Skip to content

Source. Successor Heads: Recurring, Interpretable Attention Heads In The Wild (Gould et al., 2024).

Description. One attention head maps an ordinal token to its successor, and the mechanism is a product of four weight matrices — embedding, MLP0MLP_0, the head’s OV circuit, unembedding — evaluated directly, with no activation from any prompt entering it. Mod-10 features recovered from MLP0MLP_0 by sparse autoencoders, by a linear probe and by single-neuron ablation carry the index the head increments. Evaluated on: GPT-2, Pythia and Llama-2 from 31M to 12B for behavioral recurrence, with the feature-level case study on Pythia-1.4B L12H0 and two appendix replications. Description mode: implementational-functional.

Successor heads. Readings of the claim, from the version the authors state to the version the field cites.

ReadingVerdictMissing
Primary. A single head’s effective OV circuit increments an ordinal token to its successor, and mod-10 features carry the index it acts onMechanistically SupportedNecessity measured rather than ranked (I1); a known-positive control (M5); correction for the head, feature and $$ selections (M7); and the feature account does not reproduce the greater-than bias the same weights show, which the authors state as incomplete (V2)
The head is specific to successionDisconfirmedNothing — tested and failed: on natural text acronym and greater-than behavior take 23.8% and 18.9% of winning cases, and the paper reframes the head as interpretably polysemantic (I4)
The mod-10 mechanism recurs across architectures and sizes, as the abstract statesUnderdeterminedMechanistic recurrence is one case study plus two appendix replications against broad behavioral recurrence (E4); Llama-7B fails the held-out Roman-numeral task where the Pythia models do not (E4, M6), and putting the head on synthesised Roman-numeral representations drops top-1 accuracy to 0.125 (E3, E6)
The emergence of successor heads during training explains when the incrementation behavior appearsUnderdeterminedSuccessor scores are tracked across checkpoints in two model families and the heads are seen to emerge, but nothing is plotted against a behavior or a loss curve, and the phase change the induction-head comparison predicts is absent (I11); no checkpoint interval shows the mechanism lapsing (I12)

Successor heads. Verdict: Mechanistically Supported. 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)Blocked. I6 double dissociation — Neither arm run, and the polysemantic head makes the converse hard
Validatedconstruct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5)Blocked at Triangulated. C6 complementation validity — Eight task classes scored individually and never ablated in pairs; I10 rescue reversibility — Both ablations reversible and the clean run in hand; restore never run; I12 offset coupling — The checkpoint data exists and is read in the forward direction only; M5 sensitivity — Every control a known-negative; no planted circuit recovered

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

IDCriterionStatusEvidence
Construct Validity
C1FalsifiabilityCA fixed token list and a committed threshold; some models have no head
C2Structural plausibilityCA product of four matrices tested directly, with no prompt entering it
C3Convergent validityCA probe at 0.708 cosine and single neurons, sharing nothing with the SAE
C4Discriminant validityPCTop head eight times the runner-up; no neighboring construct attempted
C5Nomological validityPCWeak universality imported, tested, and partly returned (Chughtai et al., 2023)
C6Complementation validityUEight task classes scored individually and never ablated in pairs
Measurement Validity
M1ReliabilityPCModes over a hundred autoencoders; the central score has no interval
M2Baseline separationPCCutting the head leaves under one percent and returns bigrams instead
M3StabilityPCThree perturbations, one aimed at the paper’s own premise
M4CalibrationPCA proportion against a stated threshold; importance gives rank, not magnitude
M5SensitivityUEvery control a known-negative; no planted circuit recovered
M6InvarianceCThree families over three orders of magnitude, failures kept in
M7Selection correctionUSelection at three levels, disclosed at each and corrected at none
Internal Validity
I1NecessityPCWinning head on all 64 prompts, which ranks the head without measuring it
I2SufficiencyCThe OV circuit alone ranks the successor above every alternative
I3MinimalityPCMinimal by construction; the droppable component contributes under a percent
I4SpecificityPCOff-target function measured across five ablation regimes and found present
I5Rival mechanism exclusionPCArtifact and pre-head computation both excluded by built experiments
I6Double dissociationUNeither arm run, and the polysemantic head makes the converse hard
I7Confound controlPCTokenization handled by construction; letter collisions explained, not tested
I8Confounding sensitivityUUnattempted; nothing bounds an unmeasured confounder
I9Epistatic interactionPCIndirect effect bounded under both ablation values; no feature-level grid
I10Rescue reversibilityUBoth ablations reversible and the clean run in hand; restore never run
I11Onset couplingIEmergence observed across checkpoints; coupling to behavior is not
I12Offset couplingUThe checkpoint data exists and is read in the forward direction only
External Validity
E1Intervention reachCFive intervention forms spanning removal and addition, and they agree
E2Prompt generalizationPCA constructed token list and 128 natural contexts, the harder one genuinely hard
E3Cross-task generalizationCEight tasks scored individually; Roman numerals decode though never trained
E4Cross-model recurrencePCBehavior recurs across three families; the mechanism is single-model
E5Graded responsePCA two-sided dose criterion, and every other ablation all-or-nothing
E6Novel predictionCRoman numerals fitted out of every training set and decoded anyway
Interpretive Validity
V1Level declarationCThe level is fixed by the object: a product of four weight matrices
V2Level-evidence matchPCWeights tested on weights; the recurrence claim rests mostly on behavior
V3Alternative levelPCTwo alternatives closed by experiment; the paper’s own third left open
V4Unlicensed labelingCEach label defined in the sentence that introduces its measurement
V5Scope declarationPCThree limits declared beside the results they bound; one gap remains
Total: 10 Confirmed, 18 Partially confirmed, 1 Inconclusive, 7 Untested
Verdict: Mechanistically Supported

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