Hallucination Eval

hallucination-eval

Evaluating the Quality of Hallucination Benchmarks for Large Vision-Language Models — Yan et al. (2024) (arXiv:2406.17115, 2024)

What this evaluates

Probes a model's ability to avoid generating factually incorrect statements about visual content. It measures alignment between model outputs and ground-truth visual facts using binary detection and scoring metrics.

Datasets

• POPE — total ?; splits: test (-1)
• AMBER-d — total ?; splits: test (-1)
• HallusionBench — total ?; splits: test (-1)

Metrics

• Accuracy (Acc) (primary) — range: [0, 1]
  • Proportion of correct binary predictions matching the ground truth label. Calculated as correct predictions divided by total instances.
• CHAIR — range: [0, 1]
  • Consistency-aware Hallucination Assessment Index; lower is better. Measures instance-level and sentence-level hallucination rates.
• OCH — range: [0, 1]
  • Open-Ended Consistency Hallucination metric; lower is better. Evaluates consistency in open-ended responses.
• MMHal Score — range: [0, 1]
  • Numerical score indicating hallucination quality; higher is better.
• GAVIE A-Score / R-Score — range: [0, 1]
  • Accuracy and Relevancy hallucination scores from GAVIE; higher is better.

Input / output format

Input: An image paired with a question or statement requiring a visual fact check.

Output: Binary response ('yes'/'no') or a numerical score, depending on the specific benchmark's protocol.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)

def compute_yes_rate(predictions):
    return sum(1 for p in predictions if p.lower() == 'yes') / len(predictions)

Common pitfalls

• Response bias: models may over-respond 'yes', inflating accuracy on imbalanced datasets.
• Parallel test instability: scores can vary significantly when the same model is evaluated on parallel test sets.
• Metric misalignment: automated scores (e.g., CHAIR, OCH) may not correlate well with human judgment of hallucination severity.

Evidence (verbatim from paper)

Acc denotes the accuracy, A-Score and R-Score denotes the accuracy and relevancy hallucination score in GAVIE. The top-2 results are bolded and underlined, respectively.

Citation

@misc{yan2024evaluating,
  title={Evaluating the Quality of Hallucination Benchmarks for Large Vision-Language Models},
  author={Yan et al. (2024)},
  year={2024},
  note={arXiv:2406.17115}
}

• arXiv: 2406.17115