Pre-computed stage-2 latent perturbation direction dictionaries for REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations (ICML 2026).
⚠️ Warning: This method could be misused for malicious purposes. Provided for research and educational purposes only.
REALISTA attacks a target LLM by optimizing a sparse combination of latent editing directions, each corresponding to a… See the full description on the dataset page:
https://huggingface.co/datasets/byliang/REALISTA_latent_dictionary.
