Skip to content

Source. In-context Learning and Induction Heads (Olsson et al., 2022).

Description. A two-head composition: a previous-token head writes the identity of the token before a repeat into the residual stream, and an induction head reads it to predict what followed last time. Heads are identified by two behavioral scores on sequences of repeated uniform-random tokens, prefix matching and copying, and the mechanism is derived at the weight level for the two-layer attention-only case. Evaluated on: 34 decoder-only models from one to forty layers, with causal ablation confined to the twelve small models, six attention-only and six with MLPs. Description mode: implementational-functional.

Induction heads: token copying. Readings of the claim, from the version the authors state to the version the field cites.

ReadingVerdictMissing
Primary. Induction heads implement in-context copying of a token that followed the same context earlierTriangulatedMinimality (I3), complementation validity (C6), rescue reversibility (I10) and offset coupling (I12)
The mechanism is the source of in-context learning as measured by token-loss differenceMechanistically SupportedNecessity above 345M survives a decorrelating control, but the metric is uncalibrated (M4)
The mechanism is the source of in-context learning generallyDisconfirmedNothing — tested and failed; the bridge from the small-model result to the 13B claim is analogy (V2)
The two heads compose in the stated order, each performing the role its name assertsUnderdeterminedWeight-level derivation covers the two-layer case only; role semantics exceed it elsewhere (C2)

Induction heads: token copying. Verdict: Triangulated. The claim stops at the rung marked Blocked; rungs below it are Reached and rungs above it are blocked by that rung. Criteria in bold are absent or adverse at that rung; only those the rung requires can block it.

TierRequiresMissing
Proposedconstruct (C1–C2)Reached.
Causally Suggestivemeasurement (M2); internal (I1)Reached.
Mechanistically Supportedinternal (I2, I4); external (E1)Reached.
Triangulatedconstruct (C3–C4); internal (I5–I7); external (E2, E4)Reached.
Validatedconstruct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5)Blocked. C6 complementation validity — Two named roles, ablated one at a time and never together; I3 minimality — Unattempted; the authors name the obstacle in marginal effects; I10 rescue reversibility — Unattempted; pattern-preserving ablation already caches the clean run; I12 offset coupling — Unattempted at origin; no published offset test for induction heads

Induction heads: token copying. Full 36-criterion audit. Status: C = Confirmed, PC = Partially confirmed, U = Untested, I = Inconclusive, D = Disconfirmed, N/A = Not applicable.

IDCriterionStatusEvidence
Construct Validity
C1FalsifiabilityCTwo numerical evaluators with a stated failure mode; two predictions
C2Structural plausibilityCWeight-level: copying is the OV circuit’s positive-eigenvalue property
C3Convergent validityCFive families: eigenvalues, evaluators, ablation, scaling, training
C4Discriminant validityPCBuilt to separate induction from repeated-token memorization
C5Nomological validityCDerived from the framework rather than fitted to it (Elhage et al., 2021)
C6Complementation validityUTwo named roles, ablated one at a time and never together
Measurement Validity
M1ReliabilityPCRepeated across seeds in two-layer models by later work; never for the origin measurements
M2Baseline separationPCEvery head ablated at every snapshot, so the effect has a reference
M3StabilityPCToken indices perturbed and the score shown robust to the choice
M4CalibrationPCEvaluators calibrated against weight-level quantities in the appendix
M5SensitivityPCA known-positive: the two-layer model with a derived circuit (Elhage et al., 2021)
M6InvarianceC34 models, four series, 1–40 layers, up to 13B, 200 snapshots in the small models
M7Selection correctionUUnattempted; heads are scored, then the high scorers are ablated
Internal Validity
I1NecessityCAlmost all in-context learning in small models comes from these heads
I2SufficiencyPCNo experiment reconstructs copying from induction heads alone
I3MinimalityUUnattempted; the authors name the obstacle in marginal effects
I4SpecificityPCEvery head is ablated, so off-target effects are measured throughout
I5Rival mechanism exclusionPCBasic copying heads named and argued away rather than tested
I6Double dissociationCUnattempted at origin; run later and it succeeds (Feucht et al., 2025)
I7Confound controlPCThree exogenous confounds checked, including scheduled hyperparameters
I8Confounding sensitivityUUnattempted; no sensitivity analysis for an unmeasured confounder
I9Epistatic interactionPCComposition is the mechanism; the two-head circuit is non-additive
I10Rescue reversibilityUUnattempted; pattern-preserving ablation already caches the clean run
I11Onset couplingCAcross 34 models, heads and the loss bump appear at the same point
I12Offset couplingUUnattempted at origin; no published offset test for induction heads
External Validity
E1Intervention reachPCOne primitive at origin: pattern-preserving zero ablation of a head
E2Prompt generalizationCThe defining stimulus is out of distribution by construction
E3Cross-task generalizationPCThe same heads perform literal translation, verified against definition
E4Cross-model recurrenceC34 models across four series, plus two external replications
E5Graded responsePCBoth evaluators are continuous; the ablation is not interpolated
E6Novel predictionCArgument 2 predicts an architectural change will move the bump
Interpretive Validity
V1Level declarationCDeclared early and explicitly, by behavior on the defining stimulus
V2Level-evidence matchCClose for the narrow claim; the abstract’s narrow sentence is backed
V3Alternative levelPCOne alternative addressed; the authors offer their own caveat
V4Unlicensed labelingPCThe label imports an epistemic category, and the import is declared
V5Scope declarationCThe most thoroughly scoped claim in this audit set
Total: 14 Confirmed, 16 Partially confirmed, 6 Untested
Verdict: Triangulated

An exploratory reading of this claim through the framework’s five lenses, written for this site and not part of the paper, is at Induction Heads — exploratory lens analysis.