Research

My research program is driven by a central question: How can we build AI systems that are technically powerful, legally compliant, and ethically accountable? To address this, my group pursues five interconnected research pillars — from foundational AI safety and efficient model design to applied legal reasoning and computational social science.


Hong Kong’s unique position as a jurisdiction operating under two parallel legal systems — common law and mainland Chinese law — makes it the world’s only natural laboratory for studying how AI systems can be aligned to plural legal norms simultaneously. This pillar focuses on developing benchmarks, methods, and frameworks for ensuring LLMs comply with diverse legal requirements.

Key Contributions:

ACL 2025 Findings

SafeLawBench: Towards Safe Alignment of Large Language Models

Cao, Chuxue; Zhu, Han; Ji, Jiaming; ...; Han, Sirui*; Guo, Yike*

A comprehensive benchmark for evaluating the legal safety alignment of LLMs across multiple jurisdictions.

ACL 2025

LegalReasoner: Step-wised Verification-Correction for Legal Judgment Reasoning

Shi, Weijie; Zhu, Han; ...; Han, Sirui*; Guo, Yike*.

A step-wise verification-correction framework that significantly improves the accuracy and reliability of LLM-based legal judgment reasoning.

ACL 2025

PrivaCI-Bench: Evaluating Privacy with Contextual Integrity and Legal Compliance

Li, Haoran; Hu, Wenbin; ...; Han, Sirui*; Chu, Tianshu; Hu, Peizhao; Song, Yangqiu.

A benchmark evaluating LLM privacy awareness through the lens of contextual integrity theory and legal compliance.

Other Related Work:

  • Trustworthy Legal Reasoning: A Comprehensive SurveyPreprints 2026Paper
  • Benchmarking Multi-National Value Alignment for LLMsACL 2025 FindingsarXiv
  • LRAS: Advanced Legal Reasoning with Agentic SearcharXiv 2026arXiv
  • Awesome World Law Agent — A comprehensive survey of the World Law Agent ecosystem (AI + Law) — GitHub
  • HK-O1aw — A legal reasoning assistant for Hong Kong’s common law system, built on LLaMA-3.1-8B with O1-style reasoning — GitHub

Team Members: Chuxue CAO, Han ZHU, Yujin ZHOU, Ruoxi LI, Yuyao ZHANG


Pillar 2: AI Safety & RLHF

Ensuring that AI systems behave safely and align with human values is a fundamental challenge. This pillar develops methods for safe reinforcement learning from human feedback (RLHF), multi-modal safety alignment, and out-of-distribution detection — extending safety guarantees from text-only to multi-modal and multi-turn settings.

Key Contributions:

NeurIPS 2025

Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback

Ji, Jiaming; Chen, Xinyu; ...; Han, Sirui; Guo, Yike; Yang, Yaodong.

Extending safe RLHF to the multimodal domain, enabling alignment of vision-language models with human safety preferences.

ACL 2025

PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs

Ji, Jiaming; Hong, Donghai; ...; Han, Sirui; Guo, Yike; Yang, Yaodong.

A multi-level safety alignment framework and large-scale dataset for training safer LLMs.

Other Related Work:

  • SafeMT: Multi-turn Safety for Multimodal Language ModelsarXiv 2025arXiv
  • AM3Safety: Towards Data Efficient Alignment of Multi-modal Multi-turn Safety for MLLMsarXiv 2026arXiv
  • InterMT: Multi-Turn Interleaved Preference Alignment with Human FeedbackNeurIPS 2025arXiv
  • Context Reasoner: Incentivizing Reasoning Capability for Contextualized Privacy and Safety ComplianceEMNLP 2025arXiv

Team Members: Han ZHU, Chi-Min CHAN, Pengcheng WEN, Chuxue CAO


Pillar 3: Efficient & Robust LLMs

Deploying large language models in real-world applications requires overcoming significant computational and efficiency challenges. This pillar develops novel methods for model compression, parameter-efficient fine-tuning, and inference optimization — making powerful LLMs more accessible and practical.

Key Contributions:

AAAI 2026

Sub-MoE: Efficient Mixture-of-Expert LLMs Compression via Subspace Expert Merging

Li, Lujun; Zhu, Qiyuan; ...; Han, Sirui*; Guo, Yike*.

A novel approach to compressing MoE-based LLMs through subspace expert merging.

ICCV 2025

Efficient Fine-Tuning of Large Models via Nested Low-Rank Adaptation (NoRA)

Li, Lujun; Lin, Cheng; ...; Han, Sirui*; Guo, Yike*.

A nested low-rank adaptation method that achieves superior parameter efficiency for fine-tuning large models.

Other Related Work:

  • AIRA: Activation-Informed Low-Rank AdaptationICCV 2025
  • Outlier Matters: Efficient Long-to-Short Reasoning via Outlier-Guided Model MergingAAAI 2026
  • Outlier-Aware Model Merging for Efficient Multitask InferenceACM MM 2025
  • Semantic-guided Diverse Decoding for Large Language ModelNeurIPS 2025arXiv
  • DIDS: Domain Impact-aware Data SamplingEMNLP 2025arXiv

Team Members: Lujun LI, Qiyuan ZHU, Hao GU, Zhenyuan ZHANG


Pillar 4: Multimodal Intelligence & Embodied AI

This pillar pushes the boundaries of AI systems that perceive, understand, and interact with the physical world through multiple modalities. Our work spans 3D scene understanding, robotic manipulation, motion generation, and vision-language model evaluation.

Key Contributions:

AAAI 2026

ManipDreamer3D: Synthesizing Plausible Robotic Manipulation Video with Occupancy-aware 3D Trajectory

Li, Ying; Wei, Xiaobao; Chi, Xiaowei; ...; Han, Sirui; Zhang, Shanghang.

Generating realistic robotic manipulation videos using occupancy-aware 3D trajectory synthesis.

ICLR 2026

EgoTwin: Dreaming Body and View in First Person

Xiu, Jingqiao; Hong, Fangzhou; ...; Han, Sirui*; Pan, Liang*; Liu, Ziwei.

A novel framework for first-person body and view generation through egocentric dreaming.

Other Related Work:

  • Motion-R1: Enhancing Motion Generation with Decomposed Chain-of-Thought and RL BindingICLR 2026arXiv
  • EffiVMT: Video Motion Transfer via Efficient Spatial-Temporal Decoupled FinetuningICLR 2026arXiv
  • GSRender: Deduplicated Occupancy Prediction via Weakly Supervised 3D Gaussian SplattingICRA 2026arXiv
  • IR3D-Bench: Evaluating Vision-Language Model Scene Understanding as Agentic Inverse RenderingNeurIPS 2025arXiv
  • DanceEditor: Iterative Editable Music-driven Dance GenerationICCV 2025

Team Members: Xiaowei CHI, Xingqun QI, Zhenyuan ZHANG, Yakun CUI


Pillar 5: Computational Law & Social Science

Bridging computational methods with legal and social science inquiry, this pillar examines the societal implications of AI systems — from deepfake governance and AI manipulation to corpus inequality and consumer protection in AI-governed markets.

Key Contributions:

CHI 2026

Reimagining Legal Fact Verification with GenAI: Toward Effective Human-AI Collaboration

Han, Sirui; Zhang, Yuyao; Huang, Yidan; Li, Xueyan; Liu, Chengzhong; Guo, Yike*.

Pioneering the use of generative AI for legal fact verification through effective human-AI collaboration.

Other Related Work:

  • When AI Manipulates: Context-Dependent Effect and the Psychology of University Student Trust on Social MediaICA 2026
  • Deepfakes as Systemic Risk at the Media–Law BoundaryICA 2026
  • When Data Speaks for the North: Generative AI, Corpus Inequality, and the Reinforcement of Symbolic PowerICA 2026
  • Consumer Protection in AI–Governed Credit Markets: Algorithmic Bias and the Right to ExplanationUI JLTP 2026
  • Simulated Justice: How AI Alignment Replaces Conflict with CoherenceIHRLR 2026

Journal Articles on Law & Finance:

  • ESG Disclosure, Investor Awareness, and Carbon Risk PricingInternational Review of Law and Economics, 2024
  • Harmonizing Arbitration Rules for Non-Signatories in International TradeICCLR, 2024
  • Regulating Collective Labour Disputes in ChinaJournal of Comparative Law, 2016

Team Members: Yidan HUANG, Siyu PENG, Guoying LU, Chengyi JU, Yuyao ZHANG