Figure 7c,d

Control that transfers across prompts

Reference-Fisher control reduces reference-sequence cost at matched donor changes and produces positive mean effects on unseen prompts in all 24 tested settings.

Loading recorded measurements…

Combining two reusable objectives

Loading recorded comparison…

Methods and interpretation

Four base models, two objectives and three donor targets. Four reference continuations measure cost; eight unseen prompts measure transfer. These are different evaluation sets.

Procedure in the paper: Build the fixed donor/reference/target splits, average the reference metric, solve shared updates at the donor targets and apply them to held-out prompts without using target gradients.

Reference-continuation cost

Targets are matched on the donor prompt. Four reference continuations evaluate sequence-level preservation. Compare Euclidean and donor-Fisher baselines separately using the observation labels.

A reference continuation is text used to measure preservation. It is different from the unseen prompts used to test transfer.

Effect on unseen prompts

Each mean uses eight unseen prompts. Both methods match the donor target, and their effects on unseen prompts are measured afterwards. Reference preservation cost is measured on separate reference continuations.

A donor prompt helps learn the update. An unseen prompt tests whether that update transfers beyond its training examples.

Code and data

Reusable reference-Fisher controlZIP

v0.1.0 · 72 KB · View source on GitHub ↗

The four-model donor/held-out/reference battery, three control metrics, reference sequence KL, target transfer and two-objective composition.

Interpretable steeringZIP

v0.1.0 · 62 KB · View source on GitHub ↗

SAE matching pursuit, interpretable token partitions and token-grouping versus CG flow fidelity.

Each standalone package includes code, shared helpers, required small inputs and reference results, with setup and commands in its README. Model weights and public datasets are obtained separately where needed.

Download complete source (v0.1.0) for all experiments and the companion website.

Paper and supplementary methods on arXiv ↗