Hi, I am Zhe Yu (俞哲) from Hangzhou, China. I am a Research Intern at the Binjiang Institute of Zhejiang University (IFRC Lab), where I work on trustworthy language models with Dr. Meng Han and Dr. Wenpeng Xing. I am currently pursuing my B.Eng. in Artificial Intelligence at the Communication University of Zhejiang (CUZ), supervised by Dr. Hao Zeng. During my undergraduate studies, I also spent time as a Visiting Student at Westlake University (supervised by Dr. Ziyang Zhang) and the University of Malaya.

📢 Seeking Opportunities: I am actively looking for Research Assistant (RA) positions (on-site or remote) and seeking Fall 2027 PhD opportunities. Please feel free to drop me an email if you are interested!

📄 **Download my full CV (English) 获取完整中文简历**

My research centers on trustworthy language models across three interconnected threads. First, I study the internal mechanisms of knowledge grounding — how parametric memory and retrieved evidence interact during generation, and how to detect failures like hallucinations, memory hijacking, and compositional reasoning collapse through white-box monitoring and mechanistic analysis. Second, I work on representation–action dissociation in reasoning and agentic systems — probing when and why models encode conflict information internally yet fail to route it into downstream decisions. Third, I explore verifiable model ownership and decentralized trust — combining fingerprinting, blockchain, and zero-knowledge proofs to build scalable, privacy-preserving attribution and deployment frameworks.

📚 Research Papers

DISF
Zhe Yu*, Wenpeng Xing*, Wenjie Luo, Weize Xu, Lingtong Huang, Yourong Chen, Changting Lin, Meng Han
A dual-path internal-state forcing framework that detects hallucinations in RAG by leveraging white-box activation signals.
Accepted at ACL 2026 [PDF]
KINA
Zhe Yu, Wenpeng Xing, Zhenhua Xu, Xingxing Yang, Meng Han
Shows that indirect prompt injection failure is not absent source recognition but representation–action dissociation: source role is linearly decodable early, yet tool decisions become causally controllable only in a late commitment band.
Under review at NeurIPS 2026 [PDF]
WT
Zhe Yu, Wenpeng Xing, Zhenhua Xu, Ruiqi Zhang, Meng Han
Exposes a structural source-override vulnerability in reasoning-tuned models: when the assistant-side <think> block contains a CoT for a different question, models answer the wrong question in the majority of cases.
Under review at NeurIPS 2026 [PDF]
ABS
Zhe Yu*, Wenpeng Xing*, Bo Yang, Chen Ye, Gaolei Li, Yunzhao Wei, Meng Han
Formalizes the attribution blind spot — when parametric memory and retrieved context produce identical surface text — and proposes Computational Reality Monitoring (CRM) to detect internal trajectory divergence.
Under review at ARR / EMNLP [PDF] [arXiv]
FIDES
Zhe Yu*, Wenpeng Xing*, Tiancheng Zhao, Mohan Li, Changting Lin, Meng Han
Reveals token-level conflict concentration in retrieval-memory conflict and proposes a training-free decoder that fuses three complementary internal signals for per-token selective intervention.
Under review at ARR / EMNLP [PDF]
DINR
Zhe Yu*, Wenpeng Xing*, Chen Ye, Xuyang Teng, Bo Yang, Changting Lin, Meng Han
Demonstrates a structural monitoring–control gap in RAG: models detect contradictory evidence but this awareness fails to constrain final recommendations, and single-turn diagnostics systematically overestimate multi-turn safety.
Under review at ARR / EMNLP [PDF] [arXiv]
CORDON
Zhe Yu*, Wenpeng Xing*, Gaolei Li, Shuguang Xiong, Hongzhi Wang, Xuyang Teng, Meng Han
A multi-agent compartmentalized defense that enforces the Cordon Principle architecturally, reducing knowledge-poisoning attack success rate by 92.4%.
Under review at ARR / EMNLP [PDF] [arXiv]
CC
Zhe Yu*, Wenpeng Xing*, Yunzhao Wei, Hongzhi Wang, Xuyang Teng, Meng Han
Introduces a double-gate protocol that separates atomic knowledge stability from compositional reasoning, revealing post-training recipes can diverge by >40 pp in composition failure at matched atoms.
Under review at ARR / EMNLP [PDF] [arXiv]
FV
Fingerprint Vector: Enabling Scalable and Efficient Model Fingerprint Transfer via Vector Addition
Zhenhua Xu, Qichen Liu, Zhebo Wang, Zhe Yu, Xixiang Zhao, Wenpeng Xing, Dezhang Kong, Mohan Li, Meng Han
Enables scalable and efficient model fingerprint transfer via vector addition for ownership verification.
Under review at ARR / EMNLP
LA
Zhe Yu*, Wenpeng Xing*, Meng Han
A real-time white-box auditor that measures Mahalanobis distance between residual-stream activations and evidence representations to judge RAG faithfulness at generation time.
Under review at CoLM 2026 [arXiv]
RETINA
Zhe Yu*, Wenpeng Xing*, Meng Han
A 12,522-sample evidence-grounded benchmark for diabetic retinopathy decision settings and a two-stage white-box detection framework (ECRT) for safe/unsafe risk triage with explicit subtype attribution.
Under review at MICCAI 2026 [arXiv]
ZK-FPE
ZK-FPE: Blockchain-Verifiable Model Fingerprinting with Zero-Knowledge Privacy for Ownership Attribution
Zhiguo Ma*, Zhe Yu, Wenpeng Xing, Yourong Chen, Meng Han
Combines zero-knowledge proofs and blockchain to build verifiable model ownership attribution while preserving privacy.
Accepted at ACM TURC 2026 [CFP]
IoT
Trusted Metadata-Coordinated Tiered Off-Chain Storage for Recovery-Safe and Low-Latency IoT Data Management
Weiping Yu, Weihan Wang, Mingyuan Yan, Keyang He, Zhe Yu, Wenpeng Xing, Liyuan Liu, Meng Han
Trusted metadata-coordinated tiered off-chain storage for IoT data management with recovery safety and low latency.
Electronics (MDPI)
Optics
F. Zhou, C. Chang, Q. Chang, H. Zhang, Zhe Yu, W. Liu, J. Li, J. Yang
Orthogonal salinity and temperature detection using parallel dual all-fiber interferometers.
Optics Communications, 2025 [DOI]
Biblio
Zhe Yu, H. Zeng, Y. Zhao, X. Zhang, Z. Wang, Y. Tao, M. Yuan, X. Sun
A bibliometric analysis of physical education research trends, keyword co-occurrence, and institutional collaboration in China over a decade.
ACM ICETM, 2025 [DOI]

📖 Education

  • Nov 2025 - Present, Binjiang Institute of Zhejiang University (IFRC Lab), Hangzhou, China
    • Research Intern, supervised by Dr. Meng Han and Dr. Wenpeng Xing
    • Part of the Guangdong Provincial Key R&D Program “Multimodal LLM Safety System Research and Application”
    • Part of the National Key R&D Program (Young Scientist Project) “Novel Trust System Based on Blockchain”
  • Jan 2025 - Feb 2025, University of Malaya, Kuala Lumpur, Malaysia
    • Visiting Student
  • Mar 2024 - Sep 2024, Westlake University, Hangzhou, China
    • Visiting Student, Optical Laboratory, supervised by Dr. Ziyang Zhang. Worked on dual all-fiber interferometer systems for orthogonal salinity/temperature detection (published in Optics Communications). This early cross-disciplinary research grounded my experimental rigor and shaped my approach to extracting and interpreting internal signals — a methodology central to my current work on LLM mechanistic interpretability and white-box auditing.
  • 2023 - Expected 2027, Communication University of Zhejiang, Hangzhou, China
    • B.Eng. in Artificial Intelligence, supervised by Dr. Hao Zeng

📜 Patents

  • Zhe Yu, Wenpeng Xing, Meng Han. A hallucination detection method based on dual-path internal state forcing logic for retrieval-augmented generation in large language models. Pending Patent Application No. 202610260408X (Under Review).
  • Meng Han, Zhe Yu, Jiayan Hu, Rongchang Li, Wenpeng Xing, Jingyi Yu, Zhen Hong, et al. A post-processing method, system, device, and medium for hallucination detection in large language models based on adaptive order statistics aggregation. Chinese Patent Application No. 2026107898102, filed Jun 3, 2026. (Pending)
  • Zhe Yu, Jiayan Hu, Jingyi Yu, Weihang Yu, Wenpeng Xing, Jing Xiong, Yourong Chen, Zhen Hong, et al. A hallucination detection method, system, and device for large language models based on multi-dimensional heterogeneous feature fusion. Chinese Patent Application No. 2026107899270, filed Jun 3, 2026. (Pending)