Top AI Product

Bidirectional Evolutionary Search (BES) is a new framework for self-improving language models that attacks a quiet weakness in how models search for answers: expanding from a single starting point keeps producing similar candidates. BES pairs a forward search that breeds variety with a backward search that creates feedback.

## Forward evolution, backward decomposition

In the forward direction, BES goes beyond standard expansion by adding evolutionary operators that recombine partial solution trajectories — splicing pieces of different attempts to produce candidates a single model rollout would almost never reach. In the backward direction, it recursively breaks the original task into checkable subgoals, generating dense intermediate feedback that steers the forward search instead of only judging the final answer.

## Why it matters

The paper’s theoretical argument is the sharp part: expansion-only search stays trapped in a “narrow entropy shell” — candidates that all look alike — while evolutionary operators let the search escape it. And recursive backward decomposition can exponentially cut the number of samples needed to land on a correct answer. For agentic systems that improve by searching over their own attempts, that combination matters: more diverse candidates plus cheaper verification is exactly what makes self-improvement scale instead of plateauing.