Does Location Matter? A Black-Box Causal Audit Reveals Model-Dependent Audio Grounding in LALM Hallucination

Key points

• Holding duration fixed, localization beats random removal at every cell on both models.1
• Raw detection value remains model-dependent, especially on Audio Flamingo 3.1
• The same evidence corrects decoding: contrastive decoding combining blank and language-prior contrasts, gated and ensembled, raises Qwen2-Audio accuracy from 57.1% to 71.9%, beating AAD alone.1
• A unified causal-grounding profile does not transfer between models, indicating different grounding regimes.1

Sources (1)

• [1]Does Location Matter? A Black-Box Causal Audit Reveals Model-Dependent Audio Grounding in LALM Hallucination
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 07:06 AM

  Large audio-language models (LALMs) frequently exhibit object hallucination, affirming absent sounds.

  Holding duration fixed, localization beats random removal at every cell on both models.

Extractive summary: sentences quoted from the sources.