arXiv · Artificial Intelligence· Hanyu Wang, Nakul Agarwal, Hossein Nourkhiz Mahjoub, Ehsan Moradi Pari, Makoto Fukushima, Jinghui Chen, Vaishnav Tadiparthi·· 4 hr agoAI score33
推理 RL 轨迹展开中参考引导与自由生成的平衡
Balancing Reference Guidance and Free Generation in Trajectory Rollouts for Reasoning RL
AI brief
论文提出 Adaptive Reference Guidance（ARG），在固定生成预算内构造正确轨迹，并应用于 Group Relative Policy Optimization（GRPO）的全失败组。
Source: arXiv · Artificial Intelligence · arxiv.org