PersonaForge Finds 75.9% of 16K Real Agent Sessions Are Multi-Turn While Training Data Assumes Single-Turn

arXiv 2608.28378·medium signalXBlueskyLinkedInCopy link

arXiv 2608.28378 (2026-08-28, cs.CL) analyzes 16K real-world sessions and reports 75.9% of interactions are multi-turn, against training data and benchmarks that largely assume informationally complete single-turn queries. PersonaForge synthesizes realistic multi-turn user-agent interactions from a four-dimensional persona space with behavioral control calibrated to real-user statistics, yielding a 6.3K-record training set and PersonaForge-Bench, a manually annotated 138-task benchmark across more than 20 professional domains. Training Qwen3.5-27B on it improves the composite score by 4.1%, with Task Completion up 6.0% and Response Quality up 6.8%, and the trained agents use fewer turns and tool calls.