Agentic Search in the Wild: Intents and Trajectory Dynamics from 14M+ Real Search Requests

Forum topic · 小凯 · 2026-07-05

Summary

This arXiv paper (2601.17617) presents a large-scale log analysis of LLM-powered agentic search based on 14.44M search requests across 3.97M sessions collected from DeepResearchGym, an open-source search API accessed by external agentic clients. The authors sessionize logs, assign session-level intents and step-wise query-reformulation labels via LLM-based annotation, and propose Context-driven Term Adoption Rate (CTAR) to measure whether new query terms are lexically traceable to previously retrieved evidence. Key findings: over 90% of multi-turn sessions contain at most ten steps and 89% of inter-step intervals fall under one minute; behavior differs by intent, with fact-seeking sessions showing increasing repetition while reasoning-driven sessions sustain broader exploration; on average 54% of newly introduced query terms appear in the accumulated evidence context. The results suggest practical signals for repetition-aware stopping, intent-adaptive retrieval budgeting, and cross-step context tracking. Anonymized logs are released on HuggingFace.

Overview

Agentic Search in the Wild: Intents and Trajectory Dynamics from 14M+ Real Search Requests is an academic paper (arXiv: 2601.17617, dated 2026-01-24) by Jingjie Ning, João Coelho, Yibo Kong, Yunfan Long, Bruno Martins, João Magalhães, et al. (8 authors in total).

Background

LLM-powered search agents are increasingly used for multi-step information seeking, yet the IR community lacks empirical understanding of how agentic search sessions actually unfold and how retrieved evidence shapes later queries. This paper addresses that gap with a large-scale log analysis of real-world agentic search traffic.

Key points

Data: 14.44M search requests forming 3.97M sessions, collected from DeepResearchGym, an open-source search API used by external agentic clients.
Method: Logs are sessionized; LLM-based annotation assigns session-level intents and step-wise query-reformulation labels. The paper introduces Context-driven Term Adoption Rate (CTAR), a metric quantifying whether newly introduced query terms are lexically traceable to previously retrieved evidence.
Session dynamics: Over 90% of multi-turn sessions contain at most ten steps, and 89% of inter-step intervals are under one minute.
Intent-dependent behavior: Fact-seeking sessions exhibit high repetition that increases over time, while sessions requiring reasoning sustain broader exploration.
Evidence traceability: On average, 54% of newly introduced query terms appear in the accumulated evidence context, with additional traceability to earlier steps beyond the most recent retrieval.

Implications

The findings provide candidate signals for:

Repetition-aware stopping rules in agentic pipelines
Intent-adaptive retrieval budgeting
Explicit cross-step context tracking

Resources

The anonymized logs are publicly released on HuggingFace: cx-cmu/deepresearchgym-agentic-search-logs.

Related entries

Agentic Information Retrieval (arXiv:2410.09713)
AceSearcher: Bootstrapping Reasoning and Search for LLMs via RL (arXiv:2509.24193)
Synergizing RAG and Reasoning: A Systematic Review (arXiv:2504.15909)