Skip to content

Pillar 2: Ablation

Ablation is the sufficiency test: if we keep the circuit nodes and ablate everything else (set the out-of-circuit activations to zero/mean), does the model still produce the behavior?

A circuit that survives this — behaving like the full model even with all other components stripped out — is sufficient on its own. Pillar 2 measures how much of the clean behavior the circuit alone recovers.

The Measurement

Setup: 1. Keep the circuit nodes and ablate the out-of-circuit nodes (set their activation to zero, the dataset mean, or a position-specific mean) 2. Measure the logit-difference metric on the ablated model, \(y_\text{circuit}\) 3. Normalize against the clean and corrupt baselines

The headline score is the normalized faithfulness ratio:

\[F = \frac{y_\text{circuit} - y_\text{corrupt}}{y_\text{clean} - y_\text{corrupt}}\]

where \(y_\text{clean}\) is the metric on the full (unmodified) model, \(y_\text{corrupt}\) is the metric on the corrupt baseline, and \(y_\text{circuit}\) is the metric on the ablated circuit. \(F = 1.0\) means the ablated circuit fully recovers the clean behavior; \(F = 0.0\) means it does no better than the corrupt baseline.

Degenerate baselines

When the clean and corrupt baselines are near-identical (\(|y_\text{clean} - y_\text{corrupt}|\) below a small epsilon), the denominator collapses and the ratio is undefined. In that case the pillar logs a warning and reports the sentinel score 0.0.

Running Pillar 2

# Via Pipeline
pipe.evaluate(pillars=["ablation"], n_examples=256)
print(pipe.report.ablation_score)

# Via evaluate_circuit  
results = evaluate_circuit({..., "eval": {"pillars": ["ablation"]}})

Low-level ablation strategies:

from circuitkit.evaluation.evaluate import evaluate_graph

# Zero ablation (fastest)
score = evaluate_graph(model, graph, dataloader, metrics=[metric_fn],
                       intervention="zero", quiet=True)

# Mean ablation (more principled — ablates to the dataset mean)
score = evaluate_graph(model, graph, dataloader, metrics=[metric_fn],
                       intervention="mean",
                       intervention_dataloader=dataloader,
                       quiet=True)

# Mean-positional ablation (mean conditioned on position)
score = evaluate_graph(model, graph, dataloader, metrics=[metric_fn],
                       intervention="mean-positional",
                       intervention_dataloader=dataloader,
                       quiet=True)

Ablation Strategies

Strategy What it does Use when
"zero" Sets out-of-circuit activations to 0 Fast baseline; may introduce artifacts for biased activations
"mean" Sets out-of-circuit activations to the dataset mean (computed over intervention_dataloader) Standard principled ablation
"mean-positional" Mean conditioned on position in the sequence When position matters (e.g., the indirect-object position in IOI)

"mean" and "mean-positional" require an intervention_dataloader to compute the mean activations; "zero" needs no extra data. Any other value raises ValueError.

Interpreting the Score

Faithfulness Ratio \(F\) Interpretation
≥ 0.80 Strong sufficiency. The circuit alone recovers ≥ 80% of the clean behavior.
0.50 – 0.80 Moderate sufficiency. The circuit carries much of the behavior but misses some.
< 0.50 Low sufficiency. The circuit alone is not enough — important components live outside it.

A low score means ablating the out-of-circuit nodes destroys behavior that the circuit was supposed to carry on its own — a sign the circuit is incomplete.

Pillar 1 vs. Pillar 2 Together

Both pillars probe sufficiency, from different angles: Pillar 1 patches the circuit into a corrupt run, while Pillar 2 keeps the circuit and ablates everything else.

P1 (Patching) P2 (Ablation) Interpretation
High High Faithful circuit — sufficient under both interventions
High Low Fragile under ablation — recovers behavior when patched in, but not when isolated
Low High Fragile under patching — survives isolation, but does not transfer into a corrupt run
Low Low Poor circuit — insufficient under either intervention

For a strong faithfulness claim, a circuit should score high on both Pillar 1 and Pillar 2.

Next Steps