nucleation

LIT-1579: Causal scrubbing: results on induction heads

1 min read

Properties6
StatusActive — read 2026-10-27 (NOTE-1196) in full, from the Alignment Forum's own API; worth reading as the quantified test of the textbook induction circuit in a two-layer attention-only model, pathway by pathway, and for what the refinements found: three-token matching, layer-0 heads that attend to copies of the current token, and a previous-token head that deviates on quotes, articles and punctuation (THEORY-1020). Read it knowing that its percentages are of the 0.053 nats the induction heads contribute on a filtered token subset, not of a label-shuffled loss like the paren post's, that its headline 89% is the bonus hypothesis and the main post's 86% the fourth, and that the model is not published.
Version2
Tagsrepresentation-learning, causality, anthology-candidate
ExtendsLIT-1549 — Causal Scrubbing: a method for rigorously testing interpretability hypotheses
Cited as “paper” byNOTE-1196 — Chan, Garriga-Alonso, Goldowsky-Dill, Greenblatt, Lin et al. — Causal scrubbing: results on induction heads
Cited as “source” byTHEORY-1020 — In a two-layer attention-only language model, on tokens where induction helps, the standard induction circuit (queries and values from the token embedding, keys from the previous-token head) accounts for 35% of the loss its two induction heads contribute, and a hypothesis that adds three-token matching, layer 0's average effect and three layer-0 heads acting as if they attended to the current token accounts for 86%

Graph View

Chan, L., Garriga-Alonso, A., Goldowsky-Dill, N., Greenblatt, R., Lin, T., Nitishinskaya, J., Radhakrishnan, A., Shlegeris, B. & Thomas, N. (2022). Causal scrubbing: results on induction heads. AI Alignment Forum, 3 December 2022.

Read in full on 2026-10-27 (NOTE-1196). One account was drawn from it: THEORY-1020.

Why this is here

The owner asked on 2026-10-27 to “get started on the seven works and markov IBP” that the circuits readings of 2026-10-26 had named without registering. The record’s induction line holds the circuit’s discovery (LIT-1534, LIT-1538) and its causal case in small attention-only models (THEORY-973), and the causal-scrubbing reading (NOTE-1168) said that the model, filter and details of the 35% and 86% figures were in this post, unread. This is the only quantified test in the record of how much of the induction heads’ contribution the textbook circuit accounts for, and it finds the circuit to be the core of a wider computation, not the whole of it.

Backlinks