Researchers propose Circuit-Anchored Evolution to safely evolve LLMs

Researchers Yan Liu, Jie Fu, and Tsung-Yi Ho have introduced Circuit-Anchored Evolution (CAE), a biologically inspired approach to safely evolve large language models (LLMs). The method draws parallels from developmental biology, where organisms preserve essential regulatory genes (like Hox genes) while allowing peripheral traits to adapt. Similarly, CAE identifies and anchors a tiny safety circuit—comprising less than 2% of model features—that causally mediates safety behaviors, constraining its evolution while allowing other components to optimize freely. This addresses a critical flaw in current self-evolution algorithms, which often optimize purely for capability at the expense of safety.

In experiments across three model families and two evolution algorithms, CAE demonstrated superior safety preservation with minimal capability loss. It significantly outperformed explicit reward-based constraints, both in effectiveness and efficiency. The paper suggests that just as developmental constraints prevent biological evolution from producing nonviable organisms, circuit anchoring could prevent AI models from evolving into powerful yet dangerous systems. The approach leverages mechanistic interpretability to pinpoint the safety circuit, offering a targeted, efficient method for safe AI evolution.