Explore our latest work on jailbreak attack initializations. We reveal how certain initializations encode a shared compliance direction in safety-aligned LLMs - enabling faster, more generalizable jailbreaks. We introduce CRI, a novel initialization framework that leverages this insight to accelerate jailbreak attacks by up to 100×. Check out our paper and code.

This work uses a novel method to uncover hidden biases in LLMs that standard prompts fail to reveal. By testing multiple models, we expose deeper discrepancies and introduce a benchmark, evaluation framework, and public leaderboard to track bias and fairness. Our goal is to improve transparency and support more informed use of LLMs in real-world applications.

As data privacy demands grow, effective machine unlearning for LLMs is becoming essential. This work evaluates how well current unlearning methods truly erase sensitive information without compromising model performance. We introduce a rigorous framework to test and compare unlearning strategies, highlighting key trade-offs between utility and true data removal.

LLM-based recommendation systems use language understanding to predict and personalize user preferences with high accuracy. In this work, we introduce a method to shift item representations within the model’s activation space, boosting recommendation rates for specified content - like new products or underexposed items - without requiring retraining or harming overall performance.

Computer vision has made major strides in semantic segmentation, with self-supervised learning helping reduce the need for labeled data. Yet, current methods struggle to model contextual relationships between distinct semantic regions. Approaches like MaskFormer and SemMAE rely on patches or heuristics, limiting object-level reasoning. Vision Transformers excel at local perception but fall short on capturing semantic interplay across objects. Our work introduces a strategy aimed at improving semantic reasoning by leveraging the structure of annotated scenes—achieving noticeable gains on ADE20K without additional compute.