arXiv：cs.AI· Yixuan Li, Minghao Lv, Xuenan Xu, Zihao Zheng, Mengyue Wu·· 13 小时前
SWEAR：通过自评难度实现分段高效音频推理
SWEAR: Section-Wise Efficient Audio Reasoning via Self-Rated Difficulty
AI 导读
SWEAR 是一个难度感知音频推理框架，通过离线评估得到五维难度分数与压缩推理轨迹，在监督微调中学习难度信号，并用强化学习按推理阶段动态分配推理量。在 MMAU-test-mini 和 MMAR 上，SWEAR 在提升准确率的同时大幅减少生成 token，对简单样本缩短推理且不牺牲困难样本性能。
正文
Abstract:Recent work on large audio-language models (LALMs) explores adaptive reasoning to balance accuracy and efficiency. However, existing approaches often rely on coarse-grained difficulty estimation, overlooking the diverse factors that determine reasoning effort. We address this limitation by deriving five-dimensional difficulty scores and compressed reasoning traces from complementary offline assessments. Based on this characterization, we propose SWEAR, a difficulty-aware reasoning framework that learns difficulty signals during supervised fine-tuning and uses reinforcement learning to adapt reasoning allocation across reasoning stages. Experiments on MMAU-test-mini and MMAR show that SWEAR improves accuracy while substantially reducing generated tokens, shortening reasoning for easy instances without sacrificing performance on difficult ones.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.13081 [cs.AI] |
| (or arXiv:2610.13081v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.13081 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yixuan Li [view email]
[v1]
Fri, 9 Oct 2026 06:36:43 UTC (460 KB)
来源：arXiv：cs.AI · arxiv.org