EVA-Bench 2.0 forces voice agents to handle 213 complex workflows
eva-bench 2.0 adds two domains, 121 tools and 213 scenarios to test voice agents across airlines, it and healthcare hr.
hugging face released eva-bench 2.0, expanding the voice agent benchmark from one to three enterprise domains: airline customer service, it service management, and healthcare hr. the dataset now covers 213 scenarios across 121 tools, a 4x increase in coverage, incorporating adversarial calls, authentication flows and unsatisfiable goals to better test agent robustness. validated against models like gpt-4 and claude, the open-source dataset aims to standardize evaluation for realistic, voice-first workflows.