Transfer Learning for LLMs
Transfer learning means using knowledge learned from one broad training process on new tasks.
LLMs transfer through:
- prompting
- in-context examples
- retrieval
- adapters
- fine-tuning
- distillation
Transfer options
| Method | Changes weights? | Best for |
|---|---|---|
| prompting | no | quick task adaptation |
| RAG | no | private/current facts |
| LoRA/adapters | small added weights | repeated behavior |
| full fine-tune | yes | deep specialization |
| distillation | trains smaller model | cost and latency reduction |
Decision rule
Start with the lightest method that passes evals.
Do not train a model when a better prompt, retrieval system, or tool call solves the problem.
Knowledge check
Q1: What is transfer learning?
Reusing broad learned capability for a new task.
Q2: Why is RAG often better than fine-tuning for facts?
Facts can change without retraining the model.