Achieving safe self-evolution of large language models via circuit anchoring

The researchers propose the Circuit-Anchored Evolution (CAE) method, which in experiments across 3 model families and two evolutionary algorithms achieves better safety preservation with very small capability loss, outperforming explicit reward constraints in both effectiveness and efficiency. Inspired by biological developmental constraints and by Hox genes anchoring body structure over 500 million years, the authors use mechanistic interpretability to locate a safety circuit accounting for less than 2% [of something], which causally mediates the model's safety behavior; during evolution they constrain it within a small displacement bound while the remaining features evolve freely. The authors observe that existing large language model self-evolution algorithms optimize only for capability and assume safety will be preserved by default; their experiments show that models may thus mis-evolve into powerful and dangerous systems.