Elamur Research⌕
☰
home / papers / 2026 / ProductResearch: Training E-Commerce Deep Research Agents via Multi-Agent Synthetic Trajectory Distillation

ProductResearch: Training E-Commerce Deep Research Agents via Multi-Agent Synthetic Trajectory Distillation

Paper · 2026 · Jiangyuan Wang, Kejun Xiao, Huaipeng Zhao et al.

Summary

General-purpose Deep Research agents are tuned mainly for web search and transfer poorly to online shopping. In that setting, complex pre-purchase questions require combining web evidence with structured catalog lookups and tying claims to verified product attributes. The authors (Alibaba International) built ProductResearch, a pipeline in which three LLM agents cooperate to produce long tool-use trajectories for training:

  • A User Agent mines anonymized behavioral histories (purchases, reviews, chats) to produce a persona, a hard research query and a weighted, query-specific rubric.
  • A Research Agent works through Plan→Toolcall→Report using web and product-catalog tools.
  • A Supervisor Agent, run as a three-state machine, checks every step and returns corrective feedback.

Overly short trajectories are discarded. The remaining supervisor exchanges are folded into single-role sequences ("reflective internalization") so they can be used for ordinary SFT. Fine-tuning the compact MoE model Qwen3-30B-A3B on this data substantially improves report quality and product coverage, bringing it close to proprietary deep research systems. The resulting collection of queries, rubrics and trajectories can be used for training models and also for evaluating them.

Key points

  • RACE score: the Qwen3-30B-A3B base model rose from 31.78 to 45.40 overall after SFT, with gains on all four dimensions (comprehensiveness, depth, instruction-following, readability).
  • Product coverage: effective product coverage grew from 3.58 to 12.45, more than three times the base model.
  • Human validation setup: 30 queries were scored for Gemini-DeepResearch, Qwen3-max and ProductResearch-SFT-128k, giving 90 reports. Nine expert annotators applied rubrics they wrote independently of the User Agent.
  • Human validation results: pairwise agreement between RACE and humans reached 94.1%, compared with 94.3% unanimous human agreement. The overall Pearson correlation was 0.931; for ProductResearch-SFT-128k, PAR was 0.856 and OPC was 0.957.
  • Limitations: the framework handles only single-turn queries, and the Web_Visit and Visit_Product tools need further improvement.

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

🏢Alibaba GroupHangzhou · China168 works · 15,386 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksInformation Systems · 1,437 works

Cite

Cite as Jiangyuan Wang, Kejun Xiao, Huaipeng Zhao, Tao Luo and Xiaoyi Zeng. 2026. ProductResearch: Training E-Commerce Deep Research Agents via Multi-Agent Synthetic Trajectory Distillation. 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.96

@inproceedings{wang2026productresearch,
  title = {{ProductResearch: Training E-Commerce Deep Research Agents via Multi-Agent Synthetic Trajectory Distillation}},
  author = {Jiangyuan Wang, Kejun Xiao, Huaipeng Zhao, Tao Luo and Xiaoyi Zeng},
  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.96},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.96}
}

References

18 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Gemini Deep Research — your personal research assistant. https://gemini.google/overview/deepresearch/. Qwen Deep Research. https://qwen.ai/blog?id=qwen-deepresearch. ↗
  2. Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. 2025. DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents. Preprint, arXiv:2506.11763.
  3. Jiabao Fang, Shen Gao, Pengjie Ren, Xiuying Chen, Suzan Verberne, and Zhaochun Ren. 2024. A Multi-Agent Conversational Recommender System. Preprint, arXiv:2402.01135. — [arXiv:2402.01135](https://arxiv.org/abs/2402.01135) find
  4. Ming Gong, Xucheng Huang, Chenghan Yang, Xianhan Peng, Haoxin Wang, Yang Liu, and Ling Jiang. 2025. MindFlow: Revolutionizing E-commerce Customer Support with Multimodal LLM Agents. Preprint, arXiv:2507.05330. — [arXiv:2507.05330](https://arxiv.org/abs/2507.05330) find
  5. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning. Preprint, arXiv:2503.09516.
  6. Explore and Evolve for Training Long-Horizon Web Agents. https://arxiv.org/abs/2509.06501v3. — [arXiv:2509.06501](https://arxiv.org/abs/2509.06501) ↗
  7. Yuanxing Liu, Weinan Zhang, Baohua Dong, Yan Fan, Hang Wang, Fan Feng, Yifan Chen, Ziyu Zhuang, Hengbin Cui, Yongbin Li, and Wanxiang Che. 2023. U-NEED: A Fine-grained Dataset for User Needs-Centric E-commerce Conversational Recommendation. In Proceedings of the 46th International ACM SIGIR Conferen… find
  8. Zile Qiao, Guoxin Chen, Xuanzhong Chen, Donglei Yu, Wenbiao Yin, Xinyu Wang, Zhen Zhang, Baixuan Li, Huifeng Yin, Kuan Li, Rui Min, Minpeng Liao, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. 2025. WebResearcher: Unleashing unbounded reasoning capability in Long-Horizon Agents. Preprint, arX… — [arXiv:2509.13309](https://arxiv.org/abs/2509.13309) find
  9. Rulin Shao, Akari Asai, Shannon Zejiang Shen, Hamish Ivison, Varsha Kishore, Jingming Zhuo, Xinran Zhao, Molly Park, Samuel G. Finlayson, David Sontag, Tyler Murray, Sewon Min, Pradeep Dasigi, Luca Soldaini, Faeze Brahman, Wen-tau Yih, Tongshuang Wu, Luke Zettlemoyer, Yoon Kim, and 2 others. 2025. D… — DR Tulu: Reinforcement Learning with Evolving Rubrics for Deep Research
  10. Zhengwei Tao, Jialong Wu, Wenbiao Yin, Junkai Zhang, Baixuan Li, Haiyang Shen, Kuan Li, Liwen Zhang, Xinyu Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. 2025. WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization. Preprint, arXiv:2507.15061. — [arXiv:2507.15061](https://arxiv.org/abs/2507.15061) find
  11. Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, Kuan Li, Liangcai Su, Litu Ou, Liwen Zhang, Pengjun Xie, Rui Ye, Wenbiao Yin, Xinmiao Yu, Xinyu Wang, and 38 others. 2025. Tongyi DeepResearch Technical Report. P… — [arXiv:2510.24701](https://arxiv.org/abs/2510.24701) find
  12. Jiangyuan Wang, Kejun Xiao, Qi Sun, Huaipeng Zhao, Tao Luo, Jiandong Zhang, and Xiaoyi Zeng. 2025. ShoppingBench: A Real-World Intent-Grounded Shopping Benchmark for LLM-based Agents. Preprint, arXiv:2508.04266. — [arXiv:2508.04266](https://arxiv.org/abs/2508.04266) find
  13. Zijian Li, Xin Guan, Bo Zhang, Shen Huang, Houquan Zhou, Shaopeng Lai, Ming Yan, Yong Jiang, Pengjun Xie, Fei Huang, Jun Zhang, and Jingren Zhou. 2025. WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research. https://arxiv.org/abs/2509.13312v3. — [arXiv:2509.13312](https://arxiv.org/abs/2509.13312) ↗
  14. Pei Wang, Yanan Wu, Xiaoshuai Song, Weixun Wang, Gengru Chen, Zhongwen Li, Kezhong Yan, Ken Deng, Qi Liu, Shuaibing Zhao, Shaopan Xiong, Xuepeng Liu, Xuefeng Chen, Wanxi Deng, Wenbo Su, and Bo Zheng. 2026. ShopSimulator: Evaluating and Exploring RL-Driven LLM Agent for Shopping Assistants. https://a…
  15. Junteng Liu, Yunji Li, Chi Zhang, Jingyang Li, Aili Chen, Ke Ji, Weiyu Cheng, Zijia Wu, Chengyu Du, Qidi Xu, Jiayuan Song, Zhengmao Zhu, Wenhu Chen, Pengyu Zhao, and Junxian He. 2025. find
  16. WebExplorer: Yu Xia, Sungchul Kim, Tong Yu, Ryan A. Rossi, and Julian McAuley. 2026. Multi-Agent Collaborative Filtering: Orchestrating Users and Items for Agentic Recommendations. Preprint, arXiv:2511.18413. — [arXiv:2511.18413](https://arxiv.org/abs/2511.18413) find
  17. Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents. https://arxiv.org/abs/2207.01206v4. ↗
  18. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. Preprint, arXiv:2210.03629.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-productresearch-training-e-commerce-deep/