1
Self-Improvement Can Self-Regress: The Rise-and-Collapse Failure Mode of LLM Self-Training
揭示LLM自我训练中"先升后崩"的失败模式,为模型优化后的退化问题提供关键机理分析
arXiv:2606.21090v1 Announce Type: new Abstract: Self-improvement can self-regress. In REINFORCE post-training for code, a model can quickly improve on…