What they did

The authors built PersonaForge around a four-dimensional persona space, behavioral controls calibrated to real-user statistics, and “Reverse Deep Construction,” which expands authentic seed queries into multi-turn interactions. They used the framework to generate a 6,300-record training dataset.

They also created PersonaForge-Bench, a manually annotated evaluation set containing 138 tasks across more than 20 professional domains. Agents are scored along four dimensions, including task completion and response quality, and the authors tested the training approach on Qwen3.5-27B.

Key findings

• In the analyzed 16,000 real-world sessions, 75.9% of interactions were multi-turn rather than complete single-turn requests.
• PersonaForge training improved Qwen3.5-27B’s composite benchmark score by 4.1%, with gains across all four evaluation dimensions.
• The largest reported improvements were 6.0% in task completion and 6.8% in response quality.
• Trained agents completed interactions with fewer conversational turns and tool calls; ablations attributed benefits to both the SOUL behavioral components and adaptive simulation.

Why it matters

Agent datasets commonly present complete instructions upfront, whereas real users often clarify goals, disclose constraints gradually, or change direction. PersonaForge offers a way to train and evaluate agents under this more interactive setting, potentially improving both task success and interaction efficiency.

Caveats

The benchmark contains only 138 tasks, and the reported training experiment centers on one 27-billion-parameter model, so generalization across model families and deployment settings remains unproven. The simulations are calibrated against real-session statistics and authentic queries, but synthetic personas may still omit important human behaviors; the abstract also does not provide absolute scores or numerical efficiency reductions.