Copo Hallucination Eval

copo-hallucination-eval

COPO: Causal-Oriented Policy Optimization for Hallucinations of MLLMs — Peizheng Guo et al. (2025) (arXiv:2508.04182, 2025)

What this evaluates

Evaluates the ability of Multimodal Large Language Models to generate factually grounded captions and answers while suppressing object-level hallucinations. It probes visual grounding, reasoning consistency, and alignment with human or GPT-4 preferences across multiple reasoning and perception benchmarks.

Datasets

• CHAIR — total ?; splits: test (-1)
• POPE — total ?; splits: test (-1)
• MMBench — total ?; splits: test (-1)
• MME — total ?; splits: test (-1)

Metrics

• CHAIR_I — range: percent
  • Instance-level hallucination error rate: percentage of generated captions containing objects not present in the ground truth image.
• CHAIR_S — range: percent
  • Sentence-level hallucination error rate: percentage of sentences in generated captions containing hallucinated objects.
• POPE F1 Score (primary) — range: [0, 1]
  • Harmonic mean of precision and recall on object-level probing questions in a closed-set VQA format.
• MMBench — range: percent
  • Aggregate accuracy score across perception, grounding, and reasoning tasks.
• MME — range: percent
  • Aggregate score measuring perception, grounding, and reasoning capabilities.
• GPT-4o Accuracy (A), Correctness (C), Detailedness (D) — range: [1, 5]
  • Automatic ratings provided by GPT-4o on a scale for accuracy, correctness, and detailedness of generated captions.

Input / output format

Input: Image paired with a text prompt (e.g., captioning instruction or VQA question).

Output: Generated text response (caption or answer).

Scoring recipe

def compute_metrics(predictions, golds, metric_name):
    if metric_name == 'CHAIR_I':
        return 100 * sum(1 for p, g in zip(predictions, golds) if has_hallucinated_objects(p, g)) / len(predictions)
    elif metric_name == 'CHAIR_S':
        return 100 * sum(1 for p, g in zip(predictions, golds) if any(has_hallucinated_objects(s, g) for s in p.split('.'))) / len(predictions)
    elif metric_name == 'POPE F1 Score':
        preds_labels = [1 if pred.lower() == 'yes' else 0 for pred in predictions]
        return f1_score(golds, preds_labels)
    elif metric_name == 'GPT-4o':
        return gpt4o_rate(predictions, rubric=['accuracy', 'correctness', 'detailedness'])
    return None

Common pitfalls

• CHAIR error rates are lower-is-better, which is counterintuitive compared to standard accuracy metrics.
• POPE uses a closed-set probing VQA format rather than open-ended generation, so metrics reflect detection capability rather than generation quality.
• GPT-4o evaluation scores are subjective and depend heavily on the specific prompt template and temperature settings used during rating.

Evidence (verbatim from paper)

CHAIR measures object hallucinations in captioning by computing instance-level (CHAIR_I) and sentence-level (CHAIR_S) error rates, while POPE adopts a probing-based VQA setting to evaluate object-level hallucinations under controlled conditions. As shown in Table [1] and Table [2], our approach yields lower hallucination rates in CHAIR and higher F1 scores in POPE, demonstrating the effectiveness of our approach. For GPT-4 assisted evaluation, we compare model-generated captions on images, with GPT-4o rating them on accuracy (A), correctness (C), and detailedness (D).

Citation

@misc{guo2025copo,
  title={COPO: Causal-Oriented Policy Optimization for Hallucinations of MLLMs},
  author={Peizheng Guo et al. (2025)},
  year={2025},
  note={arXiv:2508.04182}
}

• arXiv: 2508.04182