Skip to content

Pillar 6: Generalization

Generalization tests whether a circuit discovered for one task also explains a related task. If the circuit represents a genuine computational mechanism (e.g., "name-mover heads"), it should transfer partially to tasks that share that mechanism.

Preliminary status

Pillar 6 is implemented but not yet validated at scale. It has not been run on safety datasets or in a production sweep. Treat its scores as preliminary. This pillar is automatically skipped if target_task is not specified.

The Measurement

  1. Discover a circuit on the source task (e.g., IOI)
  2. Evaluate the same circuit (via Pillar 1 patching) on the source task to get its source_score
  3. Evaluate the same circuit on a target task (e.g., SVA) to get its target_score
  4. Report the transfer ratio = target_score / source_score

If the circuit generalizes, its target-task score stays close to its source-task score, so the transfer ratio approaches 1, even though the circuit was discovered for a different task.

Running Pillar 6

Pillar 6 requires target_task to be specified:

# Via Pipeline
pipe.evaluate(
    pillars=["generalization"],
    target_task="sva",         # required — the task to generalize to
    n_examples=256,
)
print(pipe.report.generalization)
# {
#   "source_task": "ioi",
#   "target_task": "sva",
#   "source_score": 0.83,             # circuit's patching score on the source task
#   "target_score": 0.61,             # circuit's patching score on the target task
#   "transfer_ratio": 0.73,           # target_score / source_score
#   "transfer_delta": 0.22,           # source_score - target_score
#   "relative_transfer_drop": 0.27,   # transfer_delta / source_score
# }

If target_task is not set, Pillar 6 is silently skipped.

Choosing Target Tasks

Transfer is most meaningful between tasks that share computational mechanisms:

Source task Related target tasks Shared mechanism
ioi sva, gender_bias, double_io Name-mover / indirect-object heads
boolq truthfulqa, mmlu Factual QA reasoning
capital_country hypernymy, ioi Lookup / relation heads
winogrande ioi, sva Coreference resolution

Interpreting the Results

Transfer Ratio Interpretation
≥ 0.70 Good transfer — circuit captures a task-general mechanism
0.50 – 0.70 Partial transfer — some overlap but task-specific components also matter
< 0.50 Poor transfer — the circuit is task-specific

Transfer ratio = target_score / source_score. A high transfer ratio means the circuit's patching score on the target task is close to its score on the source task.

Note: Even partial transfer (0.5 – 0.7) is meaningful — it indicates the algorithm found components that are not purely task-specific.

Low-Level API

At the low level, run_full_faithfulness needs both target_task_spec and target_dataloader. Passing only the spec makes Pillar 6 log a warning and skip — report.generalization comes back None. (The high-level pipe.evaluate(target_task=...) path builds the target dataloader for you, which is why the Pipeline snippet above only needs target_task.)

from circuitkit.evaluation.full import run_full_faithfulness
from circuitkit.tasks import get_task

target_spec = get_task("sva")
# Build the target dataloader yourself — omitting it silently skips Pillar 6.
target_dataloader = target_spec.build_dataloader(model, discovery_cfg, device)

report = run_full_faithfulness(
    model, graph, source_task_spec, discovery_cfg,
    pillars=["generalization"],
    target_task_spec=target_spec,
    target_dataloader=target_dataloader,
)
print(report.generalization)

Next Steps