--- title: "Test-Time Strategies for More Efficient and Accurate Agentic RAG" description: "The paper looks at weaknesses of Search-R1, an agentic RAG framework trained with reinforcement learning to alternate… 5 references." authors: ["Abhinav Sharma", "Brian Zhang", "Deepti Guntur", "Zhiyang Zuo", "Shreyas Chaudhari", "Wenlong Zhao", "Franck Dernoncourt", "Puneet Mathur", "Ryan A. Rossi", "Nedim Lipka"] year: 2026 venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)" source: "https://aclanthology.org/2026.acl-srw.41/" doi: "10.18653/v1/2026.acl-srw.41" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/sharma-2026-test-time-strategies-more-efficient/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/sharma-2026-test-time-strategies-more-efficient/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/52/) / Test-Time Strategies for More Efficient and Accurate Agentic RAG # Test-Time Strategies for More Efficient and Accurate Agentic RAG **Authors:** Abhinav Sharma, Brian Zhang, Deepti Guntur, Zhiyang Zuo, Shreyas Chaudhari, Wenlong Zhao, [Franck Dernoncourt](https://md.research.elamur.ai/free/en/authors/franck-dernoncourt/), [Puneet Mathur](https://md.research.elamur.ai/free/en/authors/puneet-mathur/), [Ryan A. Rossi](https://md.research.elamur.ai/free/en/authors/ryan-a-rossi/) and [Nedim Lipka](https://md.research.elamur.ai/free/en/authors/nedim-lipka/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.acl-srw.41/) · [PDF](https://aclanthology.org/2026.acl-srw.41.pdf) · [DOI 10.18653/v1/2026.acl-srw.41](https://doi.org/10.18653/v1/2026.acl-srw.41) · [OpenAlex](https://openalex.org/W7166868200) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary The paper looks at weaknesses of Search-R1, an agentic RAG framework trained with reinforcement learning to alternate reasoning with retrieval. Analysing the reasoning traces of its Qwen2.5-7B model, the authors found two recurring problems. The model tends to "forget" earlier retrievals and repeat queries, and it struggles to pick out the useful facts from the passages it receives. Without retraining the model, they add three inference-time modules: - **Contextualization:** GPT-4.1-mini extracts question-relevant content into a persistent memory cache. - **De-duplication:** passages already seen are swapped for the next-ranked unseen ones from the E5 retriever. - **Hybrid:** the two modules combined. Evaluation used 500 questions from HotpotQA and Natural Questions, scored by exact match, an LLM-as-a-judge "LLM Match" metric, and average retrieval count. All three variants improved accuracy. Contextualization performed best, raising EM and cutting retrievals by roughly a tenth. The work suggests that cheap test-time changes can make agentic RAG both more accurate and lighter on the retriever. The caveat is that contextualization adds an extra LLM call per turn. ## Key points - **Setup:** the backbone is Qwen2.5-7B Search-R1-base (PPO). The 3B variants were dropped because they did not reliably follow the required output format. Inference ran on a single A100 (80GB), and the code is on GitHub (a1bhinav/test-time-agentic-rag). - **Contextualization results:** EM rose from 0.464 to 0.490 (+5.6% relative) and LLM Match from 0.538 to 0.574 (+6.7%). Average searches fell from 2.392 to 2.142 (−10.5%). - **De-duplication results:** EM 0.478 and LLM Match 0.560. It needed more searches (2.498) than the baseline. - **Hybrid results:** EM 0.480, LLM Match 0.568, and 2.154 searches. - **Limitations:** the efficiency measure counts retriever calls, not latency, tokens or cost. Only one backbone was tested. GPT-4.1-mini serves as both contextualizer and judge, and these choices were not ablated. The datasets may be affected by training-data contamination. *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 - [University of Massachusetts Amherst](https://md.research.elamur.ai/free/en/universities/university-of-massachusetts-amherst/) — University, Amherst Center, United States · 114 works · 68,731 citations 📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [🗄️ Information Systems](https://md.research.elamur.ai/free/en/shelves/computer-science/information-systems/) · 1,730 works 🧭 Topics (1): [🔍 Information retrieval & search](https://md.research.elamur.ai/free/en/topics/information-retrieval/) · 552 works ## Cite Cite as: Abhinav Sharma, Brian Zhang, Deepti Guntur, Zhiyang Zuo, Shreyas Chaudhari, Wenlong Zhao, Franck Dernoncourt, Puneet Mathur, Ryan A. Rossi and Nedim Lipka. 2026. Test-Time Strategies for More Efficient and Accurate Agentic RAG. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{sharma2026test, title = {{Test-Time Strategies for More Efficient and Accurate Agentic RAG}}, author = {Abhinav Sharma, Brian Zhang, Deepti Guntur, Zhiyang Zuo, Shreyas Chaudhari, Wenlong Zhao, Franck Dernoncourt, Puneet Mathur, Ryan A. Rossi and Nedim Lipka}, booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)}, year = {2026}, doi = {10.18653/v1/2026.acl-srw.41}, url = {https://doi.org/10.18653/v1/2026.acl-srw.41} } ``` - [← previous Test-Time Scaling in Multimodal Foundation Models: A Comprehensive Survey of Generation and Reasoning](https://md.research.elamur.ai/free/en/papers/wan-2026-test-time-scaling-multimodal-foundation/) - [next → Test-Time Training for Zero-Resource Dense Retrieval Reranking](https://md.research.elamur.ai/free/en/papers/liu-2026-test-time-training-zero-resource/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/52/) ## References 5 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Zhiyu An, Xianzhong Ding, Yen-Chun Fu, Cheng-Chung Chu, Yan Li, and Wan Du. 2024. Golden-retriever: High-fidelity agentic retrieval augmented generation for industrial knowledge base. Preprint, arXiv:2408.00798. — [arXiv:2408.00798](https://arxiv.org/abs/2408.00798) [find](https://scholar.google.com/scholar?q=Zhiyu+An%2C+Xianzhong+Ding%2C+Yen%2DChun+Fu%2C+Cheng%2DChung+Chu%2C+Yan+Li%2C+and+Wan+Du%2E+2024%2E+Golden%2Dretriever%3A+High%2Dfidelity+agentic+retrieval+augmented+generation+for+industrial+knowledge+base%2E+Preprint%2C+arXiv%3A) 2. Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. 2024. [Rq-rag: Learning to refine queries for retrieval augmented generation](https://md.research.elamur.ai/free/en/papers/chan-2024-rq-rag-learning-refine-queries/). Preprint, arXiv:2404.00610. 3. Jerry Huang, Siddarth Madala, Risham Sidhu, Cheng Niu, Julia Hockenmaier, and Tong Zhang. 2025. Rag-rl: Advancing retrieval-augmented generation via rl and curriculum learning. Preprint, arXiv:2503.12759. — [arXiv:2503.12759](https://arxiv.org/abs/2503.12759) [find](https://scholar.google.com/scholar?q=Jerry+Huang%2C+Siddarth+Madala%2C+Risham+Sidhu%2C+Cheng+Niu%2C+Julia+Hockenmaier%2C+and+Tong+Zhang%2E+2025%2E+Rag%2Drl%3A+Advancing+retrieval%2Daugmented+generation+via+rl+and+curriculum+learning%2E+Preprint%2C+arXiv%3A2503%2E12) 4. 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](https://md.research.elamur.ai/free/en/papers/jin-2025-search-r1-training-llms-reason/). Preprint, arXiv:2503.09516. 5. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. [Natural…](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2019-natural-questions-benchmark-question-answering/) *Updated 2 November 2026*