各类基于决策单调性的dp优化
Training-Inference Mismatch In LLM KD
群作用
Burnside 引理
Distilling the Essence: Efficient Reasoning Distillation via Sequence Truncation
Instruction tuning with loss over instructions
Selective Reflection-Tuning: Student-Selected Data Recycling for LLM Instruction-Tuning
RETAINING BY DOING: THE ROLE OF ON-POLICY DATA IN MITIGATING FORGETTING
Proximal Gradient and Subgradients
Different Designs For LLM KD Loss(II)