Prompt Baking
6 months ago
- #Prompt Engineering
- #LLM
- #AI Safety
- Introduces 'Prompt Baking', a technique to convert prompts into weight updates for LLMs.
- Demonstrates improved performance on benchmarks like GSM8K, ASDiv, MBPP, and CommonsenseQA.
- Shows that baked models retain sensitivity to further prompting and baking, enabling iterative self-improvement ('Prompt Pursuit').
- Discusses implications for AI safety, continuous model updating, and stable AI personas.