arXiv：cs.CL· Hua-Hsuan Chen, Xuanjun Chen, Wei-Chung Lu, Chih-Kai Yang, Yu-Xiang Lin, Sung-Feng Huang, Haibin Wu, Hung-yi Lee·· 16 小时前
LALM 后训练到底学到了什么？重新审视音频理解与推理能力
What Does LALM Post-training Learn? Rethinking Audio Understanding and Reasoning Capabilities
AI 导读
研究通过控制实验分离 LALM 的音频理解提升与题型适应，发现多项模型和后训练方法中，即使没有有效音频信息，报告的性能提升仍大部分存在，而一旦更换问题格式，提升又大幅消失。这表明后训练往往让模型学会利用任务特定的文本线索和答题模式，而非真正获得可泛化的音频理解与推理能力。
正文
Abstract:Post-training often improves the performance of large audio-language models on audio understanding and reasoning benchmarks. However, these improvements may reflect adaptation to specific question formats, such as multiple-choice questions, rather than genuine advances in audio understanding and reasoning capabilities. To investigate what LALMs actually learn from post-training, we conduct controlled experiments that disentangle improvements in audio understanding from adaptation to specific question formats. Across different models and post-training methods, we find that much of the reported improvement persists even without informative audio, but largely disappears when the question format changes. These findings suggest that post-training often teaches models to exploit task-specific textual cues and answer patterns rather than developing generalizable audio understanding and reasoning capabilities.
| Comments: | Work in progress |
| Subjects: | Sound (cs.SD); Computation and Language (cs.CL); Multimedia (cs.MM); Audio and Speech Processing (eess.AS) |
| Cite as: | arXiv:2610.19476 [cs.SD] |
| (or arXiv:2610.19476v1 [cs.SD] for this version) | |
| https://doi.org/10.48550/arXiv.2610.19476 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xuanjun Chen [view email]
[v1]
Wed, 14 Oct 2026 02:54:57 UTC (4,784 KB)
来源：arXiv：cs.CL · arxiv.org