Transfer of goals beyond targeted fine-tuning

Determine whether value transplant can transfer goals other than honesty and test gaming, or can operate with goal differences that arise without targeted fine-tuning.

Background

The experiments evaluate value transplant using coding tasks and organisms deliberately fine-tuned to favor either genuine correctness or test gaming. The authors explicitly identify the generality of the intervention across other goals and naturally arising goal differences as unresolved.

References

It remains unclear whether the same approach can transfer other goals or work with goal differences that arise without this targeted fine-tuning.

— Steering Language Model Goals with Value Transplant  (2609.34056 - Jiang et al., 28 Sep 2026) in Section 5, Limitations

We have not established why this amplification is needed or whether a different axis or intervention would allow effective transfer at lower doses.

— Steering Language Model Goals with Value Transplant  (2609.34056 - Jiang et al., 28 Sep 2026) in Section 5, Limitations