strategy-distillation

Introduces Strategy-Guided Policy Optimization (SGPO) for LLM reasoning, which replaces trajectory imitation with strategy distillation, improving generalization on math benchmarks.