Most LLM search still works by sampling more rollouts or extending one path at a time. This paper's bidirectional evolutionary search does it in a smarter way. It breaks the task backward into smaller
Self-Improving Language Models with Bidirectional Evolutionary Search
By
–
