Large audio-language models (LALMs) often generate intermediate reasoning before answering, but this long-form reasoning can drift from acoustic evidence toward language-continuation patterns. The authors of a new arXiv paper propose SKAD (Sketch-Guided Audio-Contrastive Decoding), a training-free inference-time method that keeps reasoning anchored to the input audio.
SKAD first builds an acoustic evidence sketch that captures question-relevant auditory information. It then performs contrastive decoding between an audio-conditioned branch and a sketch-only branch. The sketch acts as a shared semantic scaffold, making the masked branch more informative and pushing decoding decisions to rely more on the original audio rather than on language priors.
The method was evaluated on Qwen2.5-Omni-7B and Step-Audio-R1 across three benchmarks: MMAR, MMAU, and MMSU. SKAD consistently improved average accuracy for both models on all benchmarks. Ablations showed that direct contrastive decoding without the acoustic evidence sketch gave limited or negative gains, confirming that a reliable contrastive reference is essential before applying audio-contrastive decoding.