ATLAS-Bench: An Agentic Tool-Use Benchmark for Local Area Search

ATLAS-Bench introduces a benchmark of 1,200 queries across three regions to evaluate LLM agents on local area search tasks. It differentiates between mandatory requirements and subjective preferences using real-world place attributes and reviews. The accompanying ATLASBox environment provides standardized search tools and a fixed point-of-interest dataset for consistent evaluation.