Halluaudio Eval

halluaudio-eval

HalluAudio: A Comprehensive Benchmark for Hallucination Detection in Large Audio-Language Models — Zhao et al. (2026) (arXiv:2604.19300, 2026)

What this evaluates

This benchmark probes the hallucination detection capabilities of Large Audio-Language Models (LALMs) across speech, environmental sound, and music domains. It systematically induces hallucinations using adversarial prompts and mixed-audio inputs to evaluate response correctness, affirmative bias, and refusal behavior beyond standard accuracy.

Datasets

• HalluAudio — total ?; splits: test (-1); repo https://github.com/Feiyuzhao25/halluaudio

Metrics

• Accuracy (primary) — range: percent
  • Measures the fraction of prompts with well-defined ground truth answered correctly: $\text{Accuracy}{d}=\frac{1}{|P{d}|}\sum_{p\in P_{d}}\mathbf{1}{\hat{y}{p}=y{p}}$. Captures explicit hallucination cases where predictions contradict ground-truth audio evidence.
• Yes-p Ratio — range: [0, 1]
  • Diagnoses affirmative bias in binary responses: $\text{Yes-p Ratio}=\frac{\sum_{p\in P_{\text{binary}}}\mathbf{1}{\hat{y}{p}=\text{Yes}}}{|P{\text{binary}}|}$. Higher values indicate a tendency to produce unsupported positive responses.
• Unrelated Ratio — range: [0, 1]
  • Measures the proportion of incorrect binary predictions that are affirmative: $\text{Unrelated Ratio}=\frac{\sum_{p\in P_{\text{binary}}}\mathbf{1}{\hat{y}{p}\neq y{p}\wedge\hat{y}{p}}}{\sum{p\in P_{\text{binary}}}\mathbf{1}{\hat{y}{p}\neq y{p}}}$. Highlights cases where affirmative bias leads to semantically unrelated errors.
• Conditional Accuracy — range: [0, 1]
  • Evaluates accuracy specifically on binary prompts: $\text{Conditional Accuracy}=\frac{\sum_{p\in P_{\text{binary}}}\mathbf{1}{\hat{y}{p}=y{p}}}{|P_{\text{binary}}|}$. Reveals asymmetric decision behavior where models may perform well on positives but poorly on negatives.
• False Refusal Rate (FRR) — range: [0, 1]
  • Captures cases where models abstain despite valid prompts: $\text{FRR}{d}=\frac{|{p\in P{d}:\hat{y}{p}\in\mathcal{R}}|}{|P{d}|}$. Reflects over-conservative hallucination behavior and breakdowns in evidence-based decision making.

Input / output format

Input: Audio input (speech, environmental sound, or music) paired with a text prompt. Prompts cover binary classification, multi-choice reasoning, attribute verification, and open-ended QA, often incorporating adversarial perturbations and mixed-audio contexts to induce hallucinations.

Output: Model's predicted response, typically a binary Yes/No, a classification label, or a free-text answer.

Scoring recipe

def compute_metrics(predictions, golds, domain='all'):
    P = [(p, g) for p, g in zip(predictions, golds) if g.domain == domain]
    acc = sum(1 for p, g in P if p == g) / len(P)
    binary_P = [(p, g) for p, g in P if g.task_type == 'binary']
    yes_ratio = sum(1 for p, _ in binary_P if p == 'Yes') / len(binary_P)
    wrong = [p for p, g in binary_P if p != g]
    unrelated_ratio = sum(1 for p in wrong if p == 'Yes') / len(wrong) if wrong else 0
    cond_acc = sum(1 for p, g in binary_P if p == g) / len(binary_P)
    refusal_set = {'refuse', 'cannot answer', 'i dont know'}
    frr = sum(1 for p, _ in P if p.lower() in refusal_set) / len(P)
    return {'accuracy': acc, 'yes_ratio': yes_ratio, 'unrelated_ratio': unrelated_ratio, 'conditional_accuracy': cond_acc, 'frr': frr}

Common pitfalls

• Treating a high Yes-pred ratio as good performance, whereas it often indicates unsupported affirmative bias rather than evidence-grounded reasoning.
• Ignoring False Refusal Rate (FRR) as a distinct hallucination mode; models may abstain despite clear acoustic evidence, reflecting over-conservative decision policies rather than safety alignment.
• Aggregating accuracy across speech, sound, and music domains without accounting for task-specific structural vs. semantic hallucination patterns, which vary significantly across domains.

Evidence (verbatim from paper)

Accuracy measures the fraction of prompts with well-defined ground truth that are answered correctly by LALMs: $\text{Accuracy}{d}=\frac{1}{|P{d}|}\sum_{p\in P_{d}}\mathbf{1}{\hat{y}{p}=y{p}}$

Citation

@misc{zhao2026halluaudio,
  title={HalluAudio: A Comprehensive Benchmark for Hallucination Detection in Large Audio-Language Models},
  author={Zhao et al. (2026)},
  year={2026},
  note={arXiv:2604.19300}
}

• arXiv: 2604.19300