Research
I'm interested in scaling multi-agent approaches to improve downstream task performance and expanding the real-world use cases of agentic workflows, as well as exploring critical areas of LLM safety through interpretability.
|
|
SteeringControl: Holistic Evaluation of Alignment Steering in LLMs
Vincent Siu*, Nicholas Crispino*, David Park, Nathan W. Henry, Zhun Wang, Yang Liu, Dawn Song, Chenguang Wang
in arXiv:2509.13450, 2025
arxiv
/
code
/
dataset
SteeringControl is a benchmark measuring the behavioral entanglement that occurs on out of distribution datasets when using current representation steering methods to improve performance on a single behavior, along with a modular code framework allowing for standardization of training-free steering methods.
|
|
RepIt: Representing Isolated Targets to Steer Language Models
Vincent Siu, Nathan W. Henry, Nicholas Crispino, Yang Liu, Dawn Song, Chenguang Wang
in arXiv:2509.13281, 2025
arxiv
/
code
RepIt isolates concept-specific representations of refusal that can be used to suppress refusal on target concepts while preserving the baseline model’s refusal on other concepts using steering methods in interpretability.
|
|
COSMIC: Generalized Refusal Identification in LLM Activations
Vincent Siu, Nicholas Crispino, Zihao Yu, Sam Pan, Zhun Wang, Yang Liu, Dawn Song, Chenguang Wang
in Findings of the Association for Computational Linguistics (ACL), 2025
arxiv
/
code
COSMIC improves the direction selection step of steering refusal in LLMs by choosing a direction maximizing the cosine similarity between paired harmless and harmful behavior, allowing for the more robust application of activation steering.
|
|
MLAN: Language-Based Instruction Tuning Improves Zero-Shot Generalization of Multimodal Large Language Models
Jianhong Tu, Zhuohao Ni, Nicholas Crispino, Zihao Yu, Michael Bendersky, Beliz Gunel, Ruoxi Jia, Xin Liu, Lingjuan Lyu, Dawn Song, Chenguang Wang
in arXiv preprint, 2024
arxiv
/
code
MLAN proposes focusing on text-only instances in instruction tuning to improve instruction following in both the vision and language modalities in multimodal large language models at a lower cost than traditional vision-only or vision-based methods.
|
|
Agent Instructs Large Language Models to be General Zero-Shot Reasoners
Nicholas Crispino, Kyle Montgomery, Fankun Zeng, Dawn Song, Chenguang Wang
in Forty-first International Conference on Machine Learning (ICML), 2024
arxiv
/
code
Zero-shot AgentInstruct uses an agent to generate dataset-specific instructions to improve the zero-shot performance of instruction-following large language models.
|
|
