OmniSimulator
A framework for training LLM-based user simulators. Through a five-dimensional structured behavioral reasoning schema and teacher-model distillation, it enables small models to learn structured behavioral reasoning rather than shallow pattern matching, thereby improving the fidelity of user behavior simulation.
Timeline (last 90 days)
The OmniSimulator training framework enables small models at the 8B level to reach a level competitive with Claude-Sonnet-4.5 on user simulation tasks.
The core design of OmniSimulator is a structured behavioral reasoning framework that decomposes user decisions into five dimensions for organization.
OmniSimulator constructs training data based on the historical interaction trajectories of 200 real users over three months, distilling an average of about 50 high-quality examples per user, ultimately forming 9,881 training instances.
The OmniSimulator framework introduces a teacher model to generate high-quality reasoning trajectories under a given schema, and then uses these trajectories to train small models.
Qwen3-8B improves its average score by 34.33% after applying OmniSimulator.
The root cause of the user simulation problem lies not in data volume, but in how the reasoning is organized.
The data selection step of OmniSimulator deliberately favors real behavioral scenarios to avoid synthetic data diverging from the real decision distribution.
Whether the design of the structured schema can generalize to other behavioral domains, and how to stably control the quality of trajectories generated by the teacher model, are questions that remain to be verified.
All Facts (8)
The OmniSimulator training framework enables small models at the 8B level to reach a level competitive with Claude-Sonnet-4.5 on user simulation tasks.
50%UnverifiedThe core design of OmniSimulator is a structured behavioral reasoning framework that decomposes user decisions into five dimensions for organization.
50%UnverifiedOmniSimulator constructs training data based on the historical interaction trajectories of 200 real users over three months, distilling an average of about 50 high-quality examples per user, ultimately forming 9,881 training instances.
50%UnverifiedThe OmniSimulator framework introduces a teacher model to generate high-quality reasoning trajectories under a given schema, and then uses these trajectories to train small models.
50%UnverifiedQwen3-8B improves its average score by 34.33% after applying OmniSimulator.
50%UnverifiedThe root cause of the user simulation problem lies not in data volume, but in how the reasoning is organized.
50%UnverifiedThe data selection step of OmniSimulator deliberately favors real behavioral scenarios to avoid synthetic data diverging from the real decision distribution.
50%UnverifiedWhether the design of the structured schema can generalize to other behavioral domains, and how to stably control the quality of trajectories generated by the teacher model, are questions that remain to be verified.
50%