Large language models acquire broad knowledge from web-scale pretraining, including information that could be hazardous if misused. Removing such knowledge from the model itself risks degrading general capability. A new arXiv preprint suggests an architectural solution: instead of deleting hazardous knowledge, relocate it into specific experts within a mixture-of-experts (MoE) model, then remove those experts when needed.

Because MoE models route inputs to different expert modules, the approach aims to isolate risky capabilities in a small, removable subset. The abstract indicates this can enable safety control at the architectural level while preserving overall performance. However, the abstract is truncated, so details on implementation and evaluation are not available from the source.

The work addresses a key tension in AI safety: how to make models safer without sacrificing their usefulness. Whether the method works in practice remains to be seen, as the full paper is not yet accessible from the abstract alone.