Elamur Research⌕
☰
home / papers / 2026 / Internalizing Multi-Agent Reasoning for Accurate and Efficient LLM-based Recommendation

Internalizing Multi-Agent Reasoning for Accurate and Efficient LLM-based Recommendation

Paper · 2026 · Yang Wu, Haoze Wang, Qian Li et al.

Summary

LLM-based recommenders struggle with two problems. Collaborative signals (behavioral consensus from many users) are invisible to models that read only item text. Tool-using agents that could fetch such signals are too slow because each reasoning turn adds latency. The authors address both with a two-stage approach. First, they build MARS, a multi-agent teacher that works in a plan–execute–reflect loop. It has a Planner, specialized analysis agents, a Reflector and a Ranking agent. Its Collaborative Signal Translation tool retrieves behavioral neighbours from the user-item graph and turns them into readable textual evidence. Second, MARS trajectories that pass an outcome-based filter are distilled into one Qwen3-8B student, STAR, using SFT followed by GRPO. The student learns to plan, call tools and self-reflect within a single generation. On AgentRecBench data (Amazon, Goodreads, Yelp), STAR beats its own teacher while running much faster. This suggests multi-agent reasoning can be compressed into a deployable single model.

Key points

  • STAR outperforms the MARS teacher by 8.7%–39.5% across scenarios. Its Avg HR (averaged over HR@1/3/5) is 81.9 on Amazon versus 73.0 for MARS, 78.4 versus 66.2 on Goodreads, and 21.5 versus 16.5 on Yelp.
  • On the same Qwen3-8B backbone, trajectory-based SFT lifts Amazon Avg HR from 64.5 (plain SFT) to 78.2, and adding GRPO brings it to 81.9. DeepSeek-V3 with chain-of-thought prompting reaches only 60.6.
  • Inference time falls from 461 s for the teacher to 89 s for STAR, though this is still far from sub-second serving.
  • Strict outcome filtering of teacher trajectories helps. Also accepting trajectories that only hit at rank 3 or 5 lowers results, for example Yelp drops from 20.1 to 17.9.
  • Cross-domain transfer is asymmetric. A model trained on either Amazon or Goodreads performs strongly when applied to the other, and Yelp as a target nearly matches its in-domain score. Evaluation used 20-item candidate sets, not full-corpus ranking.

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

🏢TencentShenzhen · China163 works · 21,032 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksInformation Systems · 1,730 worksDecision Sciences · 249 worksManagement Science and Operations Research · 134 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,685 works

Cite

Cite as Yang Wu, Haoze Wang, Qian Li, Jun Zhang, Huan Yu and Jie Jiang. 2026. Internalizing Multi-Agent Reasoning for Accurate and Efficient LLM-based Recommendation. 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.2134

@inproceedings{wu2026internalizing,
  title = {{Internalizing Multi-Agent Reasoning for Accurate and Efficient LLM-based Recommendation}},
  author = {Yang Wu, Haoze Wang, Qian Li, Jun Zhang, Huan Yu and Jie Jiang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2134},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2134}
}

References

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

  1. 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. Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. In Proceedings of the 17th ACM conference on recommender systems, pages 1007–1014.
  3. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, and 1 others. 2025. Deepseekr1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633–638. find
  4. Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 639–648. find
  5. Yingzhi He, Xiaohao Liu, An Zhang, Yunshan Ma, and Tat-Seng Chua. 2025. Llm2rec: Large language models are powerful embedding models for sequential recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 896–907. find
  6. Xu Huang, Jianxun Lian, Yuxuan Lei, Jing Yao, Defu Lian, and Xing Xie. 2025. Recommender ai agent: Integrating large language models for interactive recommendations. ACM Transactions on Information Systems, 43(4):1–33. find
  7. Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer, 42(8):30–37. find
  8. Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, and 1 others. 2025. How can recommender systems benefit from large language models: A survey. ACM Transactions on Information Systems, 43(2):1–47.
  9. Qidong Liu, Xian Wu, Yejing Wang, Zijian Zhang, Feng Tian, Yefeng Zheng, and Xiangyu Zhao. 2024. Llm-esr: Large language models enhancement for long-tailed sequential recommendation. Advances in Neural Information Processing Systems, 37:26701– 26727. find
  10. Yuting Liu, Jinghao Zhang, Yizhou Dang, Yuliang Liang, Qiang Liu, Guibing Guo, Jianzhe Zhao, and Xingwei Wang. 2025. Cora: Collaborative information perception by large language model’s weights for recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 122… find
  11. Bo Ma, Hang Li, ZeHua Hu, XiaoFan Gui, LuYao Liu, and Simon Lau. 2025. Agentrec: Next-generation llm-powered multi-agent collaborative recommendation with adaptive intelligence. arXiv preprint arXiv:2510.01609. — [arXiv:2510.01609](https://arxiv.org/abs/2510.01609) find
  12. Qiyao Peng, Hongtao Liu, Hua Huang, Qing Yang, and Minglai Shao. 2025. A survey on llm-powered agents for recommender systems. arXiv preprint arXiv:2502.10050. find
  13. Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation learning with large language models for recommendation. In Proceedings of the ACM web conference 2024, pages 3464–3475. find
  14. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  15. Yu Shang, Peijie Liu, Yuwei Yan, Zijing Wu, Leheng Sheng, Yuanqing Yu, Chumeng Jiang, An Zhang, Fengli Xu, Yu Wang, and 1 others. 2025. Agentrecbench: Benchmarking llm agent-based personalized recommender systems. arXiv preprint arXiv:2505.19623. — [arXiv:2505.19623](https://arxiv.org/abs/2505.19623) find
  16. An Zhang, Yuxin Chen, Leheng Sheng, Xiang Wang, and Tat-Seng Chua. 2024a. On generative agents in recommendation. In Proceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval, pages 1807– 1817. find
  17. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  18. Yang Zhang, Keqin Bao, Ming Yan, Wenjie Wang, Fuli Feng, and Xiangnan He. 2024b. Text-like encoding of collaborative information in large language models for recommendation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9181… find
  19. Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256.
  20. Jiakai Tang, Yujie Luo, Xunke Xi, Fei Sun, Xueyang Feng, Sunhao Dai, Chao Yi, Dian Chen, Zhujin Gao, Yang Li, and 1 others. 2025. Interactive recommendation agent with active user commands. arXiv preprint arXiv:2509.21317. — [arXiv:2509.21317](https://arxiv.org/abs/2509.21317) find
  21. Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Yanbin Lu, Xiaojiang Huang, and Yingzhen Yang. 2024a. Recmind: Large language model powered agent for recommendation. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4351–4364. find
  22. Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, and Min Zhang. 2024b. Macrec: A multi-agent collaboration framework for recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2760–2764. find
  23. Junda Wu, Cheng-Chun Chang, Tong Yu, Zhankui He, Jianing Wang, Yupeng Hou, and Julian McAuley. 2024. Coral: collaborative retrieval-augmented large language models improve long-tail recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3391–34… find
  24. Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He. 2025. Collm: Integrating collaborative embeddings into large language models for recommendation. IEEE Transactions on Knowledge and Data Engineering. find
  25. 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. 2024. Swift:a scalable lightweight infrastructure for fine-tuning. Preprint, arXiv:2408.05517. — [arXiv:2408.05517](https://arxiv.org/abs/2408.05517) find
  26. Xi Zhu, Yu Wang, Hang Gao, Wujiang Xu, Chen Wang, Zhiwei Liu, Kun Wang, Mingyu Jin, Linsey Pang, Qingsong Weng, and 1 others. 2025a. Recommender systems meet large language model agents: A survey. Foundations and Trends® in Privacy and Security, 7(4):247–396. find
  27. Yaochen Zhu, Chao Wan, Harald Steck, Dawen Liang, Yesu Feng, Nathan Kallus, and Jundong Li. 2025b. Collaborative retrieval for large language model-based conversational recommender systems. In Proceedings of the ACM on Web Conference 2025, pages 3323–3334. find
  28. Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collaborative large language model for recommender systems. In Proceedings of the ACM Web Conference 2024, pages 3162–3172. find
  29. Yu Xia, Sungchul Kim, Tong Yu, Ryan A Rossi, and Julian McAuely. 2025. Multi-agent collaborative filtering: Orchestrating users and items for agentic recommendations. arXiv preprint arXiv:2511.18413. — [arXiv:2511.18413](https://arxiv.org/abs/2511.18413) find
  30. Yuwei Yan, Yu Shang, Qingbin Zeng, Yu Li, Keyu Zhao, Zhiheng Zheng, Xuefei Ning, Tianji Wu, Shengen Yan, Yu Wang, and 1 others. 2025. Agentsociety challenge: Designing llm agents for user modeling and recommendation on web platforms. In Companion Proceedings of the ACM on Web Conference 2025, pages… find
  31. GRPO A • Easy: Success count is between 6 and 7. • Medium: Success count is between 3 and 5. • Hard: Success count is between 1 and 2. The RL training set is composed according to a 3:4:3 ratio (Easy:Medium:Hard), resulting in 500, 470, and 486 samples for the three datasets, respectively. A.2 Train… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-internalizing-multi-agent-reasoning-accurate/