Karina Romanova | HGRPO: Hierarchical Grouped Reward Policy Optimization for Multi-Turn Conversation

At Data Fest 2026 in Belgrade, Karina Romanova, Senior LLM Research Engineer, presented HGRPO — a hierarchical modification of GRPO for multi-turn dialogue agents. Applied to a booking agent in Yandex Alice, the method improved truthfulness by 8.0 percentage points and reduced dialogue length by 10.7%. Наши соц.сети: Как попасть в чат сообщества ODS Mattermost: https://ods.ai/tracks/mattermost #DataFest2026 #AI #ML #LLM #GenAI #MachineLearning #DataScience #MLOps #AIAgents #RAG #ComputerVision #AutonomousDriving #Yandex #Яндекс #TechTalk #Developers #ArtificialIntelligence #ReinforcementLearning #MultimodalAI