What type of study is this?

This is a Experimental Study study.

October 20, 2025Open Access

GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs

View Full Paper

Ask AI

Bookmark

GHOST achieves a hallucination success rate of over 28%, significantly higher than prior methods at 1%.
The method utilizes a diffusion model with optimized embeddings to create visually natural but misleading images.
GHOST uncovers transferable vulnerabilities, with an induction rate of 66.5% for images optimized for Qwen2.5-VL in GPT-4o.
Fine-tuning on GHOST-generated images mitigates hallucination, indicating its potential as a corrective tool.

Discussion

No takes yet. Share an insight, caveat, or question.

Implication

GHOST generates misleading images to uncover hallucination vulnerabilities in multimodal large language models, suggesting a new way to improve model reliability.

Cite This Study

Parast et al. (2025) studied this question.

synapsesocial.com/papers/68f5fcd68d54a28a75cf1e50 https://doi.org/10.48550/arxiv.2509.25178

View Full Paper

Ask AI

Bookmark

Implication

GHOST generates misleading images to uncover hallucination vulnerabilities in multimodal large language models, suggesting a new way to improve model reliability.

GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs

Key Points

Discussion

Implication

Cite This Study

GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs

Key Points

Discussion

Implication

Cite This Study

Also Consider

Also Consider