Elamur Research⌕
☰
home / papers / 2026 / DuIVRS-2: An LLM-based Interactive Voice Response System for Large-scale POI Attribute Acquisition

DuIVRS-2: An LLM-based Interactive Voice Response System for Large-scale POI Attribute Acquisition

Paper · 2026 · Le Zhang, Shengming Zhang, Rui Zha et al.

Summary

Baidu Maps must keep attributes such as business hours and addresses current for hundreds of millions of POIs. Its earlier phone-based system, DuIVRS-1, used a modular NLU–DM–NLG pipeline in which errors compound and maintenance is expensive. DuIVRS-2 replaces this pipeline with a fine-tuned compact LLM that acts as an end-to-end agent calling POI owners.

To counter the long-tail distribution of the call logs, the authors build a finite state machine from the legacy templates and historical logs. They then sample dialogue paths and user replies uniformly rather than by frequency, which yields balanced synthetic training data. A selective generation scheme with CoT keeps outputs stable and prevents invented attributes. A ReST-inspired cooperative iterative learning loop has two LLM evaluators vote on responses and sends only uncertain cases to humans.

In a two-month online A/B test, DuIVRS-2 reached 83.9% TSR, compared with 79.9% for DuIVRS-1 and 89.6% for human operators. It cost under ¥0.2 per call with a 130ms reaction time. The work is a practical blueprint for latency-constrained industrial LLM dialogue agents.

Key points

  • FSM-guided augmentation. Legacy response templates become FSM states and user replies become transitions. Paths are sampled evenly across length groups, and replies are sampled evenly between states. This gives the model training examples for infrequent dialogue states that would otherwise have almost none, and it does so without gathering any new data.

  • Model setup. The deployed model is an LLM-S with fewer than 2B parameters. Iterative learning also updates a larger fine-tuned LLM-L and refines a Black-box LLM evaluator via in-context learning. Ablations show that removing either evaluator hurts performance, and the share of data needing human judgment shrinks over iterations.

  • Deployment optimizations.

    • Static graph conversion in PaddlePaddle.
    • int8 weight quantization.
    • FastDeploy for scheduling and batching.
    • The ASR and TTS modules from DuIVRS-1 are reused.

  • Online A/B comparison (Table 2).

    SystemTSRCost per callReaction time
    Human89.6%¥1.5500ms
    DuIVRS-179.9%<¥0.215ms
    DuIVRS-283.9%<¥0.2130ms

    DuIVRS-2 reaches about 93.6% of human performance.

  • Call volume. During the test, DuIVRS-2 was allotted only about 3,000 calls per day within a one-hour window. The figure of 0.4 million calls per day in the paper is a capacity claim (human operators handle at most about 200 per day). The abstract, however, phrases it as daily processed volume.

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

🏢BaiduBeijing · China84 works · 17,681 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

Cite

Cite as Le Zhang, Shengming Zhang, Rui Zha, Wu Yunpeng, Jingbo Zhou and Jizhou Huang. 2026. DuIVRS-2: An LLM-based Interactive Voice Response System for Large-scale POI Attribute Acquisition. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.44

@inproceedings{zhang2026duivrs,
  title = {{DuIVRS-2: An LLM-based Interactive Voice Response System for Large-scale POI Attribute Acquisition}},
  author = {Le Zhang, Shengming Zhang, Rui Zha, Wu Yunpeng, Jingbo Zhou and Jizhou Huang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.44},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.44}
}

References

31 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Paweł Budzianowski and Ivan Vulić. 2019. Hello, it’s gpt-2–how can i help you? towards the use of pre-trained language models for task-oriented dialogue systems. arXiv preprint arXiv:1907.05774. — [arXiv:1907.05774](https://arxiv.org/abs/1907.05774) find
  2. Valentina Carfora, Francesca Di Massimo, Rebecca Rastelli, Patrizia Catellani, and Marco Piastra. 2020. Dialogue management in conversational agents through psychology of persuasion and machine learning. Multimedia Tools and Applications, 79:35949– 35971. find
  3. Yudong Chen, Xin Wang, Miao Fan, Jizhou Huang, Shengwen Yang, and Wenwu Zhu. 2021. Curriculum meta-learning for next poi recommendation. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 2692– 2702. find
  4. Miao Fan, Jizhou Huang, and Haifeng Wang. 2022. Dumapper: Towards automatic verification of large-scale pois with street views at baidu maps. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 3063–3071. find
  5. Miao Fan, Yibo Sun, Jizhou Huang, Haifeng Wang, and Ying Li. 2021. Meta-learned spatial-temporal poi auto-completion for the search engine at baidu maps. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 2822–2830. find
  6. Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, and 1 others. 2023. Reinforced self-training (rest) for language modeling. arXiv preprint arXiv:2308.08998.
  7. Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A simple language model for task-oriented dialogue. Advances in Neural Information Processing Systems, 33:20179– 20191.
  8. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.
  9. Jizhou Huang, Haifeng Wang, Shiqiang Ding, and Shaolei Wang. 2022a. DuIVA: An Intelligent Voice Assistant for Hands-free and Eyes-free Voice Interaction with the Baidu Maps App. In Proceedings of Jizhou Huang, Haifeng Wang, Yibo Sun, Miao Fan, Zhengjie Huang, Chunyuan Yuan, and Yawen Li. 2021. HGAMN… find
  10. Jizhou Huang, Haifeng Wang, Yibo Sun, Yunsheng Shi, Zhengjie Huang, An Zhuo, and Shikun Feng. 2022b. ERNIE-GeoL: A Geography-and-Language Pre-trained Model and its Applications in Baidu Maps. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3029–3039. find
  11. Jizhou Huang, Haifeng Wang, and Shaolei Wang. 2022c. DuIVRS: A Telephonic Interactive Voice Response System for Large-Scale POI Attribute Acquisition at Baidu Maps. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 3182–3191. find
  12. 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
  13. Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. find
  14. Manqing Mao, Paishun Ting, Yijian Xiang, Mingyang Xu, Julia Chen, and Jianzhe Lin. 2024. Multi-user chat assistant (muca): a framework using llms to facilitate group conversations. arXiv preprint arXiv:2401.04883. — [arXiv:2401.04883](https://arxiv.org/abs/2401.04883) find
  15. Sara Montagna, Stefano Ferretti, Lorenz Cuno Klopfenstein, Antonio Florio, and Martino Francesco Pengo. 2023. Data decentralisation of llm-based chatbot systems in chronic disease self-management. In Proceedings of the 2023 ACM Conference on Information Technology for Social Good, pages 205–212. find
  16. Mahdi Namazifar, Alexandros Papangelis, Gokhan Tur, and Dilek Hakkani-Tür. 2021. Language model is all you need: Natural language understanding as question answering. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7803–7807. IEEE. find
  17. Jinjie Ni, Tom Young, Vlad Pandelea, Fuzhao Xue, and Erik Cambria. 2023. Recent advances in deep learning based dialogue systems: A systematic survey. Artificial intelligence review, 56(4):3055–3155. find
  18. OpenAI. 2023. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774.
  19. Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. 2020. Few-shot natural language generation for task-oriented dialog. arXiv preprint arXiv:2002.12328.
  20. Jérôme Revaud, Minhyeok Heo, Rafael S Rezende, Chanmi You, and Seong-Gyun Jeong. 2019. Did it change? learning to detect point-of-interest changes for proactive map updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4086–4095. find
  21. Yibo Sun, Jizhou Huang, Chunyuan Yuan, Miao Fan, Haifeng Wang, Ming Liu, and Bing Qin. 2021a. Gedit: geographic-enhanced and dependency-guided tagging for joint poi and accessibility extraction at baidu maps. In Proceedings of the 30th ACM international conference on information & knowledge manageme… find
  22. Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, and 1 others. 2021b. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137.
  23. Qwen Team. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671. find
  24. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
  25. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.
  26. Fuling Wang, Le Zhang, Jingbo Zhou, Jindong Han, Ying Sun, Chuan Qin, Hengshu Zhu, and Hui Xiong. 2026. Aradd: An automatic real-world api discovery and deployment framework for ai guide service in baidu map. In Proceedings of the ACM Web Conference 2026, pages 7845–7856. others. 2022. Sustainable a… find
  27. Canwen Xu, Jing Li, Xiangyang Luo, Jiaxin Pei, Chenliang Li, and Donghong Ji. 2019. Dlocrl: A deep learning pipeline for fine-grained location recognition and linking in tweets. In The World Wide Web Conference, pages 3391–3397. find
  28. Xue Yang, Jie Gao, Xiaoyun Zheng, Mengyuan Fang, Luliang Tang, and Xia Zhang. 2022. Point-of-interest detection from weibo data for map updating. Transactions in GIS, 26(6):2716–2738. find
  29. Zixuan Yuan, Hao Liu, Yanchi Liu, Denghui Zhang, Fei Yi, Nengjun Zhu, and Hui Xiong. 2020. Spatiotemporal dual graph attention network for query-poi matching. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, pages 629–638. find
  30. Le Zhang, Yixiong Xiao, Xinjiang Lu, Jingjia Cao, Yu-sai Zhao, Jingbo Zhou, Lang An, Zikan Feng, Wanxiang Sha, Yu Shi, and 1 others. 2026. Omegause: Building a general-purpose gui agent for autonomous task execution. arXiv preprint arXiv:2601.20380. — [arXiv:2601.20380](https://arxiv.org/abs/2601.20380) find
  31. Pengpeng Zhao, Anjing Luo, Yanchi Liu, Jiajie Xu, Zhixu Li, Fuzhen Zhuang, Victor S Sheng, and Xiaofang Zhou. 2020. Where to go next: A spatiotemporal gated network for next poi recommendation. IEEE Transactions on Knowledge and Data Engineering, 34(5):2512–2524. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-duivrs-2-llm-based-interactive/