$$\mathrm{PGR} \;=\; \frac{\mathrm{perf}(f_{w\to s}) - \mathrm{perf}(f_w)}{\mathrm{perf}(f_{s\to s}) - \mathrm{perf}(f_w)}$$

Equation 22 · AI Safety

The Main Technical Approaches to AI Alignment, Compared

This equation states an equality: the expressions on both sides have the same value under the article’s assumptions.

They quantify how much of the gap this recovers with a metric called performance gap recovered, defined by its two boundary conditions: it equals one under perfect weak-to-strong generalization and zero when the weak-to-strong model does no better than the weak supervisor it learned from [ 12 ] . Writing $f_w$ for the weak supervisor, $f_{w\to s}$ for the strong model fine-tuned on the weak model’s labels, and $f_{s\to s}$ for the strong model fine-tuned on ground truth as an upper-bound ceiling, those two conditions pin down $\mathrm{PGR} \;=\; \frac{\mathrm{perf}(f_{w\to s}) - \mathrm{perf}(f_w)}{\mathrm{perf}(f_{s\to s}) - \mathrm{perf}(f_w)}$. The paper also reports that simple additional interventions help: an auxiliary confidence loss recovered performance closer to GPT-3.5 level using only…

Meanings in this article

• $f_w$: the writing.

Equation guide → · Article →