Large Language Models as Optimizers

Abstraction: OPRO uses LLMs as gradient-free optimizers via natural language prompts

Key points:

• OPRO (Optimization by PROmpting) proposes using LLMs as optimizers for tasks where gradients are unavailable
• At each step, the LLM generates new candidate solutions from a prompt containing prior solutions and their scores; best solutions accumulate iteratively
• Applied to linear regression and traveling salesman problems as proofs of concept, then to prompt optimization
• Best OPRO-optimized prompts outperform human-designed prompts by up to 8% on GSM8K and up to 50% on Big-Bench Hard tasks
• The method is model-agnostic: demonstrated with a variety of LLMs
• Code released at github.com/google-deepmind/opro

Connections: Google Deepmind · Prompt Engineering · Large Language Models · Optimization

Source: https://arxiv.org/abs/2309.03409