Скачать или смотреть LocalSearchBench: Benchmark for Agentic Local Search

In this AI Research Roundup episode, Alex discusses the paper: 'LocalSearchBench: Benchmarking Agentic Search in Real-World Local Life Services(2512.07436v1)' This work introduces LocalSearchBench, the first comprehensive benchmark focused on agentic search in local life services such as city-specific merchants and products. The benchmark includes over 150,000 high-quality entries and 300 multi-hop QA tasks derived from real user queries, emphasizing ambiguity and multi-step reasoning. The authors also present LocalPlayground, a unified environment that integrates multiple tools for agent interaction and evaluation. Experimental results show that even leading large reasoning models perform poorly on this benchmark, underscoring the need for specialized benchmarks and domain-specific training for local service search. Paper URL: https://arxiv.org/pdf/2512.07436 #AI #MachineLearning #DeepLearning #AgenticSearch #Benchmarking #LocalSearch