Dive into the Agent Matrix: A Realistic Evaluation of Self-Replication Risk in LLM Agents

[Preprint. Under Review]

Boxuan Zhang*, Yi Yu*, Jiaxuan Guo, and Jing Shao

TL;DR: We present a comprehensive evaluation framework for quantifying self-replication risks. Our framework establishes authentic production environments and realistic tasks (e.g., dynamic load balancing) to enable scenario-driven assessment of agent behaviors.

[PDF]   [BLOG]   [BibTeX]

@article{zhang2025dive,
  title={Dive into the Agent Matrix: A Realistic Evaluation of Self-Replication Risk in LLM Agents},
  author={Zhang, Boxuan and Yu, Yi and Guo, Jiaxuan and Shao, Jing},
  journal={arXiv preprint arXiv:2509.25302},
  year={2025}
} 

What Shapes a Creative Machine Mind? Comprehensively Benchmarking Creativity in Foundation Models

[Preprint. Under Review]

Zicong He*, Boxuan Zhang*, Weihao Liu*, Ruixiang Tang, and Lu Cheng

TL;DR: We introduce \C^2\-Eval, a holistic benchmark for the unified assessment of creativity in FMs. \C^2\-Eval distinguishes between two complementary forms of Creativity (\C^2\): convergent creativity, where tasks admit constrained solutions (e.g., code generation), and divergent creativity, where tasks are open-ended (e.g., story telling).

[PDF]   [BibTeX]

@article{he2025what,
  title={What Shapes a Creative Machine Mind? Comprehensively Benchmarking Creativity in Foundation Models},
  author={He, Zicong and Zhang, Boxuan and Liu, Weihao and Tang, Ruixiang and Cheng, Lu},
  journal={arXiv preprint arXiv:2510.04009},
  year={2025}
} 

Frontier AI Risk Management Framework in Practice: A Risk Analysis Technical Report

[Technical Report]

Shanghai AI Lab and Concordia AI, …, Boxuan Zhang, [30+ authors]

TL;DR: We identify critical frontier AI risks in seven areas: cyber offense, biological and chemical risks, persuasion and manipulation, uncontrolled autonomous AI R&D, strategic deception and scheming, self-replication, and collusion.

[PDF]   [Blog]   [BibTeX]

@article{shailab2025frontier,
  title={Frontier ai risk management framework (v1.0)},
  author={Shanghai AI Lab & Concordia AI},
  url={https://research.ai45.shlab.org.cn/safework-f1-framework.pdf},
  year={2025}
} 

Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs’ Decoding Layers

[Preprint. Under Review]

Zicong He*, Boxuan Zhang*, and Lu Cheng

TL;DR: Given the philosophical nature of creativity, we propose a narrow definition tailored to LLMs and introduce an evaluation framework, HCL, which quantifies Hallucination and Creativity across different Layers of LLMs during decoding.

[PDF]   [Code]   [BibTeX]

@article{he2025shakespearean,
  title={Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs' Decoding Layers},
  author={He, Zicong and Zhang, Boxuan and Cheng, Lu},
  journal={arXiv preprint arXiv:2503.02851},
  year={2025}
} 

CoT-UQ: Improving Response-wise Uncertainty Quantification in LLMs with Chain-of-Thought

[ACL 2025 Findings]

Boxuan Zhang and Ruqi Zhang

TL;DR: Propose to quantify response-wise uncertainty by integrating LLMs’ inherent reasoning capabilities through Chain-of-Thought (CoT) into the UQ process.

[PDF]   [Code]   [BibTeX]

@article{zhang2025cot,
    title={CoT-UQ: Improving Response-wise Uncertainty Quantification in LLMs with Chain-of-Thought},
    author={Zhang, Boxuan and Zhang, Ruqi},
    journal={arXiv preprint arXiv:2502.17214},
    year={2025}
} 

What If the Input is Expanded in OOD Detection?

[Neural Information Processing Systems (NeurIPS), 2024]

Boxuan Zhang*, Jianing Zhu*, Zengmao Wang, Tongliang Liu, Bo Du and Bo Han

TL;DR: Propose a novel perspective to employ different common corruptions on the input space to expand the representation dimension for OOD detection.

[PDF]   [Project Page]   [Code]   [BibTeX]

@inproceedings{zhang2024what,
    title={What If the Input is Expanded in OOD Detection?},
    author={Zhang, Boxuan and Zhu, Jianing and Wang, Zengmao and Liu, Tongliang and Du, Bo and Han, Bo},
    booktitle={The Thirty-Eighth Annual Conference on Neural Information Processing Systems},
    year={2024},
} 

Boosting Semisupervised Object Detection in Remote-Sensing Images With Active Teaching

[IEEE Geoscience and Remote Sensing Letters (GRSL), 2024]

Boxuan Zhang, Zengmao Wang and Bo Du

TL;DR: Propose to boost semi-supervised object detection with active teaching (SSOD-AT) in remote sensing images, which helps to alleviate the dependency on limited labeled images in remote sensing scenarios.

[PDF]   [Code]   [BibTeX]

@article{zhang2024boosting,
    title={Boosting Semi-Supervised Object Detection in Remote Sensing Images with Active Teaching},
    author={Zhang, Boxuan and Wang, Zengmao and Du, Bo},
    journal={IEEE Geoscience and Remote Sensing Letters},
    year={2024},
    publisher={IEEE}
}