作者位次 #3/8ACL 2026Findings· Findings of ACL
MTP - RL : Acceleration of Reinforcement Learning Rollouts with Policy-Aligned Multi-Token Prediction
Ke Wang · Aohan Zeng · Zhengxiao Du · Yuxuan Hu · Bohan Zhang · Xinyi Wang · Jie Tang · Jing Zhang
摘要
Reinforcement learning (RL) is widely applied to boost the performance of pretrained models, yet its training efficiency is severely constrained by rollout generation. While speculative decoding based on multi-token prediction (MTP) offers a potential acceleration pathway, its widespread adoption...