Elamur Research⌕
☰
home / papers / 2026 / SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning

SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning

Paper · 2026 · Peidong Wang, Zhiming Ma, Xin Dai et al.

Summary

Most phone-scam detectors work on ASR transcripts. This means recognition errors carry through to the decision, and acoustic signals such as tone, stress or staged background noise are lost. SAFE-QAQ addresses this by running a large audio-language model (LALM) directly on raw call audio, so no transcription step is needed. It handles three tasks: scenario classification, fraud detection and fraud-type classification.

Training proceeds in three stages:

  1. GRPO with rule-based rewards (accuracy, format and a reasoning-depth term) teaches the model step-by-step "slow-thinking" analysis (SAFE-RL).
  2. A Long2Short stage trims the reasoning chains using rejection sampling fine-tuning (SAFE-RS) and then length-constrained RL (SAFE-LS).
  3. Fine-tuning on audio chunks lets the model decide when it has heard enough during an ongoing call (SAFE-Real).

On TeleAntiFraud-Bench, the system reports state-of-the-art accuracy while producing much shorter reasoning and enabling early warnings. This matters because the system is already used in production, where it reduces manual review work and helps stop scams before money is lost.

Key points

  • Reported to be state of the art on TeleAntiFraud-Bench with 88.23 F1. Training and evaluation use the anonymized TeleAntiFraud-28k dataset, which the authors describe as the only open audio fraud corpus suitable for LALMs.
  • Long2Short cuts reasoning length by 48.87%. Average tokens fall from 205.76 (SAFE-RL) to 181.33 (SAFE-RS) and 104.02 (SAFE-LS), with no loss in quality.
  • SAFE-Real makes decisions after about 8.98 s of audio on average, or 1.25 dialogue turns, which is 81.4% faster than the non-real-time models.
    • Fraud detection stays strong at 88.93 F1.
    • Fraud-type F1 drops to 77.56.
  • Ablations show that SFT initialization is important: without it, SAFE-RL drops to 80.66. Skipping rejection sampling before length-constrained RL also lowers scores, from 87.49 to 86.31.
  • The system is deployed with telecom operators and processes over 70,000 calls per day. Code is available at github.com/Control-derek/SAFE-QAQ.

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

 Institutions

🎓Peking UniversityBeijing · China244 works · 39,359 citations🎓Northeastern UniversityShenyang · China42 works · 1,940 citations🏢China MobileBeijing · China9 works · 21 citations🎓Shanghai University of Electric PowerYangpu · China3 works · 13 citations

 Fields, topics and concepts

📚 Shelves 7The 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 worksSocial Sciences · 1,772 worksInformation Systems · 1,437 worksSociology and Political Science · 577 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 480 works

Cite

Cite as Peidong Wang, Zhiming Ma, Xin Dai, YongKang Liu, Shi Feng, Xiaocui Yang, Wenxing Hu, Zhihao Wang, Mingjun Pan, Li Yuan and Daling Wang. 2026. SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1201

@inproceedings{wang2026safe,
  title = {{SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning}},
  author = {Peidong Wang, Zhiming Ma, Xin Dai, YongKang Liu, Shi Feng, Xiaocui Yang, Wenxing Hu, Zhihao Wang, Mingjun Pan, Li Yuan and Daling Wang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1201},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1201}
}

References

32 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. \#\# Limitations Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  2. While SAFE-QAQ demonstrates superior performance, our experimental scope is inevitably constrained by the fact that TeleAntiFraud-28k is cur-Joymallya Chakraborty, Wei Xia, Anirban Majumder, Dan Ma, Walid Chaabene, and Naveed Janvekar. 2024. Detoxbench: Benchmarking large language models for multita… — [arXiv:2409.06072](https://arxiv.org/abs/2409.06072) find
  3. Chen-Wei Chang, Shailik Sarkar, Shutonu Mitra, Qi Zhang, Hossein Salemi, Hemant Purohit, Fengxiu Zhang, Michin Hong, Jin-Hee Cho, and Chang-Tien Lu. 2024. Exposing llm vulnerabilities: Adversarial scam detection and performance. In 2024 IEEE International Conference on Big Data (BigData), pages 3568… find
  4. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187.
  5. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, and 1 others. 2024. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759.
  6. Google DeepMind. 2024. Introducing gemini 2.0: our new ai model for the agentic era. find
  7. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  8. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  9. Sihao Hu, Tiansheng Huang, Ka-Ho Chow, Wenqi Wei, Yanzhao Wu, and Ling Liu. 2024. Zipzap: Efficient training of language models for large-scale fraud detection on blockchain. In Proceedings of the ACM Web Conference 2024, pages 2807–2816. find
  10. Ailin Huang, Boyong Wu, Bruce Wang, Chao Yan, Chen Hu, Chengli Feng, Fei Tian, Feiyu Shen, Jingbei Li, Mingrui Chen, and 1 others. 2025. Step-audio: Unified understanding and generation in intelligent speech interaction. arXiv preprint arXiv:2502.11946. length on large language models. In Findings o…
  11. Daniel Kahneman. 2011. macmillan. find
  12. Sukanth Korkanti. 2024. Enhancing financial fraud detection using llms and advanced data analytics. In 2024 2nd International Conference on Self Sustainable Artificial Intelligence Systems (ICSSAS), pages 1328–1334. IEEE. find
  13. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. find
  14. Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. In The Eleventh International Conference on Learning Representations.
  15. Zhiming Ma, Peidong Wang, Minhua Huang, Jinpeng Wang, Kai Wu, Xiangzhao Lv, Yachun Pang, Yin Yang, Wenjie Tang, and Yuchen Kang. 2025. TeleAntiFraud-28k: An audio-text slow-thinking dataset for telecom fraud detection. In Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), p… find
  16. OpenAI. 2025. Openai o3-mini. find
  17. Biqing Qi, Xinquan Chen, Junqi Gao, Dong Li, Jianxing Liu, Ligang Wu, and Bowen Zhou. 2024. Interactive continual learning: Fast and slow thinking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882–12892. find
  18. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. find
  19. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  20. Maohao Shen, Guangtao Zeng, Zhenting Qi, Zhang-Wei Hong, Zhenfang Chen, Wei Lu, Gregory Wornell, Subhro Das, David Cox, and Chuang Gan. 2025a. Satori: Reinforcement learning with chain-of-action-thought enhances llm reasoning via autoregressive search. arXiv preprint arXiv:2502.02508. find
  21. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.
  22. Zitong Shen, Sineng Yan, Youqian Zhang, Xiapu Luo, Grace Ngai, and Eugene Yujun Fu. 2025b. " it warned me just at the right moment": Exploring llm-based real-time detection of phone scams. arXiv preprint arXiv:2502.03964. — [arXiv:2502.03964](https://arxiv.org/abs/2502.03964) find
  23. Mingyu Jin, Qinkai Yu, Dong Shu, Haiyan Zhao, Wenyue Hua, Yanda Meng, Yongfeng Zhang, and Mengnan Du. 2024. The impact of reasoning step Gurjot Singh, Prabhjot Singh, and Maninder Singh. 2025. Advanced real-time fraud detection using rag-based llms. arXiv preprint arXiv:2501.15290. — [arXiv:2501.15290](https://arxiv.org/abs/2501.15290) find
  24. Maciej Świechowski, Konrad Godlewski, Bartosz Sawicki, and Jacek Mańdziuk. 2023. Monte carlo tree search: A review of recent modifications and applications. Artificial Intelligence Review, 56(3):2497– 2562. find
  25. DOUBAO TEAM. 2025. Doubao-1.5-pro. find
  26. Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, and 1 others. 2025. Llamav-o1: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186. find
  27. Luong Trung, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. 2024. Reft: Reasoning with reinforced fine-tuning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7601–7614.
  28. Zihan Wang, Xingle Xu, Hao Wang, Yiwen Ye, Yuchen Li, Linhao Wang, Hongze Tan, Peidong Wang, Shi Feng, Guoqing Chen, and 1 others. 2025. A survey on entropy mechanism in large reasoning models. Authorea Preprints. find
  29. Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. 2024. Llava-o1: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440. find
  30. Shu Yang, Shenzhe Zhu, Zeyu Wu, Keyu Wang, Junchi Yao, Junchao Wu, Lijie Hu, Mengdi Li, Derek F Wong, and Di Wang. 2025. Fraud-r1: A multi-round benchmark for assessing the robustness of llm against augmented fraud and phishing inducements. arXiv preprint arXiv:2502.12904. find
  31. Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang. 2024. Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot. arXiv preprint arXiv:2412.02612.
  32. Xinyu Zhu, Junjie Wang, Lin Zhang, Yuxiang Zhang, Ruyi Gan, Jiaxing Zhang, and Yujiu Yang. 2022. Solving math word problems via cooperative reasoning induced language models. arXiv preprint arXiv:2210.16257.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-safe-qaq-end-end-slow/