ServiceNow AI Research

🚀 Meet Apriel Voice 27B Preview: an open-weight voice AI built to get work done.
Beyond conversation: adapt to dynamic instructions, reason across modalities, and orchestrate multi-step tool calls to complete real tasks.
See Apriel Voice in action. 👇

🚀 @SpaceXSI 's Grok Voice 3.0 just set a new Pareto frontier on EVA-Bench, ServiceNow’s open benchmark for enterprise voice agents. No tradeoff between accuracy and experience; it balances both!
📊 Leaderboard: servicenow.github.io/eva/#results
#VoiceAgents #ServiceNowAIResearch #EVABench #GrokVoice #SpaceXSI

Milestone from our Document Intelligence & Multimodal Services team at ServiceNow: 3 papers, 4 workshop acceptances across COLM 2026 and EMNLP 2026. Proud to see work rooted in real production problems make it to scientific venues.

Thrilled to share 5 papers accepted at EMNLP 2026! 🥳

🤖 Agent augmentation ROI analysis: arxiv.org/pdf/2606.15017
🔄 Personalized workflow prediction: DRFLOW - arxiv.org/pdf/2606.18191
🔍 RAG advances: metadata-guided (MCompassRAG - arxiv.org/abs/2606.18508) & hierarchical retrieval (SproutRAG - arxiv.org/abs/2606.18381)
🎯 SVG generation benchmark: VectorGym - arxiv.org/abs/2603.29852

GPT-Realtime-2.1 vs 2.0, benchmarked on EVA-Bench:
📈 Accuracy: 50% → 59%
📈 Conversational Experience: 39% → 62%
🔐 User authentication: 84% → 95%

2.1 lands on the Pareto frontier 🚀
📊 Leaderboard: servicenow.github.io/eva/#results