EVA-Bench 2.0 is an open-source benchmark for evaluating enterprise voice agents, expanding from one domain to three: Airline Customer Service Management, Enterprise IT Service Management, and Healthcare HR Service Delivery. The release covers 213 evaluation scenarios across 121 tools — roughly 4x the original coverage. Scenarios are generated using SyGra, a graph-based synthetic data pipeline, and validated against three frontier models (GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6). Key design principles include voice-first scope, realism, scenario variety (single-intent, multi-intent, adversarial), authentication flows, and strict reproducibility via deterministic user goals and ground-truth database states. A multilingual extension is also previewed. All datasets are available on Hugging Face under the MIT license.
