simulation-based-training

TRACE presents a simulation-based method to train reasoning agents for causal exploration using synthesized rewards, improving reinforcement learning in diagnostic tasks. Experiments show it outperforms prompted baselines in a digital-advertising diagnostic environment.