A direction-aware diagnostic benchmark and protocol for vision-language model hallucination, accompanying the NeurIPS 2026 Datasets & Benchmarks Track submission.
TL;DR
Existing VLM hallucination benchmarks treat failure as a single scalar. HALLUCINOGEN-Signed treats it as a signed phenomenon (yes-bias vs no-bias) and provides:
4 prompt formats × 3 difficulty splits = 3,612 instances
A 13-word direction-aware adjective partition
A reproducible… See the full description on the dataset page: https://huggingface.co/datasets/jin-kwon/halluc-signed.
