Google DeepMind Introduces Mind Evolution: Enhancing Natural Language Planning with Evolutionary Search in Large Language Models

Google DeepMind's Mind Evolution introduces a novel evolutionary search strategy to enhance large language models' (LLMs) problem-solving capabilities in natural language planning. Unlike previous methods, Mind Evolution uses a genetic approach to iteratively generate, refine, and recombine candidate solutions in natural language, guided by a solution evaluator. This method avoids formalizing tasks, allowing for higher success rates in complex planning tasks. Mind Evolution integrates language-based genetic algorithms, using LLMs for operations like crossover and mutation, and employs a "Refinement through Critical Conversation" process for solution improvement. Experiments on benchmarks like TravelPlanner and StegPoet demonstrate Mind Evolution's superior performance, achieving over 95% success rates and outperforming baselines while controlling for inference costs. This approach showcases the effectiveness of evolutionary search in enhancing LLM performance without relying on formal solvers.