Source-linked AI summary
Test-Time Strategies for More Efficient and Accurate Agentic RAG
Brian Zhang, Deepti Guntur, Zhiyang Zuo, Abhinav Sharma, Shreyas Chaudhari, Wenlong Zhao, Franck Dernoncourt, Puneet Mathur, Ryan Rossi, Nedim Lipka
TL;DR
Complex multihop questions expose inefficiencies in iterative agentic RAG, including redundant retrieval and poor use of retrieved context. This paper modifies Search-R1 at test time with contextualization, de-duplication, and a hybrid combination. The contextualization variant delivers the strongest reported combination of answer accuracy and retrieval efficiency, while the de-duplication design can increase retrievals.
Problem
Search-R1 can repeatedly retrieve previously processed information and struggle to contextualize retrieved passages, increasing retrieval turns and hindering reasoning accuracy.
Method
The paper adds contextualization, de-duplication, and hybrid modules to Search-R1 at inference time, evaluating them with accuracy and retrieval-step metrics.
Results
All approaches improve Search-R1 answer accuracy; contextualization achieves the highest EM and LLM match while reducing retrievals, whereas de-duplication increases them.
Takeaways & Limitations
Contextualization provides the strongest reported balance of answer accuracy and retrieval efficiency, while de-duplication alone can make retrieval less efficient.
Takeaways & Limitations
The de-duplication expectation is conditional on information forgetting being the cause of repeated retrievals, and the approach may remove information needed for correct answers.
Abstract
from arXiv · showhide
Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.
1. Introduction
RAG systems struggle with complex, multihop questions, motivating iterative agentic frameworks such as Search-R1. This paper investigates test-time contextualization, de-duplication, and their hybrid to improve reasoning efficiency and answer accuracy.
- Traditional single-step RAG often struggles with complex or nuanced questions requiring deep contextual understanding and multihop retrieval.
- Search-R1 interleaves reasoning and retrieval, repeatedly querying a dense retriever and incorporating returned passages into the reasoning trace.
- Search-R1 inference can repeatedly retrieve processed information, increasing retrieval turns, token consumption, and latency while hindering contextualization and answer accuracy.
- Research questions: The paper asks whether concise relevant-information representations and prevention of redundant retrieval can improve efficiency and answer accuracy.
- The proposed test-time modifications process retrieved results through contextualization, de-duplication, or a hybrid combination of both.
2. Related Work
Related work addresses retrieval memory, document contextualization, hallucination, and reinforcement learning for RAG. The paper builds on these directions by contextualizing and retaining relevant information within Search-R1’s reasoning process.
- Memory Knowledge Reservoir stores previously retrieved titledocument pairs, enabling targeted queries and reportedly reducing response time by 46% while preserving baseline accuracy.
- Search-o1 extracts information from documents before integrating it into a reasoning chain, whereas this paper retains previously contextualized information alongside newly retrieved documents.
- RAG-RL trains a reasoning language model for RAG with reinforcement and curriculum learning, and benchmarks it on HotpotQA and MuSiQue.
3. Approach
The approach targets Search-R1’s information forgetting and ineffective information extraction through contextualization, de-duplication, and a hybrid pipeline. These test-time modules retain useful context, replace redundant documents, and evaluate whether the two strategies complement each other.
- Approach: Qualitative analysis identifies information forgetting and ineffective information extraction as two primary Search-R1 limitations.
- Approach: The evaluation measures Exact Match, LLM Match score, and average retrieval steps for each test-time modification.
- Contextualization: Contextualization uses an external language model to extract relevant information from retrieved documents and maintain a persistent memory cache across retrieval steps.
- Contextualization: The cache preserves previously stored information while adding relevant content, enabling reasoning over newly retrieved documents and retained context.
- Contextualization: The contextualization strategy is intended to provide concise retrieved representations and cached context without modifying the underlying model.
- De-duplication: De-duplication filters previously retrieved documents and replaces duplicates with the next-highest-ranked unseen passages.
- De-duplication: The de-duplication design tests whether retrieval overlap causes repeated queries and whether broader document exploration improves accuracy and reduces redundant searches.
- Hybrid: The hybrid approach combines contextualization with de-duplication to test whether retained extracted information and retrieval diversity jointly improve reasoning performance.
4. Experiments
Experiments evaluate Search-R1 enhancements on validation subsets using EM, LLM Match, and retrieval counts. Contextualization provides the strongest accuracy and efficiency gains, while de-duplication can increase retrievals despite adding context.
- Experimental setup: The evaluation samples 500 question-answer pairs from the HotpotQA and Natural Questions validation sets without using them for training or tuning.Retrieval uses the 2018 Wikipedia dump with the E5 retriever.
- Evaluation metrics: LLM Match supplements Exact Match by judging semantic equivalence between predicted and ground-truth answers, allowing minor phrasing differences.The metric assigns 1 for semantic equivalence and 0 for incomplete or divergent meaning.
- Results: 5.6% higher EM and 6.7% higher LLM Match make Contextualization the most accurate approach versus the Search-R1 baseline.It also records 2.142 average searches versus 2.392 for the baseline, the lowest retrieval count among the reported approaches.
- Results: De-duplication increases retrievals because returning only new documents encourages repeated queries for additional context.The additional context is rarely helpful because the needed information is often already present in the initial retrieval but not extracted by the model.
- Results: Exact Match declines as the number of retrievals increases for both Search-R1 and Contextualization, while their differences are not statistically significant at any search count.Figure 2 reports overlapping 95% confidence intervals around the compared scores.
- Results: LLM Match exceeds Exact Match by 16 to 18% across the Search-R1 baselines and three nontraining approaches.The observed discrepancies commonly involve numerical answers and shortened or abbreviated names.
5. Conclusion
The paper evaluates contextualization, de-duplication, and hybrid test-time enhancements for Search-R1. All improve answer accuracy, while contextualization alone most effectively reduces turns and the hybrid approach provides smaller gains in both accuracy and efficiency.
- The study evaluates Contextualization, Deduplication, and Hybrid test-time enhancements to the Search-R1 pipeline.
- All three approaches improve Search-R1 answer accuracy, but Contextualization reduces turns whereas De-Duplication increases them.
- The Hybrid approach improves both accuracy and retrieval efficiency, although less strongly than Contextualization alone.