arXiv · Machine Learning· Chu-Cheng Lin, Eugene Ie·· 8 小时前AI 评分53
How Fast Should a Model Commit to Supervision? 通过 Tsallis Loss Continuum 训练推理模型
How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum
AI 导读
Chu-Cheng Lin 和 Eugene Ie 提出基于 Tsallis q-logarithm 的损失函数族 J_Q，在 RLVR（q=0）与潜在轨迹的 log-marginal-likelihood（q=1）之间建立连续训练路径。
来源：arXiv · Machine Learning · arxiv.org