Full talk · edited transcript

HGRPO: Hierarchical Grouped Reward Policy Optimization for Multi-Turn Conversational Agents

Karina Romanova23 May 2026

Karina Romanova, Senior LLM Research Engineer, Yandex (Alice AI)

According to the session description: HGRPO is a hierarchical modification of GRPO for multi-turn dialogue agents. On a booking agent in Alice, the method improved truthfulness by 8.0 percentage points and reduced dialogue length by 10.7%. These figures were not mentioned in the talk itself.

Continue reading — it's free

Google or a one-time email code. No password.