TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models

TRACE enables FP4 quantization for Reinforcement Learning in Mixture-of-Experts models by aligning training and rollout execution paths. The framework uses rollout-guided quantization-aware training to minimize the discrepancy between these two paths, reducing the high memory and computational overhead typically required for LLM rollout generation.