TRACER-7B: multi-turn user simulator aligning simulated behavior with real interaction trajectories
The paper (arXiv:2609.28690v1) introduces TRACER, a multi-turn user simulator trained by supervised fine-tuning on real dialogues followed by multi-turn reinforcement learning that uses hierarchical outcome- and trajectory-level rewards and deviation-aware advantage modulation. TRACER-7B outperforms the strongest baseline by 11.4 conversion F1, achieves the lowest group-level conversion-rate error and semantic trajectory distance on real customer-service cohorts, generalizes to out-of-distribution scenarios, and passes human Turing tests near chance; the authors also present the Dynamic Marketing Benchmark, which jointly measures persuasion effectiveness and response quality and finds that higher response quality does not necessarily imply higher conversion rates.