arXiv · 2025 · Preprint
Ma et al. (Shanghai Jiao Tong University) · → Paper · Demo: ? · Code: ?
The first systematic evaluation of Chain-of-Thought (CoT) reasoning methods applied to Large Audio-Language Models, benchmarked across sound, music, and speech understanding tasks on MMAU.
Citation Stub
This paper is not a speech generation paper but is cited by the corpus. See Context in Speech Generation below for why it is relevant.
Context in Speech Generation
Audio-CoT applies three established CoT prompting strategies (Manual-CoT, Zero-Shot-CoT, and Desp-CoT) to Qwen2-Audio-7B-Instruct and evaluates them on the MMAU benchmark, covering information extraction and reasoning tasks across three audio modalities. The study finds that CoT methods improve accuracy on easy and medium difficulty questions but can degrade performance on hard questions, where explicit reasoning chains introduce confusion rather than clarity. It also identifies a positive correlation between reasoning path length and accuracy, suggesting that inference-time scaling can benefit audio-language reasoning. The findings are relevant to speech LMs that incorporate reasoning capabilities, as they reveal both the promise and the current ceiling of CoT-based inference in multimodal audio systems.
Wiki Connections
- spoken-language-model — the paper evaluates CoT reasoning in Large Audio-Language Models, which are architecturally related to spoken language models that process audio with LLM backbones.
- evaluation-metrics — the paper uses the MMAU benchmark as its primary evaluation framework, contributing empirical data on how CoT affects task-level accuracy across difficulty tiers.
- 2311.07919 (Qwen-Audio) — the paper’s evaluation model, Qwen2-Audio-7B-Instruct, is a successor to Qwen-Audio; the study builds directly on this model family’s audio-language capabilities.
- 2407.10759 (Qwen2-Audio Technical Report) — the specific model used for all Audio-CoT experiments, adopted for its strong baseline performance across audio benchmarks.
- 2402.13236 (Towards Audio Language Modeling: An Overview) — a survey cited as background on the LALM landscape that motivates investigating reasoning as the next capability frontier.