Elamur Research⌕
☰
home / papers / 2026 / ReflectiveRAG: Rethinking Adaptivity in Retrieval-Augmented Generation

ReflectiveRAG: Rethinking Adaptivity in Retrieval-Augmented Generation

Paper · 2026 · Akshay Verma, Swapnil Gupta, Siddharth Pillai et al.

Summary

Retrieval-augmented generation tends to break down when the retrieved context is dominated by irrelevant or duplicated passages. Existing remedies, such as fixed top-k retrieval, cross-encoder reranking, or RL-trained retrieval policies like DeepRAG, either rely on rigid heuristics or require expensive training. The authors, from Amazon, propose ReflectiveRAG, a training-free pipeline built from two modules.

  • Self-Reflective Retrieval (SRR): a small language model checks whether the retrieved evidence is enough to answer the question. If not, it rewrites the query using verbalized sampling and stops once further gains become marginal. Retrieval uses hybrid BM25 plus dense scoring.
  • Noise Removal (NR): a contrastive, embedding-based filter that keeps distinctive, query-relevant chunks and discards overlapping ones.

The system was tested on WebQuestions, HotpotQA (distractor setting) and a proprietary InternalQA set, with 50M Common Crawl distractor documents added. It outperforms DeepRAG and other baselines in accuracy and evidence compactness at a small latency cost. The work matters because it shows that explicit control logic can rival learned policy optimization in noisy retrieval settings.

Key points

  • On WebQuestions, ReflectiveRAG beats DeepRAG by +2.7 pp EM and +2.5 pp F1. On InternalQA the gains are +4.9 pp EM and +4.1 F1.
  • Evidence redundancy falls by 30.88%, measured as mean pairwise cosine similarity among selected chunks. Added latency is only about 18 ms.
  • In the WebQuestions ablation, the full model reaches 63.1 EM and 77.7 F1. Removing SRR drops EM to 53.2, and removing NR drops it to 55.9 while raising redundancy from 0.47 to 0.61.
  • Running a fixed 3-step SRR instead of adaptive stopping trails the full model by about 4 pp EM and has the highest latency (1,090 ms vs 838 ms).
  • Neither the retriever nor the generator (e.g., GPT-4-turbo) is fine-tuned. The authors note that performance still depends on the retriever surfacing at least one relevant document.

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

 Fields, topics and concepts

📚 Shelves 5The 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 worksInformation Systems · 1,437 works

🧭 Topics 1Research subjects and languages the work is about.Information retrieval & search · 473 works

Cite

Cite as Akshay Verma, Swapnil Gupta, Siddharth Pillai, Prateek Sircar and Deepak Gupta. 2026. ReflectiveRAG: Rethinking Adaptivity in Retrieval-Augmented Generation. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-industry.27

@inproceedings{verma2026reflectiverag,
  title = {{ReflectiveRAG: Rethinking Adaptivity in Retrieval-Augmented Generation}},
  author = {Akshay Verma, Swapnil Gupta, Siddharth Pillai, Prateek Sircar and Deepak Gupta},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-industry.27},
  url = {https://doi.org/10.18653/v1/2026.eacl-industry.27}
}

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. Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on Freebase from question-answer pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1533–1544, Seattle, Washington, USA. Association for Computational Linguistics.
  2. Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yi-Ting Guo, and Jie Fu. 2024. Rq-rag: Learning to refine queries for retrieval augmented generation. ArXiv, abs/2404.00610. find
  3. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. ArXiv, abs/2005.11401.
  4. Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, and Ji rong Wen. 2021. Rocketqav2: A joint training method for dense passage retrieval and passage re-ranking. ArXiv, abs/2110.07367.
  5. Wonduk Seo, Hyunjin An, and Seunghyun Lee. 2025. A new query expansion approach via agent-mediated dialogic inquiry. find
  6. Ishneet Sukhvinder Singh, Ritvik Aggarwal, Ibrahim Allahverdiyev, Muhammad Taha, Aslihan Akalin, Kevin Zhu, and Sean O’Brien. 2024. Chunkrag: Novel llm-chunk filtering method for rag systems. ArXiv, abs/2410.19572. — [arXiv:2410.19572](https://arxiv.org/abs/2410.19572) find
  7. Sambeet Tiady, Anirban Majumder, and Deepak Gupta. 2023. Prodigy: Product design guidance at scale. In CIKM, pages 4836–4842. find
  8. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p…
  9. Xinyan Guan, Jiali Zeng, Fandong Meng, Chunlei Xin, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, and Jie Zhou. 2025. Deeprag: Thinking to retrieve step by step for large language models. find
  10. Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.
  11. Trans. Mach. Learn. Res., 2022. find
  12. Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane A. Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. J. Mach. Learn. Res., 24:251:1– 251:43. find
  13. Arihant Jain, Purav Aggarwal, and Anoop Saladi. 2025. AutoChunker: Structured text chunking and its evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), pages 983–995, Vienna, Austria. Association for Computational Linguis… find
  14. Vinay Kumar Verma, Shreyas Sunil Kulkarni, Happy Mittal, and Deepak Gupta. 2025. Moemoe: Question guided dense and scalable sparse mixture-of-expert for multi-source multi-modal answering. arXiv preprint arXiv:2503.06296. find
  15. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Conference on Empirical Methods in Natural Language Processing.
  16. Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R. Tomz, Christopher D. Manning, and Weiyan Shi. 2025. Verbalized sampling: How to mitigate mode collapse and unlock llm diversity. find
  17. A Anant Khandelwal, Happy Mittal, Shreyas Sunil Kulkarni, and Deepak Gupta. 2023. Large scale generative multimodal attribute extraction for e-commerce attributes. CoRR, abs/2306.00379. find
  18. Dongkyu Kim, Byoungwook Kim, Donggeon Han, and Matouvs Eibich. 2024. Autorag: Automated framework for optimization of retrieval augmented generation pipeline. ArXiv, abs/2410.20878. — [arXiv:2410.20878](https://arxiv.org/abs/2410.20878) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/verma-2026-reflectiverag-rethinking-adaptivity-retrieval-augmented/