Newest papers

Towards User Simulators That Stay On Task And Still Sound Natural

Mustafa Omer Gul, Philippe Laban, Tanya Goyal, Jennifer Neville · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.31512 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los usuarios dependen cada vez más de sistemas de IA para tareas ampliadas y multi-turn, sin embargo, la evidencia indica que el rendimiento del sistema basado en LLM se degrada sobre la interacción prolongada. Para evaluar y mejorar los asistentes de IA en tales configuraciones, necesitamos simuladores de usuarios que sigan fielmente las intenciones complejas del usuario mientras exhiben un comportamiento de conversación realista. Los enfoques existentes manifiestan un compromiso: los asistentes promptados pueden seguir las intenciones complejas del usuario pero carecen de fidelidad comportamental, mientras que los simuladores de usuarios especializados aprenden mejor el comportamiento de conversación pero luchan para mantenerse fieles a las intenciones complejas. En este artículo, proponemos una receta de aprendizaje de reforzo (RL) que recompensa conjuntamente la adhesión a la

Machine translation (Meta M2M100); the English original follows.

Abstract

Users increasingly rely on AI systems for extended, multi-turn tasks, yet evidence indicates that LLM-based system performance degrades over prolonged interaction. To evaluate and improve AI assistants in such settings, we need user simulators that faithfully follow complex user intents while exhibiting realistic conversational behavior. Existing approaches manifest a trade-off: prompted assistants can follow complex user intents but lack behavioral fidelity, while specialized user simulators better approximate conversational behavior but struggle to remain faithful to complex intents. In this paper, we propose a reinforcement learning (RL) recipe that jointly rewards task intent adherence and behavioral naturalness. A crucial contribution of the method is operationalizing naturalness with a lightweight discriminator trained in an alternating, adversarial loop with the user simulator. We release the resulting model, UserLM-2.0, which produces valid simulated conversations (i.e., meeting both task-faithfulness and naturalness criteria) 46.2% of the time, over three times the rate of the strongest prior user simulator (14.1%).

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Naturalness, Computer science, Human–computer interaction, Task (project management), Operationalization, User modeling

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange