Hong
Zhejiang
George
Griffith
Wake
AlibabaInternal representation
Understand how models encode knowledge, retrieved evidence, source provenance, conflict, and hallucination risk through white-box monitoring and hidden-state analysis.
Bio. I am a Research Assistant at the Hong Kong University of Science and Technology (Guangzhou), advised by Prof. Chengwei Qin. I am also a Research Intern at Zhejiang University (IFRC Lab), working on trustworthy language models with Prof. Meng Han and Dr. Wenpeng Xing. Additionally, I am honored to receive guidance on AI safety research from Prof. Simin Chen (George Mason University), Prof. Yi Liu (Griffith University), and Prof. Ying Zhang (Wake Forest University).
Beyond research, I gained valuable industry experience through an AI agent development internship (July–October 2025). I am grateful to my mentor, Wenjie Zhang, a P9 engineer at Alibaba Group, for his guidance and support.
Research. From latent knowledge to reliable action, memory, and verifiable trust. I study how models internally represent knowledge, evidence provenance, conflict, and risk; why those signals fail to guide generation, tool use, and memory; and how monitoring, system safeguards, and verification can make AI systems more reliable.
Research Agenda
Trustworthy agents must not only represent provenance, conflict, and risk internally. These signals should reliably inform what agents say, do, remember, reuse, and learn, while the resulting models and artifacts remain externally verifiable.
Action and memory
Study why decodable safety signals fail to guide generation, tool use, and memory updates, and turn monitoring into reliable system safeguards.
Verifiable trust
Build accountable deployment mechanisms through information-flow control, model fingerprinting, blockchain, and privacy-preserving verification.
Current themes
- Multi-agent memory consistency: write validity, contamination, persistence, propagation, and repair.
- Poisoning defense and credit: separating the source, amplifier, executor, and best remediation target.
- Runtime monitoring of agent skills: identifying risky skills before unsafe tool actions.
My longer-term goal is to understand environment–verifier–memory co-evolution in lifelong agents: how tasks, evaluation, memory, and oversight should adapt together as agents act, learn, and update over time.
Academic Services
Reviewer
- EMNLP 2026
- NeurIPS 2026
- AAAI 2027
- ICLR 2027
Publications
* Equal contribution.
DISF: Detecting Hallucinations in Retrieval-Augmented Generation via Dual-path Internal State Forcing Framework
Accepted at ACL 2026
Detects RAG hallucinations through dual-path internal-state forcing and white-box activation signals.
FIDES: Faithful Inference via Deep Evidence Signals for Retrieval-Memory Conflict in RAG
Accepted at EMNLP 2026
Uses three complementary internal signals for token-level, training-free intervention under retrieval-memory conflict.
Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning
Accepted at EMNLP 2026
Introduces a double-gate protocol that separates atomic knowledge stability from compositional reasoning.
PLAIT: From Hallucination Scores to Parent-Preserving Audit Plans
Under review at AAAI 2027
PLAIT converts hallucination scores into parent-preserving response–claim–span audit plans, using whole-plan learning and exact budgeted optimization to capture more unsupported content per review minute.
The Attribution Blind Spot: Detecting When Language Models Rely on Memory Rather Than Retrieved Context
Under review at AAAI 2027
Separates identical surface outputs that follow different internal parametric-memory and retrieved-evidence routes.
Calibration Does Not Certify a Contrast: Identification and Correction for LLM Evaluators
Under review at AAAI 2027
GAVEL shows item calibration cannot validate experimental contrasts: paired audits reveal that Qwen judges greatly overstate a reminder intervention.
Knowing Is Not Acting: Representation–Action Dissociation in Indirect Prompt Injection
Under review at NeurIPS 2026
Shows that source role is decodable early while tool decisions become controllable only in a late commitment band.
Whose Thoughts? Chain-of-Thought Override in Reasoning-Tuned Language Models
Under review at NeurIPS 2026
Exposes a source-override vulnerability in which assistant-side reasoning for another question can redirect the final answer.
Cordon-MAS: Defending RAG against Knowledge Poisoning via Information-Flow Control
To be submitted to ICLR 2027
A compartmentalized multi-agent defense that restricts how untrusted evidence reaches final synthesis.
Fingerprint Vector: Enabling Scalable and Efficient Model Fingerprint Transfer via Vector Addition
Under review at ARR (October 2026)
Enables scalable model fingerprint transfer through vector addition for ownership verification.
LatentAudit: Real-Time White-Box Faithfulness Monitoring for Retrieval-Augmented Generation with Verifiable Deployment
Accepted at NeurIPS 2026 Workshop
Monitors faithfulness using distances between residual-stream activations and evidence representations.
From Retinal Evidence to Safe Decisions: RETINA-SAFE and ECRT for Hallucination Risk Triage in Medical LLMs
Accepted at NeurIPS 2026 Workshop
A 12,522-sample evidence-grounded benchmark and a two-stage white-box hallucination-risk triage framework.
ZK-FPE: Blockchain-Verifiable Model Fingerprinting with Zero-Knowledge Privacy for Ownership Attribution
Accepted at ACM TURC 2026
Combines zero-knowledge proofs and blockchain for privacy-preserving model ownership attribution.
IoTStorage
Trusted Metadata-Coordinated Tiered Off-Chain Storage for Recovery-Safe and Low-Latency IoT Data Management
Electronics (MDPI)
Trusted metadata coordination and tiered off-chain storage for recovery-safe, low-latency IoT data management.
OPTICSSensing
Orthogonal salinity and temperature detection via paralleled dual all-fiber interferometers
Optics Communications, 2025
Parallel dual all-fiber interferometers for orthogonal salinity and temperature detection.
BIBLIOICETM
Bibliometric analysis of physical education research in China from 2014 to 2024
ACM ICETM, 2025
A bibliometric analysis of research trends, keyword co-occurrence, and institutional collaboration.
Experience & Education
Hong Kong University of Science and Technology (Guangzhou)
Guangzhou, China
Research Assistant, advised by Prof. Chengwei Qin.
Binjiang Institute of Zhejiang University · IFRC Lab
Hangzhou, China
Research Intern, supervised by Prof. Meng Han and Dr. Wenpeng Xing.
University of Malaya
Kuala Lumpur, Malaysia
Visiting Student.
Westlake University
Hangzhou, China
Research Assistant, supervised by Dr. Ziyang Zhang.
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. 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. Application No. 2026107898102, filed June 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. Application No. 2026107899270, filed June 3, 2026. Pending.