Elamur Research⌕
☰
home / papers / 2026 / Understanding the Behaviors of Environment-aware Information Retrieval

Understanding the Behaviors of Environment-aware Information Retrieval

Paper · 2026 · Ruifeng Yuan, Chaohao Yuan, David Dai et al.

Summary

Agentic RAG systems often assume a single web search engine. In practice, they also query local backends such as BM25 or dense embedders, and each backend may respond best to a differently phrased query. This paper asks whether an LLM query rewriter can learn retriever-specific query styles through reinforcement learning. The rewriter is trained with GRPO, using nDCG@10 of the retrieved documents as the reward. Training uses a deduplicated retrieval corpus built from RAGBench, and transfer is tested zero-shot on BEIR.

RL does adapt the rewriter to each retriever, and the learned styles differ sharply:

  • Contriever prefers long, document-like rewrites.
  • Qwen3-Embedding favours question-like queries.
  • BM25 favours short keyword strings.

Because of this, a policy tuned for one retriever transfers poorly to another, which the authors call "Structural Drift". Performance improves with more detailed retriever-specific human guidance in the prompt, with larger models, and with a second retrieval turn. To keep two-turn training stable, the authors propose a branching rollout scheme. The work offers practical guidance for building retriever-aware RAG pipelines.

Key points

  • The reward is nDCG@10 against the gold passage. It was chosen over recall so that the rewriter learns to rank the correct document highly, not just retrieve it.
  • The data come from RAGBench (over 100,000 samples). The authors pooled all documents into one corpus and removed near-duplicates, defined as Qwen-Embedding-0.6B similarity above 0.99. Out-of-domain evaluation is zero-shot on BEIR.
  • Prompts were tested at three levels of human guidance, ranging from general rewriting advice to expert, retriever-specific instructions. More specific guidance gave better results. Larger models sometimes found strategies that the human experts had not foreseen.
  • The branching rollout samples 4 first-turn rewrites and 4 continuations for each (4×4), instead of 16 independent trajectories (16×1). The total reward is 0.5·R_turn1 + R_turn2.
    • Grouping continuations that share a first turn cancels first-turn noise in the second-turn advantage.
    • Averaging over the branches gives a Monte Carlo estimate of each first-turn rewrite's value.
    • Together, these prevent the training collapse seen with 16×1.
  • Limitations: only text retrievers were studied, and at most two retrieval steps were tested. Code is released at LCO-Embedding/Envs-aware-Information-Retrieval.

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

🎓Stanford UniversityStanford · United States506 works · 330,053 citations🏢Alibaba GroupHangzhou · China201 works · 17,173 citations🎓Chinese University of Hong KongHong Kong · China159 works · 16,154 citations🎓University of MacauMacao · China43 works · 2,368 citations🎓Shanghai University of Finance and EconomicsShanghai · China15 works · 825 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksInformation Systems · 1,730 works

🧭 Topics 1Research subjects and languages the work is about.Information retrieval & search · 552 works

Cite

Cite as Ruifeng Yuan, Chaohao Yuan, David Dai, Yu Rong, Hong Cheng, Hou Pong Chan and Chenghao Xiao. 2026. Understanding the Behaviors of Environment-aware Information Retrieval. 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.2013

@inproceedings{yuan2026understanding,
  title = {{Understanding the Behaviors of Environment-aware Information Retrieval}},
  author = {Ruifeng Yuan, Chaohao Yuan, David Dai, Yu Rong, Hong Cheng, Hou Pong Chan and Chenghao Xiao},
  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.2013},
  url = {https://doi.org/10.18653/v1/2026.acl-long.2013}
}

References

29 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Dan Hendrycks, Collin Burns, Anya Chen, and Spencer Ball. 2021. CUAD: An expert-annotated NLP dataset for legal contract review. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1).
  2. Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-rag: Learning to retrieve, generate, and critique through self-reflection.
  3. Robert Friel, Masha Belyi, and Atindriyo Sanyal. 2024. Ragbench: Explainable benchmark for retrieval-augmented generation systems. arXiv preprint arXiv:2407.11005. — [arXiv:2407.11005](https://arxiv.org/abs/2407.11005) find
  4. Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762–1777.
  5. Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning.
  6. Trans. Mach. Learn. Res., 2022. find
  7. 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. arXiv preprint arXiv:2503.09516.
  8. Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. 2019. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language proce…
  9. Ehsan Kamalloo, Aref Jafari, Xinyu Zhang, Nandan Thakur, and Jimmy Lin. 2023. Hagrid: A human-llm collaborative dataset for generative information-seeking with attribution. arXiv preprint arXiv:2307.16883. — [arXiv:2307.16883](https://arxiv.org/abs/2307.16883) find
  10. Zixuan Ke, Weize Kong, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. 2024. Bridging the preference gap between retrievers and llms. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10438–10451.
  11. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459–…
  12. Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting in retrieval-augmented large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5303–5315.
  13. Chaitanya Malaviya, Subin Lee, Sihao Chen, Elizabeth Sieber, Mark Yatskar, and Dan Roth. 2024. Expertqa: Expert-curated questions and attributed answers. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies…
  14. Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st annual meeting of the association for computational linguistics (v… find
  15. Timo Möller, Anthony Reina, Raghavan Jayakumar, and Malte Pietsch. 2020. Covid-qa: A question answering dataset for covid-19. In Proceedings of the 1st Workshop on NLP for COVID-19 at ACL 2020. find
  16. Abhilash Nandy, Soumya Sharma, Shubham Mad-dhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. 2021. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages… find
  17. Rodrigo Nogueira and Kyunghyun Cho. 2017. Task-oriented query reformulation with reinforcement learning. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 574–583. find
  18. Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Found. Trends Inf. Retr., 3(4):333–389. find
  19. Sentence-Transformers. 2025. all-minilm-l6-v2: Sentence embeddings model. https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2. Accessed: 2026-01-06. ↗
  20. Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. 2023. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9248–9274.
  21. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. CoRR, abs/2402.03300.
  22. Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663.
  23. Chenghao Xiao, Hou Pong Chan, Hao Zhang, Mahani Aljunied, Lidong Bing, Noura Al Moubayed, and Yu Rong. 2025a. Analyzing llms’ knowledge boundary cognition across languages through the lens of internal representations. In Proceedings of the 63rd Annual Meeting of the Association for Computational Lin… find
  24. Chenghao Xiao, Hou Pong Chan, Hao Zhang, Weiwen Xu, Mahani Aljunied, and Yu Rong. 2025b. Scaling language-centric omnimodal representation learning. arXiv preprint arXiv:2510.11693. — [arXiv:2510.11693](https://arxiv.org/abs/2510.11693) find
  25. Chenghao Xiao, Isaac Chung, Imene Kerboua, Jamie Stirling, Xin Zhang, Márton Kardos, Roman Solomatin, Noura Al Moubayed, Kenneth Enevoldsen, and Niklas Muennighoff. 2025c. Mieb: Massive image embedding benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22187… find
  26. Chenghao Xiao, Yizhi Li, G Hudson, Chenghua Lin, and Noura Al Moubayed. 2023. Length is a curse and a blessing for document-level semantics. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1385–1396. find
  27. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2369–…
  28. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, and 1 others. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176.
  29. Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. 2021. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yuan-2026-understanding-behaviors-environment-aware-information/