THEORY-1313: In two 7–8B chat models, the shift that finetuning induces along a difference-of-means persona vector correlates with how strongly the finetuned model then expresses that trait, r = 0.76–0.97 across synthetic datasets for evil, sycophancy and hallucination, above cross-trait baselines of 0.34–0.86; the last-prompt-token projection predicts a response's trait at r = 0.75–0.83 overall but mainly by telling prompt types apart

Source

Chen, Arditi, Sleight, Evans and Lindsey (2025), LIT-1998, arXiv:2507.21509v3; read in NOTE-1524.

What was actually shown

Finetuning shift (§4.2, Figure 6). For each finetuned model, the mean last-prompt-token activation change from the base model, projected on the unit persona vector, against the judge’s post-finetuning trait score: r = 0.76–0.97 over the datasets of Figure 5, for Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct. Cross-trait correlations are r = 0.34–0.86 (footnote 6, Appendix G.2), so the vectors are specific but not cleanly so; negative traits and humor shift together.

Prompt-time projection (§3.3, Appendix C.2). Overall r = 0.747 (evil), 0.798 (sycophancy), 0.830 (hallucination) under system-prompt interpolation; within a condition 0.511, 0.669, 0.245. The authors: the correlation arises “primarily from distinguishing between different prompt types”, and persona vectors “may be less reliable for more subtle behavioral changes” (p. 5).

The finetuning correlation could have failed: emergent-misalignment datasets that contain no trait expression could have moved behaviour without moving the vector. In the reported data they did not.

What this does not say

• That the vector is the trait’s mechanism: penalizing movement along it during training fails, suspected because the trait is re-encoded in other directions (p. 9).
• That it holds at other scales or in models that refuse the eliciting prompts (§8).
• That a persona vector is a component of a persona distribution: the paper fits no distribution over personas.