---
title: "Step-Audio-R1 Technical Report — paper by Tian et al. (2025)"
description: "The StepFun team argues that the cause is \"textual surrogate reasoning\". 29 references, cited by 3 papers here."
authors: ["Fei Tian", "Xiangyu Tony Zhang", "Yuxin Zhang", "Haoyang Zhang", "Yuxin Li", "Daijiao Liu", "Yayue Deng", "Donghang Wu", "Jun Chen", "Liang Zhao", "Chengyuan Yao", "Hexin Liu", "Eng Siong Chng", "Xuerui Yang", "Xiangyu Zhang", "Daxin Jiang", "Gang Yu"]
year: 2025
venue: "arXiv:2511.15848"
source: "https://arxiv.org/abs/2511.15848"
doi: "10.48550/arxiv.2511.15848"
arxiv: "2511.15848"
licence: "arXiv — non-exclusive licence to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/tian-2025-step-audio-r1-technical/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/tian-2025-step-audio-r1-technical/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/13/) / Step-Audio-R1 Technical Report

# Step-Audio-R1 Technical Report

**Authors:** [Fei Tian](https://md.research.elamur.ai/free/en/authors/fei-tian/), Xiangyu Tony Zhang, Yuxin Zhang, [Haoyang Zhang](https://md.research.elamur.ai/free/en/authors/haoyang-zhang/), [Yuxin Li](https://md.research.elamur.ai/free/en/authors/yuxin-li/), Daijiao Liu, [Yayue Deng](https://md.research.elamur.ai/free/en/authors/yayue-deng/), [Donghang Wu](https://md.research.elamur.ai/free/en/authors/donghang-wu/), [Jun Chen](https://md.research.elamur.ai/free/en/authors/jun-chen/), [Liang Zhao](https://md.research.elamur.ai/free/en/authors/liang-zhao/), [Chengyuan Yao](https://md.research.elamur.ai/free/en/authors/chengyuan-yao/), [Hexin Liu](https://md.research.elamur.ai/free/en/authors/hexin-liu/), Eng Siong Chng, Xuerui Yang, [Xiangyu Zhang](https://md.research.elamur.ai/free/en/authors/xiangyu-zhang/), [Daxin Jiang](https://md.research.elamur.ai/free/en/authors/daxin-jiang/) and [Gang Yu](https://md.research.elamur.ai/free/en/authors/gang-yu/) · **Year:** 2025 · **Venue:** arXiv:2511.15848 · **Licence:** [arXiv — non-exclusive licence to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2511.15848) · [PDF](https://arxiv.org/pdf/2511.15848) · [DOI 10.48550/arxiv.2511.15848](https://doi.org/10.48550/arxiv.2511.15848) · [arXiv 2511.15848](https://arxiv.org/abs/2511.15848) · [OpenAlex](https://openalex.org/W7106292994)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [3](https://md.research.elamur.ai/free/en/papers/tian-2025-step-audio-r1-technical/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
Audio language models usually perform worse when they reason at length, which runs against the test-time compute scaling seen in text and vision models. The StepFun team argues that the cause is "textual surrogate reasoning". Models learn their chain-of-thought from text-derived data, so they reason over transcripts or captions instead of the sound itself. Step-Audio-R1 is built on the Step-Audio 2 design: a frozen Qwen2 audio encoder, an adaptor and a Qwen2.5 32B decoder. Training starts with a cold start that combines SFT and RLVR. Next comes Modality-Grounded Reasoning Distillation (MGRD), in which the model repeatedly self-distills on perception-heavy audio questions. Only reasoning chains that are explicitly tied to acoustic cues such as timbre, pitch and rhythm are kept. The authors report that the resulting model beats Gemini 2.5 Pro and is roughly on par with Gemini 3 Pro on speech, sound and music benchmarks. This makes it the first audio model shown to gain from extended deliberation. The result suggests that reasoning can carry over across modalities if it is grounded in the right input.

## Key points
- **Architecture:** The Qwen2 audio encoder runs at 25 Hz and stays frozen throughout training. An adaptor downsamples by 2 to reach 12.5 Hz, and a Qwen2.5 32B LLM decoder produces a reasoning block followed by the answer.
- **Cold-start data:** about 5M samples.
  - The token budget is roughly 1B text tokens and 4B audio-side tokens.
  - Self-distilled audio CoT makes up 10% of the audio data.
  - Samples without CoT are given an empty `<think>` block.
- **RL stage:**
  - It uses a compact set of 5,000 samples: 2,000 text math/code items and 3,000 augmented speech QA items.
  - Optimization is PPO with binary verified rewards and no KL penalty.
- **Self-cognition fix:** This addresses the model wrongly claiming it cannot hear audio. On a 5,000-sample test set, the error rate fell in three steps:

  | Stage | Error rate |
  |---|---|
  | Base model | 6.76% |
  | After filtered iterative self-distillation | 2.63% |
  | After DPO with 8,000 preference pairs | 0.02% |

- **Variant:** A Step-Audio-R1-Realtime version is also evaluated on speech-to-speech benchmarks. It is compared with GPT-Realtime and Gemini-2.5-Flash native audio dialog.

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

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 493 works

🧭 Topics (2): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,685 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 519 works

## Cite

Cite as: Fei Tian, Xiangyu Tony Zhang, Yuxin Zhang, Haoyang Zhang, Yuxin Li, Daijiao Liu, Yayue Deng, Donghang Wu, Jun Chen, Liang Zhao, Chengyuan Yao, Hexin Liu, Eng Siong Chng, Xuerui Yang, Xiangyu Zhang, Daxin Jiang and Gang Yu. 2025. Step-Audio-R1 Technical Report. arXiv:2511.15848.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{tian2025step,
  title = {{Step-Audio-R1 Technical Report}},
  author = {Fei Tian, Xiangyu Tony Zhang, Yuxin Zhang, Haoyang Zhang, Yuxin Li, Daijiao Liu, Yayue Deng, Donghang Wu, Jun Chen, Liang Zhao, Chengyuan Yao, Hexin Liu, Eng Siong Chng, Xuerui Yang, Xiangyu Zhang, Daxin Jiang and Gang Yu},
  year = {2025},
  doi = {10.48550/arxiv.2511.15848},
  eprint = {2511.15848},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2511.15848}
}
```

- [← previous Speech Recognition for Automatically Assessing Afrikaans and isiXhosa Preschool Oral Narratives](https://md.research.elamur.ai/free/en/papers/jacobs-2025-speech-recognition-automatically-assessing-afrikaans/)
- [next → TafBERTa: Learning Grammatical Rules from Small-Scale Language Acquisition Data in Hebrew](https://md.research.elamur.ai/free/en/papers/gelboim-2025-tafberta-learning-grammatical-rules-small/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/13/)

## References

29 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/). Advances in neural information processing systems, 35:24824–24837, 2022.
2. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. [Gpt-4 technical report](https://md.research.elamur.ai/free/en/papers/openai-2023-gpt-4-technical-report/). arXiv preprint arXiv:2303.08774, 2023.
3. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. [Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning](https://md.research.elamur.ai/free/en/papers/deepseek-ai-2025-deepseek-r1-incentivizing-reasoning-capability/). arXiv preprint arXiv:2501.12948, 2025.
4. Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. [Gemini: a family of highly capable multimodal models](https://md.research.elamur.ai/free/en/papers/gemini-team-2023-gemini-family-highly-capable-multimodal/). arXiv preprint arXiv:2312.11805, 2023.
5. Bin Wang, Bojun Wang, Changyi Wan, Guanzhe Huang, Hanpeng Hu, Haonan Jia, Hao Nie, Mingliang Li, Nuo Chen, Siyu Chen, et al. [Step-3 is large yet affordable: Model-system co-design for cost-effective decoding](https://md.research.elamur.ai/free/en/papers/stepfun-2025-step-3-large-yet/). arXiv preprint arXiv:2507.19427, 2025.
6. Wei Shen, Jiangbo Pei, Yi Peng, Xuchen Song, Yang Liu, Jian Peng, Haofeng Sun, Yunzhuo Hao, Peiyu Wang, Jianhao Zhang, et al. [Skywork-r1v3 technical report](https://md.research.elamur.ai/free/en/papers/shen-2025-skywork-r1v3-technical-report/). arXiv preprint arXiv:2507.06167, 2025.
7. Zhuosheng Zhang, Aston Zhang, Mu Li, George Karypis, Alex Smola, et al. [Multimodal chain-of-thought reasoning in language models](https://md.research.elamur.ai/free/en/papers/zhang-2023-multimodal-chain-thought-reasoning-language/). Transactions on Machine Learning Research.
8. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations, 2025. [find](https://scholar.google.com/scholar?q=Charlie+Victor+Snell%2C+Jaehoon+Lee%2C+Kelvin+Xu%2C+and+Aviral+Kumar%2E+Scaling+llm+test%2Dtime+compute+optimally+can+be+more+effective+than+scaling+parameters+for+reasoning%2E+In+The+Thirteenth+International+Con)
9. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. [Openai o1 system card](https://md.research.elamur.ai/free/en/papers/openai-2024-openai-o1-system-card/). arXiv preprint arXiv:2412.16720, 2024.
10. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in University of New South Wales † Project Leader language models. In The Eleventh International Conference on Learning Representatio… [find](https://scholar.google.com/scholar?q=Xuezhi+Wang%2C+Jason+Wei%2C+Dale+Schuurmans%2C+Quoc+V+Le%2C+Ed+H+Chi%2C+Sharan+Narang%2C+Aakanksha+Chowdhery%2C+and+Denny+Zhou%2E+Self%2Dconsistency+improves+chain+of+thought+reasoning+in+University+of+New+South+Wales+)
11. Gang Li, Jizhong Liu, Heinrich Dinkel, Yadong Niu, Junbo Zhang, and Jian Luan. Reinforcement learning outperforms supervised fine-tuning: A case study on audio question answering. arXiv preprint arXiv:2503.11197, 2025. — [arXiv:2503.11197](https://arxiv.org/abs/2503.11197) [find](https://scholar.google.com/scholar?q=Gang+Li%2C+Jizhong+Liu%2C+Heinrich+Dinkel%2C+Yadong+Niu%2C+Junbo+Zhang%2C+and+Jian+Luan%2E+Reinforcement+learning+outperforms+supervised+fine%2Dtuning%3A+A+case+study+on+audio+question+answering%2E+arXiv+preprint+arXiv)
12. Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, et al. [Qwen3-omni technical report](https://md.research.elamur.ai/free/en/papers/xu-2025-qwen3-omni-technical-report/). arXiv preprint arXiv:2509.17765, 2025.
13. Boyong Wu, Chao Yan, Chen Hu, Cheng Yi, Chengli Feng, Fei Tian, Feiyu Shen, Gang Yu, Haoyang Zhang, Jingbei Li, et al. [Step-audio 2 technical report](https://md.research.elamur.ai/free/en/papers/wu-2025-step-audio-2-technical/). arXiv preprint arXiv:2507.16632, 2025.
14. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. [Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities](https://md.research.elamur.ai/free/en/papers/comanici-2025-gemini-2-5-pushing-frontier/). arXiv preprint a…
15. Jiajun Fan, Roger Ren, Jingyuan Li, Rahul Pandey, Prashanth Gurunath Shivakumar, Ivan Bulyko, Ankur Gandhe, Ge Liu, and Yile Gu. Incentivizing consistent, effective and scalable reasoning capability in audio llms via reasoning process rewards. arXiv preprint arXiv:2510.20867, 2025. — [arXiv:2510.20867](https://arxiv.org/abs/2510.20867) [find](https://scholar.google.com/scholar?q=Jiajun+Fan%2C+Roger+Ren%2C+Jingyuan+Li%2C+Rahul+Pandey%2C+Prashanth+Gurunath+Shivakumar%2C+Ivan+Bulyko%2C+Ankur+Gandhe%2C+Ge+Liu%2C+and+Yile+Gu%2E+Incentivizing+consistent%2C+effective+and+scalable+reasoning+capability+i)
16. Shu Wu, Chenxing Li, Wenfu Wang, Hao Zhang, Hualei Wang, Meng Yu, and Dong Yu. Audio-thinker: Guiding audio language model when and how to think via reinforcement learning. arXiv preprint arXiv:2508.08039, 2025. — [arXiv:2508.08039](https://arxiv.org/abs/2508.08039) [find](https://scholar.google.com/scholar?q=Shu+Wu%2C+Chenxing+Li%2C+Wenfu+Wang%2C+Hao+Zhang%2C+Hualei+Wang%2C+Meng+Yu%2C+and+Dong+Yu%2E+Audio%2Dthinker%3A+Guiding+audio+language+model+when+and+how+to+think+via+reinforcement+learning%2E+arXiv+preprint+arXiv%3A2508%2E0)
17. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/), 2023. URL https://arxiv.org/abs/2201.11903. [↗](https://arxiv.org/abs/2201.11903)
18. Google DeepMind. Gemini 3. [https://deepmind.google/models/gemini/](https://deepmind.google/models/gemini/), 2025. Accessed: 2025. [↗](https://deepmind.google/models/gemini/)
19. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. [Qwen2-audio technical report](https://md.research.elamur.ai/free/en/papers/chu-2024-qwen2-audio-technical-report/). arXiv preprint arXiv:2407.10759, 2024.
20. Qwen Team et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2(3), 2024. [find](https://scholar.google.com/scholar?q=Qwen+Team+et+al%2E+Qwen2+technical+report%2E+arXiv+preprint+arXiv%3A2407%2E10671%2C+2%283%29%2C+2024%2E)
21. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2023. — [Let's Verify Step by Step](https://md.research.elamur.ai/free/en/papers/lightman-2023-let-s-verify-step-step/)
22. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. [find](https://scholar.google.com/scholar?q=John+Schulman%2C+Filip+Wolski%2C+Prafulla+Dhariwal%2C+Alec+Radford%2C+and+Oleg+Klimov%2E+Proximal+policy+optimization+algorithms%2E+arXiv+preprint+arXiv%3A1707%2E06347%2C+2017%2E)
23. Dingdong Wang, Jincenzi Wu, Junan Li, Dongchao Yang, Xueyuan Chen, Tianhua Zhang, and Helen Meng. Mmsu: A massive multi-task spoken language understanding and reasoning benchmark. arXiv preprint arXiv:2506.04779, 2025. — [arXiv:2506.04779](https://arxiv.org/abs/2506.04779) [find](https://scholar.google.com/scholar?q=Dingdong+Wang%2C+Jincenzi+Wu%2C+Junan+Li%2C+Dongchao+Yang%2C+Xueyuan+Chen%2C+Tianhua+Zhang%2C+and+Helen+Meng%2E+Mmsu%3A+A+massive+multi%2Dtask+spoken+language+understanding+and+reasoning+benchmark%2E+arXiv+preprint+arXiv)
24. S Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha. [Mmau: A massive multi-task audio understanding and reasoning benchmark](https://md.research.elamur.ai/free/en/papers/sakshi-2024-mmau-massive-multi-task-audio/). arXiv preprint arXiv:2410.19168, 2024.
25. Chengwei Wei, Bin Wang, Jung-jae Kim, and Nancy F Chen. Towards spoken mathematical reasoning: Benchmarking speech-based models over multi-faceted math problems. arXiv preprint arXiv:2505.15000, 2025. — [arXiv:2505.15000](https://arxiv.org/abs/2505.15000) [find](https://scholar.google.com/scholar?q=Chengwei+Wei%2C+Bin+Wang%2C+Jung%2Djae+Kim%2C+and+Nancy+F+Chen%2E+Towards+spoken+mathematical+reasoning%3A+Benchmarking+speech%2Dbased+models+over+multi%2Dfaceted+math+problems%2E+arXiv+preprint+arXiv%3A2505%2E15000%2C+2025%2E)
26. Jian Zhang, Linhao Zhang, Bokai Lei, Chuhan Wu, Wei Jia, and Xiao Zhou. Wildspeech-bench: Benchmarking audio llms in natural speech conversation. arXiv preprint arXiv:2506.21875, 2025. — [arXiv:2506.21875](https://arxiv.org/abs/2506.21875) [find](https://scholar.google.com/scholar?q=Jian+Zhang%2C+Linhao+Zhang%2C+Bokai+Lei%2C+Chuhan+Wu%2C+Wei+Jia%2C+and+Xiao+Zhou%2E+Wildspeech%2Dbench%3A+Benchmarking+audio+llms+in+natural+speech+conversation%2E+arXiv+preprint+arXiv%3A2506%2E21875%2C+2025%2E+%E2%80%94+%5BarXiv%3A2506%2E2)
27. Donghang Wu, Haoyang Zhang, Chen Chen, Tianyu Zhang, Fei Tian, Xuerui Yang, Gang Yu, Hexin Liu, Nana Hou, Yuchen Hu, et al. [Chronological thinking in full-duplex spoken dialogue language models](https://md.research.elamur.ai/free/en/papers/wu-2026-chronological-thinking-full-duplex-spoken/). arXiv preprint arXiv:2510.05150, 2025.
28. Donghang Wu, Haoyang Zhang, Jun Chen, Hexin Liu, Eng Siong Chng, Fei Tian, Xuerui Yang, Xiangyu Zhang, Daxin Jiang, Gang Yu, et al. Mind-paced speaking: A dual-brain approach to real-time reasoning in spoken language models. arXiv preprint arXiv:2510.09592, 2025. — [arXiv:2510.09592](https://arxiv.org/abs/2510.09592) [find](https://scholar.google.com/scholar?q=Donghang+Wu%2C+Haoyang+Zhang%2C+Jun+Chen%2C+Hexin+Liu%2C+Eng+Siong+Chng%2C+Fei+Tian%2C+Xuerui+Yang%2C+Xiangyu+Zhang%2C+Daxin+Jiang%2C+Gang+Yu%2C+et+al%2E+Mind%2Dpaced+speaking%3A+A+dual%2Dbrain+approach+to+real%2Dtime+reasoning+in)
29. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. [Direct preference optimization: Your language model is secretly a reward model](https://md.research.elamur.ai/free/en/papers/rafailov-2023-direct-preference-optimization-language-model/). Advances in neural information processing systems, 36:53728–53741, 2023.

*Updated 2 November 2026*