Xinpeng Wang

I am a PhD student at the Munich AI & NLP (MaiNLP) research lab at LMU Munich . My supervisor is Prof. Barbara Plank . I'm currently a visiting researcher at New York University advised by Prof. He He .

Previously, I completed my M.Sc. degree in Robitics, Cognition, Intelligence at Technical University of Munich, where I was a student researcher at Visual Computing & AI lab at TUM working on indoor scene synthesis. I was also a teaching assistant of the course Introduction to Deep Learning (IN2346).

My research currently focuses on Human-Centric AI and Alignment.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  Twitter  /  CV

Research

Refusal Direction is Universal Across Safety-Aligned Languages

Refusal directions in LLMs work across languages, revealing shared jailbreak mechanisms and raising the need for stronger multilingual safety.

Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation

We propose a surgical and flexible approach to mitigate the false refusal in LLMs with minimal effect on performance and inference cost.

Seeing the Big through the Small: Can LLMs Approximate Human Judgment Distributions on NLI from a Few Explanations?

This study proposes to exploit LLMs to approximate human judgment distributions using a small number of expert labels and explanations on NLI.

The Potential and Challenges of Evaluating Attitudes, Opinions, and Values in Large Language Models

We review recent research on evaluating Attitudes, Opinions, and Values in LLMs, highlighting the potential and challenges and offering suggestions for future research.

Look at the Text: Instruction-Tuned Language Models are More Robust Multiple Choice Selectors than You Think

We showed that text answers are more robust than first token answer in instruction-tuned language models, even debiased with SOTA first-token debiasing method.

"My Answer is C": First-Token Probabilities Do Not Match Text Answers in Instruction-Tuned Language Models

We showed that the first-token probability evaluation does not match text answers in instruction-tuned language models.

ACTOR: Active Learning with Annotator-specific Classification Heads to Embrace Human Label Variation

We proposed an active learning framework that utilizes a muli-head model to model individual annotators. We designed different acquisition functions and showed our active learning setup achieved performance comparable to full-scale training while saving up to 70% of the annotation budget.

How to Distill your BERT: An Empirical Study on the Impact of Weight Initialisation and Distillation Objectives

We showed that using lower teacher layers for pre-loading student model gives significant performance improvement compared to higher layers. We also studied the robustness of different distillation objectives under various initialisation choices.

Sceneformer: Indoor Scene Generation with Transformers

We proposed a transformer model for scene generation conditioned on room layout and text description.

Projects

These include coursework and practical course projects.

Domain Specific Multi-Lingually Aligned Word Embeddings

Curiosity Driven Reinforcement Learning

Evaluated and compared the count-based and prediction-based curiosity driven learning on different Atari game environments.

Teaching

Introduction to Deep Learning (IN2346)

Design and source code from Jon Barron's website