4 papers
Resisting Harmful Instructions through Training-Free Weight Editing
Ali Ebrahimpour-Boroojeny, Xiaocong Yang, Samuel Schapiro +4
Instruction-following language models can comply with requests that conflict with intended behaviors and safety alignment. We introduce a training-free weight-editing approach for…
You Can't Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning
Yian Wang, Ali Ebrahimpour-Boroojeny, Hari Sundaram +1
Concept unlearning in text-to-image diffusion models aims to suppress a target concept (e.g., \texttt{horse}) while preserving related but distinct content (e.g., \texttt{donkey}),…
AgentAbstain: Do LLM Agents Know When Not to Act?
Xun Liu, Yi Evie Zhang, Vira Kasprova +5
Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, yet existing evaluations mostly focus on task success rather than whether agents…
On the Necessity of Output Distribution Reweighting for Effective Class Unlearning
Ali Ebrahimpour-Boroojeny, Yian Wang, Hari Sundaram
In this paper, we reveal a significant shortcoming in class unlearning evaluations: overlooking the underlying class geometry can cause information leakage about the forgotten clas…