Elamur Research⌕
☰
home / papers / 2026 / Think Smart, Not Hard: Difficulty Adaptive Reasoning for Large Audio Language Models

Think Smart, Not Hard: Difficulty Adaptive Reasoning for Large Audio Language Models

Paper · 2026 · Zhichao Sheng, Shilin Zhou, Chen Gong et al.

Summary

Large Audio Language Models (LALMs) that use chain-of-thought tend to apply the same reasoning depth to every question. They overthink easy items and under-reason on hard ones. Existing fixes either decide only whether to reason or impose fixed length penalties. The authors first compare SFT with GRPO on Qwen2-Audio-7B-Instruct and Qwen2.5-Omni-7B, using the FS and AVQA training sets. They find that GRPO's explicit reasoning helps on medium and hard questions but slightly hurts easy ones. They then propose a length-based reward whose strength depends on how difficult the model itself finds a question. Difficulty is estimated in two ways: GRDR uses rollout group accuracy, and GA²DR uses audio attention entropy. Both reuse quantities already computed during training. On MMAU, MMAU-v0515 and MMAR, the rewards beat plain GRPO while cutting average reasoning length by at least half. This shows that efficiency and accuracy can improve together when reasoning depth tracks difficulty.

Key points

  • Training data: FS, which combines AVQA (36,036 items) with AudioGrounding, VocalSound, TUT2017 and Clotho-AQA, is about twice the size of AVQA. Evaluation uses MMAU-Test-Mini, MMAU-v0515 and MMAR, with Airbench-Foundation also used in the analysis.
  • GRDR: With a rollout group of G = 8, a question counts as easy (γ = 0) at 6 or more correct answers, medium (γ = 0.5) at 3–5, and hard (γ = 1) below 3.
  • GA²DR: Difficulty is measured from the entropy of the last token's attention over audio tokens, averaged across heads and min-max normalized per batch. The final attention layer worked best.
  • Entropy as a difficulty signal: On MMAU-test-mini, accuracy falls from 72.85 to 69.52 as entropy rises across three intervals.
  • Fixed vs. adaptive rewards: Fixed length rewards do beat plain GRPO but still trail the difficulty-adaptive versions. Code is released on GitHub (zcs-hlt).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 worksSpeech & spoken language · 480 works

Cite

Cite as Zhichao Sheng, Shilin Zhou, Chen Gong and Zhenghua Li. 2026. Think Smart, Not Hard: Difficulty Adaptive Reasoning for Large Audio Language Models. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1640

@inproceedings{sheng2026think,
  title = {{Think Smart, Not Hard: Difficulty Adaptive Reasoning for Large Audio Language Models}},
  author = {Zhichao Sheng, Shilin Zhou, Chen Gong and Zhenghua Li},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1640},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1640}
}

References

27 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Daman Arora and Andrea Zanette. 2025. Training language models to reason efficiently. CoRR, abs/2502.04463. find
  2. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou. 2024. Qwen2-audio technical report. CoRR, abs/2407.10759.
  3. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 81 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llm…
  4. Yuan Gong, Jin Yu, and James R. Glass. 2022. Vocalsound: A dataset for improving human vocal sounds recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022. IEEE. find
  5. KimiTeam, Ding Ding, Zeqian Ju, Yichong Leng, Songxiang Liu, Tong Liu, Zeyu Shang, Kai Shen, Wei Song, Xu Tan, Heyi Tang, Zhengtao Wang, Chu Wei, Yifei Xin, Xinran Xu, Jianwei Yu, Yutao Zhang, Xinyu Zhou, Y. Charles, and 21 others. 2025. Kimi-audio technical report. CoRR, abs/2504.18425.
  6. Zhifeng Kong, Arushi Goel, Rohan Badlani, Wei Ping, Rafael Valle, and Bryan Catanzaro. 2024. Audio flamingo: A novel audio language model with few-shot learning and dialogue abilities. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. find
  7. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP 2023, Koblenz…
  8. Gang Li, Jizhong Liu, Heinrich Dinkel, Yadong Niu, Junbo Zhang, and Jian Luan. 2025. Reinforcement learning outperforms supervised fine-tuning: A case study on audio question answering. CoRR, abs/2503.11197. — [arXiv:2503.11197](https://arxiv.org/abs/2503.11197) find
  9. Samuel Lipping, Parthasaarathy Sudarsanam, Konstantinos Drossos, and Tuomas Virtanen. 2022. Clothoaqa: A crowdsourced dataset for audio question answering. In 30th European Signal Processing Conference, EUSIPCO 2022, Belgrade, Serbia, August 29 - Sept. 2, 2022. IEEE. find
  10. Wei Liu, Ruochen Zhou, Yiyun Deng, Yuzhen Huang, Junteng Liu, Yuntian Deng, Yizhe Zhang, and Junxian He. 2025. Learn to reason efficiently with adaptive length-based reward shaping. CoRR, abs/2505.15612. — [arXiv:2505.15612](https://arxiv.org/abs/2505.15612) find
  11. Ziyang Ma, Yinghao Ma, Yanqiao Zhu, Chen Yang, Yi-Wen Chao, Ruiyang Xu, Wenxi Chen, Yuanzhe Chen, Zhuo Chen, Jian Cong, Kai Li, Keliang Li, Siyou Li, Xinfeng Li, Xiquan Li, Zheng Lian, Yuzhe Liang, Minghao Liu, Zhikang Niu, and 15 others. 2025. MMAR: A challenging benchmark for deep reasoning in spe… — [arXiv:2505.13032](https://arxiv.org/abs/2505.13032) find
  12. Annamaria Mesaros, Toni Heittola, Aleksandr Diment, Benjamin Elizalde, Ankit Shah, Emmanuel Vincent, Bhiksha Raj, and Tuomas Virtanen. 2017. DCASE2017 challenge setup: Tasks, datasets and baseline system. In Proceedings of the Workshop on Detection and Classification of Acoustic Scenes and Events, D… find
  13. Ruotian Peng, Yi Ren, Zhouliang Yu, Weiyang Liu, and Yandong Wen. 2025. Simko: Simple pass@ k policy optimization. arXiv preprint arXiv:2510.14807. — [arXiv:2510.14807](https://arxiv.org/abs/2510.14807) find
  14. Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, Peng Li, Wei Wei, Jing Shao, Chaochao Lu, Yue Zhang, Xian-Sheng Hua, Bowen Zhou, and Yu Cheng. 2025. A survey of efficient reasoning for large reasoning models: Language, multimo…
  15. Andrew Rouditchenko, Saurabhchand Bhati, Edson Araujo, Samuel Thomas, Hilde Kuehne, Rogério Feris, and James R. Glass. 2025. Omni-r1: Do you really need audio to fine-tune your audio llm? CoRR, abs/2505.09439. — [arXiv:2505.09439](https://arxiv.org/abs/2505.09439) find
  16. S. Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha. 2025. MMAU: A massive multi-task audio understanding and reasoning benchmark. In The Thirteenth International Conference on Learning Representations, ICLR 20… find
  17. Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. 2025. DAST: difficulty-adaptive slow-thinking for large reasoning models. CoRR, abs/2503.04472. find
  18. Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2024. SALMONN: towards generic hearing abilities for large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024.
  19. Dingdong Wang, Jincenzi Wu, Junan Li, Dongchao Yang, Xueyuan Chen, Tianhua Zhang, and Helen Meng. 2025. MMSU: A massive multi-task spoken language understanding and reasoning benchmark. CoRR. find
  20. Shu Wu, Chenxing Li, Wenfu Wang, Hao Zhang, Hualei Wang, Meng Yu, and Dong Yu. 2025. Audio-thinker: Guiding audio language model when and how to think via reinforcement learning. arXiv preprint arXiv:2508.08039. — [arXiv:2508.08039](https://arxiv.org/abs/2508.08039) find
  21. Zhifei Xie, Mingbao Lin, Zihang Liu, Pengcheng Wu, Shuicheng Yan, and Chunyan Miao. 2025. Audio-reasoner: Improving reasoning capability in large audio language models. CoRR, abs/2503.02318.
  22. Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. 2025. Qwen2.5-omni technical report. CoRR, abs/2503.20215.
  23. Xuenan Xu, Heinrich Dinkel, Mengyue Wu, and Kai Yu. 2021. Text-to-audio grounding: Building correspondence between captions and sound events. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2021, Toronto, ON, Canada, June 6-11, 2021. IEEE. find
  24. Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. 2022. AVQA: A dataset for audio-visual question answering on videos. In MM ’22: The 30th ACM International Conference on Multimedia, Lisboa, Portugal, October 10 - 14, 2022. ACM. find
  25. Qian Yang, Jin Xu, Wenrui Liu, Yunfei Chu, Ziyue Jiang, Xiaohuan Zhou, Yichong Leng, Yuanjun Lv, Zhou Zhao, Chang Zhou, and Jingren Zhou. 2024. Air-bench: Benchmarking large audio-language models via generative comprehension. In Proceedings of the 62nd Annual Meeting of the Association for Computati…
  26. Zhisong Zhang, Yan Wang, Xinting Huang, Tianqing Fang, Hongming Zhang, Chenlong Deng, Shuaiyi Li, and Dong Yu. 2025. Attention entropy is a key factor: An analysis of parallel context encoding with full-attention-based pre-trained language models. In Proceedings of the 63rd Annual Meeting of the Ass…
  27. Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, Wenmeng Zhou, and Yingda Chen. 2025. SWIFT: A scalable lightweight infrastructure for fine-tuning. In AAAI-25, Sponsored by the Association for the Advancement of Artificial Inte… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sheng-2026-think-smart-not-hard-difficulty/