arXiv cs.AI· Yunuo Wang, Maoyuan Chen, Yuanjin Huang, Xing Liu·· 13 小时前AI 评分26
音频语言模型答案变化评估：同类别匹配对照协议
Beyond Answer Changes: Matched Same-Class Controls for Local Audio Evidence
AI 导读
针对音频语言模型，研究提出同类别匹配对照协议，区分答案变化源于事件类别响应还是波形改写敏感性。在600个合成声景和两个架构家族的三个模型上，不同类别替换比同类别替换多引发49.7至60.9个百分点的答案变化，同类别准确率与原准确率差距在5.2个百分点以内。在原始与同类别答案均正确的试验中，插入类别的声景加权选择性准确率为73.7%至81.9%。
正文 · 原文
Abstract:For audio-language models, an answer change after a local sound-event edit can reflect a response to the event's class, sensitivity to waveform rewriting, or both. We compare matched same-class and different-class replacements of a queried event. Both use the same interval, level estimate, donor fitting, and crossfade, while the same-class replacement preserves the correct answer. Across 600 synthetic soundscapes and three models from two architecture families, different-class replacements produce 49.7 to 60.9 percentage points more answer changes than same-class replacements. Same-class accuracy remains within 5.2 percentage points of original accuracy. Among trials with correct original and same-class answers, soundscape-weighted selective accuracy for the inserted class is 73.7 to 81.9%, averaging eligible trials within each soundscape and then weighting eligible soundscapes equally. This protocol tests how model answers respond to local changes in event class.
| Comments: | Under review at ICASSP 2027. Code and frozen evidence: this https URL |
| Subjects: | Sound (cs.SD); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.37605 [cs.SD] |
| (or arXiv:2610.37605v1 [cs.SD] for this version) | |
| https://doi.org/10.48550/arXiv.2610.37605 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yunuo Wang [view email]
[v1]
Tue, 27 Oct 2026 17:36:06 UTC (113 KB)
来源：arXiv cs.AI · arxiv.org