Self-Improving Language Models with Bidirectional Evolutionary Search

Researchers introduced Bidirectional Evolutionary Search (BES), a new method for improving language model search and generation.

BES combines evolutionary recombination of partial candidates with recursive backward decomposition into checkable subgoals.

The authors say this addresses sparse feedback and weak exploration in best-of-N and tree search.

They report gains on post-training tasks and inference benchmarks, suggesting a stronger framework for problem solving and self-improvement.