Omnidpo Hallucination Eval

omnidpo-hallucination-eval

OmniDPO: A Preference Optimization Framework to Address Omni-Modal Hallucination — Junzhe Chen et al. (2025) (arXiv:2509.00723, 2025)

What this evaluates

This evaluation protocol assesses the ability of omni-modal large language models to avoid hallucinating non-existent audio or visual content when presented with contradictory or incomplete multimodal inputs. It specifically probes whether models can correctly identify real elements while resisting false affirmations of missing ones across text, vision, and audio modalities.

Datasets

• AVHBench — total ?; splits: test (-1)
• CMM — total ?; splits: test (-1)

Metrics

• Accuracy — range: percent
  • Standard classification accuracy: proportion of correct predictions out of total predictions.
• Precision — range: percent
  • Standard classification precision: proportion of true positives among all positive predictions.
• Recall — range: percent
  • Standard classification recall: proportion of true positives among all actual positives.
• F1 Score (primary) — range: percent
  • Harmonic mean of Precision and Recall: 2 * (Precision * Recall) / (Precision + Recall).
• Perception Accuracy (PA) — range: percent
  • PA = # correctly predicted 'yes' / # ground-truth 'yes'. Measures the model's ability to correctly detect real elements.
• Hallucination Resistance (HR) (primary) — range: percent
  • HR = # correctly predicted 'no' / # ground-truth 'no'. Quantifies the model's robustness in rejecting non-existent elements.

Input / output format

Input: Multimodal inputs (audio, video, and/or text) where one or more modalities may contradict each other or be degraded. Each sample is paired with probing questions targeting either an existing object/event (ground-truth 'yes') or a non-existent one (ground-truth 'no').

Output: Binary classification response: 'yes' or 'no'.

Scoring recipe

def compute_pa_hr(predictions, gold):
    actual_yes = sum(1 for g in gold if g == 'yes')
    actual_no = sum(1 for g in gold if g == 'no')
    correct_yes = sum(1 for p, g in zip(predictions, gold) if p == 'yes' and g == 'yes')
    correct_no = sum(1 for p, g in zip(predictions, gold) if p == 'no' and g == 'no')
    pa = correct_yes / actual_yes if actual_yes > 0 else 0.0
    hr = correct_no / actual_no if actual_no > 0 else 0.0
    return pa, hr

Common pitfalls

• Confusing PA (detecting real elements) with HR (rejecting non-existent ones), as they measure opposite error types.
• Assuming text-only preference optimization (DPO) generalizes to audio-visual hallucination mitigation without modality-aware conditioning.
• Ignoring the strong 'yes' bias in base models, which inflates accuracy but masks hallucination tendencies.

Evidence (verbatim from paper)

CMM (Curse of Multi-Modalities) *[[34]]*is a benchmark specifically designed to evaluate hallucination behavior in OLLMs across text, vision, and audio modalities. It focuses on measuring a model’s tendency to generate modality-inconsistent outputs by presenting it with inputs where one or more modalities contradict the others. Each sample is paired with two probing questions—one targeting an existing object or event (ground-truth answer "yes") and one targeting a non-existent one (ground-truth "no")—covering both visual and auditory modalities. CMM uses two core metrics: Perception Accuracy (PA) and Hallucination Resistance (HR), defined as follows: | PA = # correctly predicted "yes" / # ground-truth "yes", HR = # correctly predicted "no" / # ground-truth "no" | where PA measures the model’s ability to correctly detect real elements, while HR quantifies its robustness in rejecting non-existent ones.

Citation

@misc{chen2025omnidpo,
  title={OmniDPO: A Preference Optimization Framework to Address Omni-Modal Hallucination},
  author={Junzhe Chen et al. (2025)},
  year={2025},
  note={arXiv:2509.00723}
}

• arXiv: 2509.00723