Independent research,
open by default.
Xuehang Cang is an independent researcher working on LLM safety and mechanistic interpretability. The core project, PICK, locates and surgically removes refusal behavior from language models using component-level attribution and rank-1 LoRA. Everything ships in the open — code, datasets, experiment logs, and the results that didn't work.
End-to-end ownership of the research pipeline: dataset curation (safe_prompt, unsafe_prompt), model analysis (activation patching, residual stream attribution), component ablation (KL-constrained LoRA), and evaluation (refusal rate, KL divergence, capability retention). Every experiment is logged; every finding is published. The work is the work.
Selective Component Ablation for Refusal Removal. Locate the attention heads and MLP neurons that cause refusal, then surgically remove them with rank-1 LoRA.
github.com/XuehangCang/pick↗02All code, training scripts, and experiment logs. Public repositories for every project.
github.com/XuehangCang↗03Published evaluation datasets (safe_prompt, unsafe_prompt) for LLM safety benchmarking. All open-access.
huggingface.co/XuehangCang↗