Minjun Zhu

SafeReview: Building a Robust Deep Review Assistant Against Prompt Injection

DeepScientist: Advancing Frontier-Pushing Scientific Findings Progressively

AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations

Safetylock: Guarding LLM againt FuneTuning Risks with Efficient Inference-time Addon

CycleResearcher: Improving Automated Research via Automated Review

Personality Alignment of Large Language Models

Direct Preference Optimization Using Sparse Feature-level Constraints

Constrain Alignment with Sparse Autoencoders

Locking Down the Finetuned LLMs Safety