Table 1; Extended Data 6

Sparse-feature cost and selectivity

Fisher cost predicts ablation KL with rank correlation 0.997. Fisher causal efficiency predicts selected-token-set specificity at 0.935. Activation and attribution comparators measure different properties.

Loading recorded measurements…

How much structure is needed for a clean intervention?

Loading recorded comparison…

Methods and interpretation

303 held-out zero-ablations from 57 active features and 18 context clusters at Pythia-410M. Attribution patching remains stronger for raw effect magnitude; specificity is a separate target.

Procedure in the paper: Acquire the specified pretrained sparse autoencoder at its fixed revision, freeze feature selection on calibration data, zero each active feature on held-out contexts and measure full-vocabulary KL and effects on the selected 16-token sets.

Fisher cost prediction

303 held-out feature–context zero-ablations. This compares the same KL quantity on both axes.

A sparse-autoencoder feature is a learned direction in activation space. Setting it to zero tests its causal effect in a specific context.

Activation and cost

The horizontal axis measures feature activation strength; the vertical axis measures output disturbance in KL units.

Fisher selectivity

Selectivity asks how concentrated the causal effect is, a different endpoint from raw change magnitude.

Attribution and selectivity

This comparison tests prediction of token-set specificity. Attribution patching is the stronger predictor in the separate raw-effect comparison.

Code and data

Sparse-feature ablationZIP

v0.1.0 · 70 KB · View source on GitHub ↗

The 303 held-out zero-ablations using the specified pretrained Pythia-410M sparse autoencoder.

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.