Figure 7b
Finite intervention paths
The 72 model–objective–target geometric-mean cost ratios span 11.5–127.1, with positive paired intervals in every cell and target.
Loading recorded measurements…
Methods and interpretation
864 observations from 216 prompts, 16 models and four targets. Path cost is the sum of KL between successive accepted outputs.
Procedure in the paper: Apply the frozen trust-region and stopping rules, recompute both directions after each accepted step, match the four final objective changes and sum the accepted local KL increments.
Cumulative local KL
864 matched-target observations, comprising 216 prompt cases at four targets. Path cost sums KL between successive accepted predictions, with directions recalculated at each step.
Cumulative local KL adds the divergence between successive accepted predictions along a path.
Code and data
v0.1.0 · 152 KB · View source on GitHub ↗
The complete 18-cell, 16-model finite-path battery with relinearized Fisher and Euclidean control, cumulative and endpoint KL, and paired prompt bootstrap.
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.