TRACER
A multi-turn user simulator that explicitly models user intent evolution through two-stage training (supervised fine-tuning plus multi-turn reinforcement learning). It addresses the problem of surface-level style alignment without behavioral consistency, employing hierarchical rewards and bias-aware advantage modulation mechanisms.
Timeline (last 90 days)
TRACER is a multi-turn user simulator that explicitly models users' evolving intent and learns to align simulated behavior with real interaction trajectories.
TRACER employs a two-stage training process: the first stage performs supervised fine-tuning (SFT) on real user conversations, and the second stage introduces multi-turn reinforcement learning (RL).
TRACER designs hierarchical rewards that provide reward signals at both the outcome level and the trajectory level, combined with bias-aware advantage modulation to mitigate reward sparsity and credit assignment problems.
TRACER achieves the lowest values on both group-level conversion rate error and semantic trajectory distance metrics, and generalizes to out-of-distribution scenarios.
In a human Turing test, subjects' accuracy in identifying conversations generated by TRACER was close to random-guessing level.
Based on TRACER, the authors propose the Dynamic Marketing Benchmark, which jointly evaluates two dimensions of large language models: persuasion effectiveness and response quality.
All Facts (6)
TRACER is a multi-turn user simulator that explicitly models users' evolving intent and learns to align simulated behavior with real interaction trajectories.
50%UnverifiedTRACER employs a two-stage training process: the first stage performs supervised fine-tuning (SFT) on real user conversations, and the second stage introduces multi-turn reinforcement learning (RL).
50%UnverifiedTRACER designs hierarchical rewards that provide reward signals at both the outcome level and the trajectory level, combined with bias-aware advantage modulation to mitigate reward sparsity and credit assignment problems.
50%UnverifiedTRACER achieves the lowest values on both group-level conversion rate error and semantic trajectory distance metrics, and generalizes to out-of-distribution scenarios.
50%UnverifiedIn a human Turing test, subjects' accuracy in identifying conversations generated by TRACER was close to random-guessing level.
50%UnverifiedBased on TRACER, the authors propose the Dynamic Marketing Benchmark, which jointly evaluates two dimensions of large language models: persuasion effectiveness and response quality.
50%