The authors analyzed 16K real-world user sessions and found that 75.9% of agent interactions are multi-turn, yet most training data and assume single-turn, informationally complete queries. To close this gap, they built PersonaForge, a framework that synthesizes realistic multi-turn user-agent interactions using a four-dimensional persona space, SOUL-driven behavioral control calibrated to real-user statistics, and Reverse Deep Construction based on authentic seed queries.
Using PersonaForge, they generated a 6.3K-record training dataset and PersonaForge-Bench, a manually annotated 138-task benchmark across over 20 professional domains with four-dimensional scoring. Fine-tuning Qwen3.5-27B on PersonaForge improved the composite score by +4.1%, with the largest gains in Task Completion (+6.0%) and Response Quality (+6.8%). Trained agents also used fewer turns and tool calls, suggesting better interaction efficiency, and ablations confirmed the contribution of SOUL components and adaptive simulation.