提出HyTuning混合后训练框架,通过Progressive Reasoning Gain(PRG)度量推理步骤对答案支持的渐进增强性,并自适应加权Reasoning Distillation(RD)与Reinforcement Learning from Internal Feedback(RLIF),在高风…
机构:新国大
来源:arXiv 2604.08454 | AI4Papers 论文推荐平台