SAGE: Benchmarking and Improving Retrieval for Deep Research Agents (arXiv 2602.05975)

Forum topic · 小凯 · 2026-07-05

Summary

SAGE is a benchmark and improvement framework for evaluating retrieval in deep research agents, presented in a February 2026 arXiv paper (2602.05975) by Tiansheng Hu, Yilun Zhao, Canyu Zhang, Arman Cohan, and Chen Zhao. The work addresses a key gap in agentic search: as LLM-based deep research systems increasingly perform iterative, multi-hop retrieval over external knowledge sources, existing static retrieval benchmarks fail to capture the planning, query decomposition, and evidence-synthesis behaviors these agents exhibit. SAGE provides a reproducible evaluation protocol for comparing retrievers, rerankers, and agentic retrieval policies within end-to-end research pipelines, covering aspects such as retrieval depth, iterative query formulation, citation quality, and task success. Beyond benchmarking, the paper proposes methods to improve retrieval for research agents and situates the work relative to broader trends in retrieval-augmented generation (RAG), LLM tool use, and multi-agent collaboration. For engineers and researchers, SAGE offers a standardized way to measure how retrieval quality translates into downstream research-task performance, and highlights open problems including evaluation reliability, latency and cost constraints, hallucination control, and safety when agents search the open web. Readers should consult the original PDF for exact experimental numbers.

SAGE: Benchmarking and Improving Retrieval for Deep Research Agents

Source: arXiv:2602.05975 · February 2026 Authors: Tiansheng Hu, Yilun Zhao, Canyu Zhang, Arman Cohan, Chen Zhao

Overview

SAGE targets a core weakness in modern deep research agents: retrieval. While LLM-based research agents can plan, decompose questions, and synthesize reports, their final output quality is bounded by how well they retrieve and verify evidence. SAGE provides both a benchmark for measuring retrieval quality in agentic research settings and methods for improving it.

Key points

Problem: Static retrieval metrics (e.g., nDCG on fixed query–document pairs) do not reflect how deep research agents actually retrieve — through iterative queries, sub-question decomposition, and multi-hop evidence chains.
Benchmark: SAGE offers a reproducible evaluation protocol covering retrievers, rerankers, and agentic retrieval policies inside end-to-end research pipelines.
Evaluation dimensions: retrieval depth, iterative query formulation, citation/evidence quality, and downstream task success — moving beyond single-shot ranking scores.
Improvements: The paper proposes techniques to strengthen retrieval for research agents, treating retrieval strategy as a learnable, plannable component rather than a fixed pipeline stage.

Context in the field

SAGE sits at the intersection of agentic search and large-scale retrieval systems. Classical search stacks treat recall, reranking, and generation as separate stages; the LLM era introduces two new variables — reasoning budget and action space (whether to retrieve, how many times, and which tools to call). Related work includes surveys of deep research systems, LLM-based scientific agents, and reasoning-aware retrieval frameworks such as AgentIR.

Practical takeaways

1. Architecture: Cascaded retrieve–rerank–generate remains mainstream, but agentic paradigms make retrieval count and policy themselves learnable. 2. Data: High-quality instruction data and synthetic queries matter, but beware of knowledge leakage and distribution shift. 3. Evaluation: Offline retrieval metrics diverge from end-task satisfaction; process-level metrics (citation accuracy, multi-hop chain completeness) are increasingly important. 4. Deployment: Latency, token cost, interpretability, and safety (e.g., poisoning via open-web retrieval) are hard constraints for real systems.

Limitations and open problems

The paper points to open questions including evaluation trustworthiness, latency/cost trade-offs, hallucination and safety risks of agents searching the open web, and cross-lingual/multimodal generalization. Quantitative results should be verified against the original PDF before citation.

References

Original paper: SAGE: Benchmarking and Improving Retrieval for Deep Research Agents
Related: surveys of deep research systems (arXiv 2506.12594), LLM-based search agents (arXiv 2508.05668), and AgentIR (arXiv 2603.04384).