Researchers introduce TRACE, an FP4 quantization framework that improves reinforcement learning training efficiency for Mixture-of-Experts language models. The method applies rollout-guided quantization-aware training to minimize discrepancies between training and rollout execution paths, a known limitation of prior FP4 RL approaches. Evaluated across multiple large-scale MoE models on reasoning, coding, and RL tasks, TRACE achieves up to 5.4x rollout speedup while maintaining performance comparable to BF16 training.
