A new arXiv paper proposes Circuit-Anchored Evolution, a method meant to keep self-evolving language models from gaining capability while losing safety behavior.
The authors compare unconstrained model evolution to biological mutation without developmental constraints. Their concern is that self-evolution algorithms may optimize for capability and assume safety will remain intact. In experiments, they say models can “misevolve” into more powerful but more dangerous versions.
Their proposed method uses mechanistic interpretability to identify a small safety circuit, described as less than 2 percent of model features, and constrains that circuit during evolution. The rest of the model can still change more freely, mirroring the idea that biological systems preserve core regulatory structures while adapting elsewhere.
The approach is research, not a proven safety standard. Its value is that it treats safety as something with internal mechanisms that must be preserved during optimization, rather than as a final behavior check after a model has already changed.