New TED framework enables AI models to learn from teachers without expensive retraining by storing reasoning patterns directly in prompts

TED (Training-Free Experience Distillation) transfers knowledge from teacher to student models without parameter updates or large-scale training data

Student models generate multiple reasoning trajectories while teachers independently solve problems and extract generalized reasoning patterns

Extracted experiences are continuously refined and updated over time, stored as in-context information in prompts rather than model weights

Approach designed for resource-constrained environments where traditional supervised or reinforcement-based knowledge distillation is impractical