arXiv · NLP· Yinghao Hu, Qian Xiao, Yuting Huang, Kaisong Song, Yiquan Wu, Ang Li, Junjie Cao, Jun Lin, Kun Kuang, Fei Wu·· 7 天前AI 评分26
JurisCoT-SFT 与 Lifecycle-Aware Backtrackable Policy Optimization 提升法律 LLM 推理
Towards Better Legal Reasoning LLMs: Signal Balancing and Reward Scheduling
AI 导读
针对法律领域 LLM 训练痛点,提出两阶段框架:Stage I 采用 JurisCoT-SFT 平衡推理轨迹与答案信号,Stage II 引入 Lifecycle-Aware Backtrackable Policy Optimization 动态管理奖励信号。
来源:arXiv · NLP · arxiv.org