Visual Language Large Model Hallucination Management from the Perspective of Service Design
Year: 2025
Type: conference-paper
Abstract: Purpose To tackle object hallucination in Large Vision-Language Models (LVLMs), this study adopts a humancentered governance approach rather than purely technical fixes. Method Drawing on cross-disciplinary insights from service design and AI governance, we introduce a “design intervention” model featuring a collaboratively built Design Semantic Constraint Layer (DSCL). This layer bridges human expertise and model behavior by encoding expert judgments into machinereadable constraints, dynamically integrated into LVLM workflows. Conclusion Comparative experiments confirm that our approach reduces object hallucination and improves output accuracy and coherence. The study establishes a human-machine service paradigm that balances AI capability with human creativity, positioning service design as a value converter in AI collaboration and offering a practical framework for reliable generative AI.
Authors: Nian Liu, Pingting Zhang
Institution: Hubei Institute of Fine Arts
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204795663