Value steering

Can a value be dialled up by telling the model it holds that value? GPT-4o is given a system prompt naming one Schwartz value and its definition, then evaluated on Value Portrait. Each cell below is the change in a measured value relative to unsteered GPT-4o; the diagonal is how far the targeted value itself moved.

steering targets

8 / 10

targets where the target value rose by ≥ 0.3

+1.08

Security shift when steering toward Benevolence

−1.10

Universalism shift when steering toward Power

Steering matrix: change in each value when GPT-4o is steered toward the row value

Rows: steering target. Columns: measured value. ★ row maximum, † row minimum. Click a row to inspect it.

Table 17 of the paper. In 8 of 10 rows the targeted value rises by at least 0.3; the exceptions are Benevolence (+0.11) and Achievement (+0.19).

Steering toward Universalism

Target

Schwartz-circle expectations for this target

System prompt used

What steering shows

Most values steer as intended. Universalism (+0.66), Hedonism (+0.77), Power (+0.55), Tradition (+0.70), Security (+0.56) and Self-Direction (+0.48) all rise when targeted, and for Universalism, Power, Hedonism and Self-Direction the targeted value is the largest change in its row.
The Schwartz circle is respected. Steering toward Power lowers Universalism (−1.10) and raises Achievement (+0.49); Hedonism steering lowers Tradition and Conformity and raises Stimulation; Security steering lowers Stimulation (−0.67). Values that are adjacent on the circle move together and opposing values move apart, as the theory predicts.
Benevolence is not understood as its own value. Steering toward Benevolence raises Security by 1.08 and Self-Direction by 0.44 but Benevolence itself by only 0.11. GPT-4o's picture of "caring for one's in-group" seems to be dominated by safety and stability rather than by the Benevolence items people who hold that value endorse.
Achievement steering leaks into Self-Direction. The Achievement prompt moves Self-Direction (+0.63) more than Achievement (+0.19) and lowers Universalism (−0.67).

Together with the demographic analysis, this suggests that prompting can control some value dimensions reliably while others interact in ways a developer would not anticipate from the prompt alone, which is what a benchmark anchored in human value holders is designed to reveal.

Value definitions given to the model