Self-Improving Language Models with Bidirectional Evolutionary Search

Researchers introduced Bidirectional Evolutionary Search (BES), a new framework for improving language-model self-search.

BES combines forward recombination of candidates with recursive backward subgoal decomposition to generate better options and denser feedback.

The paper argues BES has theoretical advantages over expansion-only search methods.

Experiments show gains on hard post-training tasks and open-ended problem-solving benchmarks, especially where standard search stalls.