Formalizes hallucination-to-action conversion: an unsupported perceptual claim becomes the basis for a privileged action. Evidence-carrying agents decompose tool calls into action-critical predicates, check them with constrained DOM, OCR, and accessibility-tree verifiers, and authorize actions through a deterministic gate.
BibTeX citation
@misc{zhang2026eca,
title = {Hallucination as Exploit: Evidence-Carrying Multimodal Agents},
author = {Guijia Zhang and Hao Zheng and Harry Yang},
year = {2026},
eprint = {2605.19192},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.19192}
}
Studies invocation-time risk estimation from the user request, candidate skill, and runtime context. The work combines static capability priors with contextual scoring and introduces SIA-Bench, a 3,000-record benchmark with source-group splits and a held-out indirect-prompt-injection split.
BibTeX citation
@misc{zhang2026stars,
title = {{STARS}: Skill-Triggered Audit for Request-Conditioned Invocation Safety in Agent Systems},
author = {Guijia Zhang and Shu Yang and Xilin Gong and Di Wang},
year = {2026},
eprint = {2604.10286},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2604.10286}
}
Document Image Tampering Localization Based on Foundation Models
Guijia Zhang — third author
International Conference on Document Analysis and Recognition (ICDAR), 2025
A Diffusion-Based Approach for Restoring Face-Swapped Images
Guijia Zhang — third author
Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2024
Preprints
Agent Skill Regulation as a Procedural Knowledge Representation Problem
Guijia Zhang — co-first author, with collaborators at HKUST
Preprint
About this work
Models agent skills as procedural knowledge with execution commitments, state constraints, and recovery actions. The work studies the relationship between formal prior libraries and procedural evidence in human-written skills across syntax, semantics, and execution.
Uses paired single-channel interventions to test whether GUI agents ground state beliefs in pixels or serialized structure. The Perception-Fusion Gap measures cases where correct visual perception is overridden by conflicting structural information. The updated preprint studies 735 probes from web, mobile, and desktop interfaces.
BibTeX citation
@misc{zhang2026gui,
title = {Do {GUI} Agents Believe Their Eyes? Diagnosing State-Belief Reliance on Pixels versus Structure},
author = {Guijia Zhang and Yuxun Chen and Yuheng Qi and Harry Yang},
year = {2026},
eprint = {2607.04334},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.04334}
}