About LOCA-bench
LOCA-bench evaluates language agents under extreme and controllable context growth scenarios. Given a task prompt, it leverages automated and scalable control of environment states to regulate the agent's context length from 8K to 256K tokens.
Tasks use mock MCP servers simulating real services: Google Calendar, Canvas LMS, Email, BigQuery, Google Sheets, Snowflake, and WooCommerce.
Citation
Use the following citation when referencing LOCA-bench in your research
@article{zeng2026loca,
title={LOCA-bench: Benchmarking Language Agents Under Controllable and Extreme Context Growth},
author={Zeng, Weihao and Huang, Yuzhen and He, Junxian},
journal={arXiv preprint arXiv:2602.07962},
year={2026}
}