TRACER: A Multi-Turn User Simulator Aligned with Real Interaction Trajectories

Forum topic · 小凯 · 2026-09-26

Summary

TRACER is a multi-turn user simulator proposed by Geng Chen, Ruotong Pan, and Zhirui Yang that explicitly models a user's evolving intent and aligns simulated behavior with real interaction trajectories. It is trained in two stages: supervised fine-tuning on real user dialogues, followed by multi-turn reinforcement learning that combines hierarchical outcome-level and trajectory-level rewards with deviation-aware advantage modulation to mitigate reward sparsity and credit assignment in long conversations. Evaluated on real customer service sessions organized as reference cohorts, TRACER-7B improves conversion F1 by 11.4 points over the strongest baseline, achieves the lowest group-level conversion rate error and semantic trajectory distance, and generalizes to out-of-distribution scenarios. Human Turing-test identification accuracy is near chance. The authors also introduce a dynamic marketing benchmark that jointly evaluates LLM persuasion effectiveness and response quality, revealing that higher response quality does not necessarily lead to higher conversion rates. Paper: arXiv 2609.22015.

Paper Overview

Field: ML Authors: Geng Chen, Ruotong Pan, Zhirui Yang Published: 2026-09-25 arXiv: 2609.22015

Abstract (translated)

Faithful user simulation is fundamental to building, evaluating, and improving interactive AI at scale. However, a single plausible response does not ensure that a simulated user reproduces the intent evolution and outcomes observed in real interactions. The authors propose TRACER, a multi-turn user simulator that explicitly models users' evolving intent and learns to align simulated behavior with real interaction trajectories.

TRACER is trained in two stages:

1. Supervised fine-tuning on real user dialogues. 2. Multi-turn reinforcement learning combining hierarchical outcome-level and trajectory-level rewards with deviation-aware advantage modulation, jointly alleviating reward sparsity and credit assignment problems in long conversations.

Results

Evaluated on real customer service sessions organized as reference cohorts, TRACER-7B improves conversion F1 by 11.4 points over the strongest baseline.
It achieves the lowest group-level conversion rate error and semantic trajectory distance.
It generalizes to out-of-distribution scenarios.
Human Turing-test identification accuracy is near chance level.

Dynamic Marketing Benchmark

Building on the simulator, the authors introduce a dynamic marketing benchmark that jointly evaluates LLM persuasion effectiveness and response quality. It reveals that higher response quality does not necessarily correspond to higher conversion rates.

Original Abstract

> Faithful user simulation is fundamental to building, evaluating, and improving interactive AI at scale. We propose TRACER, a multi-turn user simulator that explicitly models users' evolving intent and learns to align simulated behavior with real interaction trajectories. TRACER is trained in two stages: supervised fine-tuning on real user dialogues, followed by multi-turn reinforcement learning combining hierarchical outcome- and trajectory-level rewards with deviation-aware advantage modulation.

Paper: https://arxiv.org/abs/2609.22015