I discovered that AI models can extract their own knowledge through
self-reflection and transfer it to other models via natural language
strategies. The student models actually surpass their teachers.
Results from scheduling task:
- Student baseline: 12%
- Teacher (Claude 3.5): 82%
- Student after transfer: 86.7%
The framework lets you transfer ANY capability (Python coding, math,
reasoning, etc.) in ~50 lines of code. Includes working examples and
complete documentation.
Technical approach: Teacher learns via Linguistic RL (reflection →
strategy extraction), then student learns via LoRA fine-tuning on
teacher-generated curriculum.
I discovered that AI models can extract their own knowledge through self-reflection and transfer it to other models via natural language strategies. The student models actually surpass their teachers.
Results from scheduling task: - Student baseline: 12% - Teacher (Claude 3.5): 82% - Student after transfer: 86.7%
The framework lets you transfer ANY capability (Python coding, math, reasoning, etc.) in ~50 lines of code. Includes working examples and complete documentation.
Technical approach: Teacher learns via Linguistic RL (reflection → strategy extraction), then student learns via LoRA fine-tuning on teacher-generated curriculum.