robustbench-tc/

RobustBench-TC

Anonymous reviewer-facing release for a NeurIPS 2026 D&B-track submission. All identifying information has been redacted; full author / institution attribution will be added with the camera-ready release.

RobustBench-TC is a sim-to-real robustness benchmark for tool-use language agents. It augments five public single-turn tool-calling benchmarks (BFCL V3, API-Bank, RoTBench, ToolAlpaca, ToolEyes) with 22 perturbation types organized along the four components of the… See the full description on the dataset page: https://huggingface.co/datasets/robustbench-tc/RobustBench-TC.