Experimental results show that SKAD consistently improves the average accuracy of both models across all benchmarks, and ablations show that direct contrastive decoding without the acoustic evidence sketch provides limited or negative gains, confirming the importance of constructing a reliable contrastive reference before audio-contrastive decoding.
Abstract
Large Audio-Language Models (LALMs) have shown increasing potential in audio understanding and reasoning by generating intermediate reasoning steps before answering. However, directly eliciting long-form reasoning from audio does not always lead to more reliable predictions. Even reasoning-oriented LALMs may gradually shift from acoustic evidence to language-continuation patterns during generation. A natural solution is to use contrastive decoding with an audio-masked branch, but directly applying this strategy to plain chain-of-thought reasoning is often ineffective, since the masked branch lacks sufficient audio-related context and provides an unstable reference. To address this issue, we propose \textbf{SKAD} (\textbf{Sk}etch-Guided \textbf{A}udio-Contrastive \textbf{D}ecoding), a training-free inference-time method for audio reasoning. SKAD first constructs an acoustic evidence sketch that summarizes question-relevant auditory information. It then performs contrastive decoding between an audio-conditioned reasoning branch and a sketch-only branch. By using the sketch as a shared semantic scaffold, SKAD makes the masked branch more informative and encourages decoding decisions that are more dependent on the original audio. We evaluate SKAD on Qwen2.5-Omni-7B and Step-Audio-R1 across three audio understanding and reasoning benchmarks: MMAR, MMAU, and MMSU. Experimental results show that SKAD consistently improves the average accuracy of both models across all benchmarks. Further ablations show that direct contrastive decoding without the acoustic evidence sketch provides limited or negative gains, confirming the importance of constructing a reliable contrastive reference before audio-contrastive decoding.
Large audio language models answer questions about speech, sound, and music, yet their accuracy drops sharply on tasks that need fine-grained acoustic analysis. Judging which of two speakers has the higher pitch demands iterative signal-level reasoning that a content question does not. Current models spend the same com...
JELAR, a Joint-Embedding Predictive Architecture (JEPA)-based latent reasoning framework that conditions latent reasoning supervision on acoustic representations learned from raw waveforms, demonstrates the effectiveness of JEPA-conditioned latent reasoning as an alternative to explicit textual CoT supervision.
Donghang Wu, Haoyang Zhang, Yizhou Peng et al.· 0 citations
Large audio-language models (LALMs) excel at understanding and reasoning tasks over atomic sound events, yet their ability to infer higher-level human activities from such fine-grained events remains largely unexamined. Everyday human actions and activities, such as setting a table, cleaning the house, or preparing a b...
Michel Olvera, Paraskevas Stamatiadis, Chang-Hong Wang et al.· Conference on Empirical Meth...· 0 citations
Large audio-language models (LALMs) have advanced music understanding and driven progress in music information retrieval (MIR), which requires integrating low-level acoustic perception and high-level semantic interpretation. These two levels impose markedly different reasoning demands: over-deliberation on perceptual t...
Wei Lan, Yin-Yi Feng, Zu-Xuan Wu· Proceedings of the 34th ACM...· 0 citations
Perception-Grounded TTRL (PG-TTRL) is proposed, which aligns label-free test-time optimisation with perceptually grounded reasoning, encouraging the model to structure its reasoning more strongly on the audio input, showing the value of perceptual-grounding optimisation for test-time audio reasoning.
Jia-Heng Dong, Xiao-Feng Yu, Jean Honorio et al.· arXiv.org· 0 citations
Mizar, a 159.3M-parameter ALM, is introduced, a recipe that brings together architecture, data, and three-stage training to build Mizar, a 159.3M-parameter ALM that surpasses the previous best-performing ALM below 200M parameters on all three benchmarks.