Figure viewer
← Back to article
A constant-length arrow rotates through nested Fisher ellipses while a cost bar grows. Beside it, measured KL and local predictions agree at small doses and separate at full dose.
Local geometry · Measurement + analytic model

Equal-length moves can have different local costs (left: mathematical example). In E3 (right), Fisher predicts small KL changes well but loses accuracy at the original edit strength.

Settings, scope and source files
Left: equal-length arrows cross different cost contours of the exact quadratic form F=diag⁡(1,3)F=\operatorname{diag}(1,3); this is not measured Qwen geometry. The rotating arrow keeps its Euclidean length while its quadratic cost changes. Right: E3 on 32 evaluation prefixes, with Fisher estimated separately for each prefix. Gold points show all measured prefixes at the highlighted dose. Highlights step between tested doses; lines join their medians. The full-dose statistic is the median of per-prefix prediction ratios, not the ratio of medians.
Plays once when in view · pause at any time