TRACE：对大音频语言模型进行源级反事实审计，发现正确答案背后的证据不一致
Right Answers, Inconsistent Evidence: Source-Level Counterfactual Auditing of Large Audio-Language Models
AI 导读
研究者提出 TRACE 方法，通过逐个删除音频中的目标声源和干扰源，审计大音频语言模型（LALM）是否真正区分目标的存在与缺失。在 8 个 LALM 和 3,456 个受控混合音频上，6.3% 的正确答案无法通过该排序检验，仅用 target-only 分数检验会漏掉其中 42.5%；当目标 SNR 从 +5 降至 -5 dB 时，失败比例从 4.9% 升至 8.4%。
正文 · 原文
Abstract:A large audio-language model (LALM) can answer "is a cat present?" about a mixture correctly, and its confidence can drop when the cat is removed, yet this does not show that it can tell the cat's presence from its absence. We introduce TRACE. It deletes the queried source and each rival in turn and asks whether every view still containing that sound outscores the view with it removed. Across 8 LALMs and 3,456 controlled mixtures, 6.3% of correct answers fail this ordering: the target-absent view scores at or above some target-present view, a conflict no shared threshold repairs. A target-only score check misses 42.5% of them. The share rises from 4.9% to 8.4% as target SNR falls from +5 to -5 dB, even on units correct at every level, and survives loudness and re-pairing controls. Accuracy certifies the answer; TRACE audits whether it orders presence above absence. Code and data: this https URL.
| Comments: | 5 pages, 3 figures, 1 table. Submitted to ICASSP 2027 |
| Subjects: | Sound (cs.SD); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.15463 [cs.SD] |
| (or arXiv:2611.15463v1 [cs.SD] for this version) | |
| https://doi.org/10.48550/arXiv.2611.15463 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yi Zhang [view email]
[v1]
Wed, 11 Nov 2026 13:32:23 UTC (185 KB)
来源：arXiv cs.AI · arxiv.org