EVA-Bench - Faithfulness: leaderboard

Metric: Faithfulness (%; LLM-judge rating on a 1-3 scale mapped to 0-1 of whether actions and statements stay grounded in instructions, policy, tool results and user input; mean over trials; 213 scenarios in three enterprise domains, five clean-audio trials each, averaged equally over the airline, healthcare HR and IT service management domains). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.

Top models

#ModelScore
1gpt-realtime-1.536
2Ultravox-Realtime29.2
3Gemini 3.1 Flash Live23.8
4GPT-Realtime-mini12.5

Interactive version: theaggregate.ai/benchmark?slug=eva-bench-faithfulness · How It Works · Data refreshed daily, snapshot 2026-09-26.