TRACE Paper: Rollout-Guided FP4 QAT Matches BF16 Rollout for MoE RL, 5.4x Faster

TRACE pairs rollout-side FP4 quantization decisions with the training-side rounding step for RL post-training of MoE language models, directly collapsing the train/rollout discrepancy that broke prior FP4 RL stacks. On four large MoE models across reasoning, coding, and long-horizon tasks, joint FP4 weight/activation and FP4 KV-cache rollout matches BF16 rollout performance while delivering up to 5.4x rollout speedup over post-hoc FP4 quantization of BF16 policies.