Robust honesty under emergent misalignment

Determine whether CATCH trained on a base model can improve honesty on a broadly misaligned model, thereby characterizing the robustness of CATCH to emergent misalignment.

Background

CATCH is evaluated on base models and transferred across several dishonesty benchmarks, including instructed dishonesty, incentive-driven dishonesty, and trained-in concealed behaviors. The paper does not evaluate whether the learned honesty intervention remains effective when the underlying model has broader emergent misalignment.

The authors identify this as a concrete direction for future work: testing whether an adapter trained on a base model can still improve honesty when applied to a broadly misaligned model.

References

Finally, future work could also evaluate CATCH’s robustness to emergent misalignment (Betley et al., 2026), for example, by evaluating whether CATCH trained on the base model can still improve honesty on a broadly misaligned model.

— CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty  (2610.22292 - Chang et al., 15 Oct 2026) in Section 6, page 8