Attention Smoothing Unlearning
A framework that constructs a forget-teacher by smoothing a model’s attention, then uses self-distillation to suppress unwanted factual recall while preserving coherent responses and model utility.
Projects
Research in practice.
A project-oriented view of the methods and questions represented in my publication record.
Automatic Calibration for Membership Inference
Research on automatically calibrating membership inference attacks for large language models.
Not All Tokens Are Meant to Be Forgotten
Work investigating token-aware approaches to forgetting in large language models.
Adversarial In-Context Learning
Research examining how adversarial demonstrations in context can hijack large language models.
Safety Alignment in LRMs
Ongoing work on deceptive safety alignment and counter-aligned few-shot conversation exposure in large reasoning models.
Manuscripts under review
Downstream Manipulation
Research on poisoning large language models during training for downstream behavioral manipulation.
Read the preprint Under review