| title | Joint Training of Multi-Token Prediction in Reinforcement Learning via Optimal Coefficient Calibration |
|---|---|
| authors | Zili Wang, Jiajun Chai, Lin Chen, Xiaohan Wang, Shiming Xiang, Guojun Yin |
| date | 2026-05-27 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.28184 |
| upvotes | 4 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as the standard paradigm for improving reasoning capability of large language models, while Multi-Token Prediction (MTP) has been a widely adopted module in pretraining. Combining them is a natural approach, yet current RL practices detach MTP gradients because joint training degrades the performance. We revisit this failure from an optimization perspective. We show that the per-step effect of MTP on the RL objective can be decomposed into two terms: a first-order correlation and a second-order perturbation penalty. This decomposition unifies three MTP training regimes: Detach, Cross-Entropy loss, and Policy loss, and explains why each succeeds or fails. Further analysis of policy loss reveals that, although it aligns with intuition, performance still degrades: the correlation term decays while the quadratic penalty persists. Guided by the analysis, we propose Optimal Coefficient Calibration (OCC), an adaptive scheme that tracks the optimal coefficient online via a log-probability proxy at negligible cost. Across six competition-level mathematical reasoning benchmarks, OCC consistently matches or exceeds the detach baseline, delivering improved joint MTP-RL training performance. |
cite
@misc{wang2026joint,
title={Joint Training of Multi-Token Prediction in Reinforcement Learning via Optimal Coefficient Calibration},
author={Zili Wang and Jiajun Chai and Lin Chen and Xiaohan Wang and Shiming Xiang and Guojun Yin},
year={2026},
eprint={2605.28184},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.28184}
}