GHOST

Successful hallucination-inducing images from GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs (ICLR 2026).

These are the generated images for the main COCO object-hallucination experiment. Each file is a sample where the victim model answered “Yes” for a target object that OWLv2 did not detect. The original COCO images are not included.

Layout

qwen/<class>/*.png
llava/<class>/*.png

Classes: boat, bottle, bus, carrot, clock, knife, suitcase, toilet, traffic_light, vase.

Model	Images
Qwen2.5-VL-7B-Instruct	2816
LLaVA-v1.6-Mistral-7B	2462

Filenames are {index}_{coco_id}_{step}_{generation}.png.

Attack settings

Both sets use the main-paper settings: learning rate 0.1, 100 steps, threshold 0.8, 4 generations, guidance scale 5, λ_clip 15, λ_reg 10, OWLv2 threshold 0.5, CLIP-sorted negatives, diffusion noise level 30, 50 inference steps.

• Qwen projector: batch 32, learning rate 2e-4, 10 epochs, context 4096, hidden 2048
• LLaVA projector: batch 64, learning rate 2e-4, 10 epochs, context 4096, hidden 1024

Citation

@article{parast2025ghost,
  title={GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs},
  author={Parast, Aryan Yazdan and Hosseini, Parsa and Asadollahzadeh, Hesam and Moakhar, Arshia Soltani and Azam, Basim and Feizi, Soheil and Akhtar, Naveed},
  journal={arXiv preprint arXiv:2509.25178},
  year={2025}
}