Recent large audio‑language models (LALMs) have explored adaptive reasoning to trade off accuracy and efficiency, yet most approaches rely on coarse difficulty estimates and ignore the many factors that drive reasoning effort. To address this, we derive five‑dimensional difficulty scores and compressed reasoning traces from complementary offline assessments, providing a fine‑grained characterization of each audio instance. Leveraging this representation, we introduce SWEAR, a difficulty‑aware reasoning framework that learns difficulty signals during supervised fine‑tuning and employs reinforcement learning to allocate reasoning effort across stages. Experiments on MMAU‑test‑mini and MMAR demonstrate that SWEAR improves overall accuracy while substantially reducing generated tokens, shortening reasoning for easy cases without sacrificing performance on hard ones.
Review