities in LLM post-training pipelines that are invisible when analyzing individual stages.

The paper proposes a framework to harvest unused computation resources on AI chips for general-purpose tasks using neural architecture search and approximation techniques.