{"ID":5552860,"CreatedAt":"2026-07-02T01:54:51.863792489Z","UpdatedAt":"2026-07-03T22:44:45.074508885Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2607.00247","arxiv_id":"2607.00247","title":"Adaptive Perturbation Selection for Contrastive Audio Decoding","abstract":"Large audio-language models (LALMs) frequently hallucinate by overriding acoustic evidence with language priors. While contrastive decoding (CD) offers training-free mitigation, existing methods rely on blunt perturbations like masking or noise, leaving structured audio transformations unexplored. We explore this design space by evaluating a diverse library of targeted audio perturbations and adaptively selecting the optimal negative branch for each task and example. First, we improve upon earlier prompt engineering by showing that a simple binary yes/no constraint reduces the model's tendency to falsely confirm absent audio features. Second, evaluating our library across temporal, spectral, frequency, and amplitude domains reveals that optimal transformations are highly task-dependent; for instance, reversing the audio array disrupts temporal coherence, raising accuracy on the temporal order task from 74.7% to 81.4%. Finally, we trained a light-weight perturbation selector on model hidden states to dynamically route negative branches, yielding an additional +4.3% gain on the existence task.","short_abstract":"Large audio-language models (LALMs) frequently hallucinate by overriding acoustic evidence with language priors. While contrastive decoding (CD) offers training-free mitigation, existing methods rely on blunt perturbations like masking or noise, leaving structured audio transformations unexplored. We explore this desig...","url_abs":"https://arxiv.org/abs/2607.00247","url_pdf":"https://arxiv.org/pdf/2607.00247v1","authors":"[\"Aaron Isidore Grace\",\"Zhouyuan Huo\",\"Weiran Wang\"]","published":"2026-06-30T22:55:22Z","proceeding":"cs.SD","tasks":"[\"cs.SD\",\"cs.AI\"]","methods":"[\"Language Model\"]","has_code":false}