🚏 LocalSearchBench

Benchmarking agentic local-life search with nine high-density Chinese cities.
Live demo, leaderboard, and playground: localsearchbench.github.io

📝 Abstract

Recent advances in large reasoning models (LRMs) have enabled agentic search systems to perform complex multi-step reasoning across multiple sources. However, most research focuses on general information retrieval and rarely explores vertical domains with unique challenges. In this work, we focus on local life services and introduce LocalSearchBench, which encompasses diverse and complex business scenarios. Real-world queries in this domain are often ambiguous and require multi-hop reasoning across merchants and products, remaining challenging and not fully addressed. As the first comprehensive benchmark for agentic search in local life services, LocalSearchBench includes over 1,354,185 high-quality merchant entries spanning six business scenarios and nine cities. We construct 900 multi-hop QA tasks (3–5 hops, L3–L4 difficulty) based on real user intents, challenging agents to decompose requirements, retrieve information, and synthesize itineraries. We also developed LocalPlayground, a unified evaluation interface integrating retrieval, reranking, and tool-use traces. Experiments show that even state-of-the-art LRMs struggle on LocalSearchBench: the best model (DeepSeek-V3.1) reaches only 34.34% correctness, with average completeness and faithfulness below 80% and 62%, respectively. This highlights the need for specialized benchmarks and domain-specific agent training in local life services.

⭐ Key Features

• Six core scenarios: Dining, entertainment, shopping, hotel, travel errands, and mixed lifestyle tasks reflect the platform’s dominant traffic mix.
• Nine major cities: Shanghai, Beijing, Guangzhou, Shenzhen, Hangzhou, Suzhou, Chengdu, Chongqing, and Wuhan ensure geographic and economic diversity.
• Multi-hop supervision: Each QA example bundles a full Multi-hop search path trace that records tool calls and intermediate evidence.
• Difficulty grading (L3–L4): First industry-aligned grading system for local-life agentic search, measuring requirement complexity and plan length.
• Merchant-scale grounding: 1.35M merchants with anonymization, augmentation, and quality filtering support faithful retrieval.

📊 Benchmark At a Glance

• Merchant database: 1,354,185 POIs collected via multi-agent crawling and QA.
• QA tasks: 900 samples → 100 tasks per city, 3–5 hops each.
• Models evaluated: GPT-4.1, Gemini-2.5-Pro, Qwen3 series, GLM-4.5, DeepSeek-V3.1, LongCat-Large, Hunyuan-T1, etc.
• Metrics: Correctness, completeness, fluency, faithfulness, safety, avg. tool calls, avg. rounds (matches leaderboard section on the website).

🧱 Data Fields

Field	Type	Description
Hop Count	int64	Number of chained search hops used to solve the query (3–5).
Difficulty	string	L3 (standard multi-hop) or L4 (complex constraints).
City	string	One of the nine supported Chinese cities.
Question	string	Natural-language request describing the user’s goal.
Multi-hop search path	string	Serialized trace of each hop’s query + retrieved evidence.
Answer	string	Final recommendation sequence with addresses, hours, and pricing.

🚀 Sample Usage

from datasets import load_dataset

ds = load_dataset("localsearchbench/localsearchbench")
print(ds["train"][0]["Question"])
print(ds["train"][0]["Multi-hop search path"])

🏅 Leaderboard Snapshot

Model	Avg tool calls	Avg rounds	Correctness	Completeness	Faithfulness
DeepSeek-V3.1	3.43	4.02	34.34	80.00	60.80
Qwen-Plus-Latest	2.59	3.12	32.79	80.94	68.68
LongCat-Large-32K	2.73	3.22	33.19	80.51	60.80
GPT-4.1	1.72	2.70	26.76	75.42	72.63
Gemini-2.5-Pro	1.89	2.86	26.09	77.93	78.26

Full interactive leaderboard (with fluency and safety columns) is available on the project page.

📎 Citation

@misc{he2025localsearchbenchbenchmarkingagenticsearch,
  title={LocalSearchBench: Benchmarking Agentic Search in Real-World Local Life Services},
  author={Hang He and Chuhuai Yue and Chengqi Dong and Mingxue Tian and Zhenfeng Liu and Jiajun Chai and Xiaohan Wang and Yufei Zhang and Qun Liao and Guojun Yin and Wei Lin and Chengcheng Wan and Haiying Sun and Ting Su},
  year={2025},
  eprint={2512.07436},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2512.07436}
}

✅ License

MIT License. Commercial and research use permitted with attribution to LocalSearchBench.