Skip to content

Source. Copy Suppression: Comprehensively Understanding an Attention Head (McDougall et al., 2023).

Description. Negative Head L10H7 in GPT-2 Small suppresses tokens that earlier components have already predicted, in three steps: an early component predicts a token that already appears in context, the head attends back to that token, and its OV circuit writes a negative logit for it. The QK and OV halves are read off the weights before any ablation, with 84.70% of tokens carrying their OV diagonal among the ten most negative entries of its column and 95.72% of QK rows carrying the diagonal as the largest entry. A structured ablation that deletes every function of the head except those two preserves 76.9% of its effect on OpenWebText, and copy-suppression scores correlate head by head with anti-induction scores across GPT, Pythia and SoLU models. Description mode: implementational-functional.

Copy suppression. Readings of the claim, from the version the authors state to the version the field cites.

ReadingVerdictMissing
Primary. Copy suppression is the main role of Negative Head L10H7 in GPT-2 Small, covering at least 76.9% of its direct effectMechanistically SupportedA dissociation from the backup-head mechanism the same ablation moves (I6), a known-strength positive control fixing the method’s floor (M5), and correction for the top-5% slices the headline figures are computed on (M7)
The token the head suppresses is the source token itselfDisconfirmedNothing — tested and failed: in 42.00% of large-attention pairs the source token is suppressed without being the most suppressed, and a semantically related token takes first place in 90% of those (M6); the authors conclude the mechanism is better read as semantic copy suppression and keep the coarser name elsewhere (V3)
Copy suppression is what separates L10H7 from the other heads in its layersUnderdeterminedEither half of the instrument alone clears 50% recovered KL for many layer 9–11 heads, so the conjunction rather than the mechanism does the separating (I4, C4); the queryside direction perpendicular to the IO unembedding matters more than the parallel one and is left uncharacterized (I5)
Negative Heads do copy suppression across modelsProposedThe structured ablation is never run outside GPT-2 Small (E4); GPT-2 Medium recovers two of its three most negative heads, Pythia’s copy suppression is weaker, and Stanford GPT-2 Small E’s analogue attends to IO and S2 equally (M1)

Copy suppression. 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 — Materials present in Table 2, never assembled into a dissociation
Validatedconstruct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5)Blocked at Triangulated. C6 complementation validity — OV and QK split within one head; the across-head trans test is absent; I10 rescue reversibility — Clean activations already in hand; the restore is never run; I12 offset coupling — Pythia checkpoints already in use; no offset observed anywhere; M5 sensitivity — Every control is a known-negative; no planted mechanism recovered

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

IDCriterionStatusEvidence
Construct Validity
C1FalsifiabilityCThree conjunctive conditions, 80% pass; weight predictions stated first
C2Structural plausibilityCRead off the weights: 84.70% OV, 95.72% QK, before any ablation
C3Convergent validityCFour instruments differing in kind agree: weights, coding, ablation, OOD
C4Discriminant validityPCThree rivals separated; only the conjunction discriminates, not either half
C5Nomological validityPCUnifies negative name movers with anti-induction; formation unexplained
C6Complementation validityUOV and QK split within one head; the across-head trans test is absent
Measurement Validity
M1ReliabilityPCSpread across data in 100 percentiles; no interval on any estimate
M2Baseline separationPCMean ablation is the denominator; no null constructed for 76.9% itself
M3StabilityPCPerturbations run and disclosed where they fail the idealization
M4CalibrationPCKL’s failure mode named; the effect is small and sign-unstable
M5SensitivityUEvery control is a known-negative; no planted mechanism recovered
M6InvariancePCAll of OpenWebText under three metrics; no domain or frequency split
M7Selection correctionUSelection disclosed at every step and corrected at none
Internal Validity
I1NecessityPCDirect path localized cleanly; the absolute effect is a thousandth of loss
I2SufficiencyPCCSPA preserves 76.9%, but the figure spans 25–95% across metrics
I3MinimalityPCEach half mild alone and the conjunction costs; nothing left to prune
I4SpecificityPCCSPA discriminates L10H7; its components clear 50% for many heads
I5Rival mechanism exclusionPCTwo rivals addressed; the perpendicular queryside driver left unknown
I6Double dissociationUMaterials present in Table 2, never assembled into a dissociation
I7Confound controlPCTied embeddings and LayerNorm handled; data confounds untouched
I8Confounding sensitivityUUnattempted; nothing bounds an unmeasured confounder
I9Epistatic interactionPCOne interaction quantified, copy suppression on itself; no grid
I10Rescue reversibilityUClean activations already in hand; the restore is never run
I11Onset couplingPCNo onset experiment at origin; run later in a list sorter (Urdshals & Urdshals, 2025)
I12Offset couplingUPythia checkpoints already in use; no offset observed anywhere
External Validity
E1Intervention reachPCFour intervention forms, all activation edits on one head
E2Prompt generalizationCAveraged over the pretraining distribution, not a template set
E3Cross-task generalizationCIOI and random-token scores correlate across three model families
E4Cross-model recurrencePCPhenomenon replicates in three systems; the mechanism does not
E5Graded responsePCA genuine dose-response at the query; every other intervention binary
E6Novel predictionCAnti-induction on random tokens predicted, then confirmed
Interpretive Validity
V1Level declarationCThree numbered steps mapped to matrices, with the remainder named
V2Level-evidence matchPCMatched on mechanism and coverage; when and how much it fires is not
V3Alternative levelPCAppendix K favors semantic copy suppression over the paper’s own name
V4Unlicensed labelingCEvery label redescribes a measured object; no intentional vocabulary
V5Scope declarationCScope declared where each claim is made, not collected at the end
Total: 9 Confirmed, 20 Partially confirmed, 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 Copy Suppression — exploratory lens analysis.