Yutao Mou (牟宇滔)

I am a first-year PhD student at KCL Lab, Peking University, and supervised by professors Wei Ye and Shikun Zhang. I received the Master and B.S. degree from Beijing University of Posts and Telecommunications, in 2024 and 2021.

I have been working on building safe, reliable and scalable artificial intelligence systems. Currently, my research interests lie in two main areas:

(1) Safety Evaluation and Red Teaming of Large Language Models (LLMs):
We focus on building comprehensive safety evaluation benchmarks across various scenarios (e.g., text generation, code agents, multimodal tasks), designing jailbreak and automated red-teaming methods, and studying the interpretability of model vulnerabilities.

(2) Post-Training and Safety Alignment of LLMs:
Our goal is to incorporate safety-related data, reward signals, and training objectives into the post-training stage of LLMs to enhance safety without compromising general capabilities. We are particularly interested in safety alignment of large reasoning models, safety-oriented reward models, and the automated synthesis and data augmentation of safety fine-tuning data. In the future, we also plan to explore multimodal safety alignment.

Feel free to contact me for communication and collaboration.

Email  /  Scholar  /  Github

Publications

In this study, we address the prevalent issue of shallow alignment in current LLM safety alignment methods by proposing the Safety-oriented Reasoning Optimization Framework (SaRO). SaRO introduces a System 2-style alignment to internalize a deliberative, slow-thinking reasoning approach, enabling models to reflect on safety policies more effectively. This framework enhances safety performance without compromising the general capabilities of the model.

In this paper, we construct a benchmark dataset named CoV-Eval, which encompasses various task types and covers a wide range of code security vulnerabilities. From the perspective of code security, we evaluate LLMs and various coding assistants, focusing on assessing the security vulnerabilities in code generated by LLMs, as well as their capabilities in vulnerability detection and repair.

This paper introduces SG-Bench, a new benchmark for evaluating LLM safety, which pioneers the concept of safety generalization—assessing how variations in task and prompt types affect model safety. SG-Bench combines both generative and discriminative tasks and expands test data to evaluate the impact of prompt engineering and jailbreak attacks

This paper first propose a new paradigm: knowledge rekindle, which aims to re-incorporate the fine-tuned expert model into the training cycle and break through the performance upper bounds of experts without introducing additional annotated data. Then we further propose a unified expert-guided pre-training (UEGP) framework for knowledge rekindle.

This paper focuses on the generalized intent discovery task, and proposes a decoupled prototype learning framework (DPL) to decouple pseudo label disambiguation and representation learning.

This paper focuses on new intent discovery and clustering task, and propose a unified K-nearest neighbor contrastive learning framework to discover OOD intents. Specifically, we design a novel K-nearest neighbor contrastive learning objective (KCL) for in-domain pre-training, and a hard negative mining strategy for self-supervised representation learning on unlabeled out-of-domain data.

This paper focuses on out-of-domain (OOD) intent detection task, and propose a unified neighborhood learning framework (UniNL) to detect OOD intents. propose a unified K-nearest neighbor contrastive learning framework to discover OOD intents. Specifically, we design a K-nearest neighbor contrastive learning objective for training and introduce a KNN-based scoring function for confidence estimation. We aim to align training objective with confidence function in inference stage.

This paper defines a new task, Generalized Intent Discovery (GID), which aims to extend an IND intent classifier to an open-world intent set including IND and OOD intents. We hope to simultaneously classify a set of labeled IND intent classes while discovering and recognizing new unlabeled OOD types incrementally. We construct three public datasets for different application scenarios and propose two kinds of frameworks, pipeline-based and end-to-end for future work.

Discovering Out-of-Domain (OOD) intents is essential for developing new skills in a task-oriented dialogue system. The key challenge is how to transfer prior IND knowledge to OOD clustering. This paper proposes a decoupled knowledge transfer framework for new intent discovery and clustering, which unifies the two-stage learning process of in-domain and out-of-domain data into instance discrimination and clustering discrimination tasks and bridges the gap between in-domain and out-of-domain data.

Internships

Selected Honors

Service