🤖 AI Summary
This study addresses the challenge of identifying attention-head circuits responsible for specific tasks within pretrained Transformers under fully unsupervised conditions. The authors propose a three-step methodology: first, attention heads are ranked without supervision using a time-integrated participation ratio derived from spectral signals; second, candidate circuits are selected by aligning with task-specific activation patterns; and third, causal relevance is validated through grouped ablation against randomized controls. This approach is the first to identify causal circuits without requiring labels or gradients, consistently uncovering 2–6 essential inductive circuits across models ranging from 51M to 7B parameters and diverse architectures—ablating which degrades performance by 94%–100%. The work further reveals that the proportion of specialized computational heads remains conserved across models, falling within a narrow range of 17%–19%.
📝 Abstract
We present a three-step recipe for identifying attention-head circuits in pretrained transformers. A per-head spectral signal -- the time-integrated participation ratio of each head's attention output -- ranks heads doing sustained content-dependent computation without labels or attribution gradients. A task-pattern screen filters this general indicator into a task-specific candidate circuit, and group ablation against a matched-random control completes the causal claim. We validate across an 8x parameter range (51M to 1B-active / 7B-total), two architecture families (dense, mixture-of-experts), and four pretraining pipelines. The recipe ports: a 2-6 head induction circuit is causally necessary in every model tested, with a 94-100% drop in synthetic-induction top-1 after ablation. The spectral signal is predictive without supervision: on six independent seeds of a 51M-parameter probe model, the same computation identifies the seed-specific circuit on each seed. The fraction of heads doing identifiable specialized computation is conserved at 17-19% across the Pythia family (124M to 410M), while specific induction circuits stay 3-11 heads -- sublinear in total head count. This paper is the methodology anchor of a three-paper program; companion papers extend the recipe to developmental trajectories during pretraining and to composed-task circuits where pattern selectivity decouples from task-causal structure.