arXiv · AI· Jingxuan Xu, Gang Wu, Yanan Wu, Yutao Mou, Songwei Yu, Tianzhuang He, Zhengshuo Gong, Zhao Liu, Zihang Xu, Wenqiang Zhu, Xinping Lei, Weihao Li, Ziqi Miao, Yuhui Bai, Zhongqiu Wang, Yan Wu, Ariel Deng·· 2 天前AI 评分26
FLARE:通过生成式奖励模型实现长程编码智能体全生命周期密集监督
FLARE: A Full-Lifecycle Dense Supervision Paradigm for Long-Horizon Coding Agents via Generative Reward Model
AI 导读
FLARE 提出基于轻量级生成式奖励模型(GRM)的密集监督范式,解决长程编码智能体中稀疏二元奖励导致的信用分配危机。FLARE (N=1) 比 Global Rollout (N=5) 性能更优且 Token 消耗减少 5 倍。在 SFT 中带来 19.13% 相对性能增益,在 RL 中带来 9.19% 性能提升。
来源:arXiv · AI · arxiv.org