safety security
The Circuit Metric Preferred the Worse Circuit
Intervention faithfulness misranked equally sized mechanisms even when their held-out behavior was less faithful.
Summary
Intervention faithfulness misranked equally sized mechanisms even when their held-out behavior was less faithful.
Mechanistic-interpretability searches assume the objective will recognize a better circuit once it appears. This study shows an objective-level recovery gap: under ordinary resampling, KL-based faithfulness misranked 9.4 to 41.2 percent of candidate pairs across several discovery methods on human-reference tasks. Restoring selected recipient-side signals repaired 96 of 100 persistent misrankings without changing the circuits, suggesting that context distortion in the evaluation can reward the wrong mechanism.
Why it matters
Intervention faithfulness misranked equally sized mechanisms even when their held-out behavior was less faithful.
Limits and context
No additional limitation was separately recorded.
Key claims
Intervention faithfulness misranked equally sized mechanisms even when their held-out behavior was less faithful.
Evidence: source-2026-10-03-009
Sources
- arXiv preprint 2610.02098arXiv · primary research
Corrections
No corrections have been recorded for this story.