Copy Suppression
Section titled “Copy Suppression”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.
Readings of the claim
Section titled “Readings of the claim”Copy suppression. Readings of the claim, from the version the authors state to the version the field cites.
| Reading | Verdict | Missing |
|---|---|---|
| Primary. Copy suppression is the main role of Negative Head L10H7 in GPT-2 Small, covering at least 76.9% of its direct effect | Mechanistically Supported | A 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 itself | Disconfirmed | Nothing — 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 layers | Underdetermined | Either 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 models | Proposed | The 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) |
Verdict
Section titled “Verdict”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.
| Tier | Requires | Missing |
|---|---|---|
| Proposed | construct (C1–C2) | Reached. |
| Causally Suggestive | measurement (M2); internal (I1) | Reached. |
| Mechanistically Supported | internal (I2, I4); external (E1) | Reached. |
| Triangulated | construct (C3–C4); internal (I5–I7); external (E2, E4) | Blocked. I6 double dissociation — Materials present in Table 2, never assembled into a dissociation |
| Validated | construct (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 |
36-criterion audit
Section titled “36-criterion audit”Copy suppression. Full 36-criterion audit. Status: C = Confirmed, PC = Partially confirmed, U = Untested, I = Inconclusive, D = Disconfirmed, N/A = Not applicable.
| ID | Criterion | Status | Evidence |
|---|---|---|---|
| Construct Validity | |||
| C1 | Falsifiability | C | Three conjunctive conditions, 80% pass; weight predictions stated first |
| C2 | Structural plausibility | C | Read off the weights: 84.70% OV, 95.72% QK, before any ablation |
| C3 | Convergent validity | C | Four instruments differing in kind agree: weights, coding, ablation, OOD |
| C4 | Discriminant validity | PC | Three rivals separated; only the conjunction discriminates, not either half |
| C5 | Nomological validity | PC | Unifies negative name movers with anti-induction; formation unexplained |
| C6 | Complementation validity | U | OV and QK split within one head; the across-head trans test is absent |
| Measurement Validity | |||
| M1 | Reliability | PC | Spread across data in 100 percentiles; no interval on any estimate |
| M2 | Baseline separation | PC | Mean ablation is the denominator; no null constructed for 76.9% itself |
| M3 | Stability | PC | Perturbations run and disclosed where they fail the idealization |
| M4 | Calibration | PC | KL’s failure mode named; the effect is small and sign-unstable |
| M5 | Sensitivity | U | Every control is a known-negative; no planted mechanism recovered |
| M6 | Invariance | PC | All of OpenWebText under three metrics; no domain or frequency split |
| M7 | Selection correction | U | Selection disclosed at every step and corrected at none |
| Internal Validity | |||
| I1 | Necessity | PC | Direct path localized cleanly; the absolute effect is a thousandth of loss |
| I2 | Sufficiency | PC | CSPA preserves 76.9%, but the figure spans 25–95% across metrics |
| I3 | Minimality | PC | Each half mild alone and the conjunction costs; nothing left to prune |
| I4 | Specificity | PC | CSPA discriminates L10H7; its components clear 50% for many heads |
| I5 | Rival mechanism exclusion | PC | Two rivals addressed; the perpendicular queryside driver left unknown |
| I6 | Double dissociation | U | Materials present in Table 2, never assembled into a dissociation |
| I7 | Confound control | PC | Tied embeddings and LayerNorm handled; data confounds untouched |
| I8 | Confounding sensitivity | U | Unattempted; nothing bounds an unmeasured confounder |
| I9 | Epistatic interaction | PC | One interaction quantified, copy suppression on itself; no grid |
| I10 | Rescue reversibility | U | Clean activations already in hand; the restore is never run |
| I11 | Onset coupling | PC | No onset experiment at origin; run later in a list sorter (Urdshals & Urdshals, 2025) |
| I12 | Offset coupling | U | Pythia checkpoints already in use; no offset observed anywhere |
| External Validity | |||
| E1 | Intervention reach | PC | Four intervention forms, all activation edits on one head |
| E2 | Prompt generalization | C | Averaged over the pretraining distribution, not a template set |
| E3 | Cross-task generalization | C | IOI and random-token scores correlate across three model families |
| E4 | Cross-model recurrence | PC | Phenomenon replicates in three systems; the mechanism does not |
| E5 | Graded response | PC | A genuine dose-response at the query; every other intervention binary |
| E6 | Novel prediction | C | Anti-induction on random tokens predicted, then confirmed |
| Interpretive Validity | |||
| V1 | Level declaration | C | Three numbered steps mapped to matrices, with the remainder named |
| V2 | Level-evidence match | PC | Matched on mechanism and coverage; when and how much it fires is not |
| V3 | Alternative level | PC | Appendix K favors semantic copy suppression over the paper’s own name |
| V4 | Unlicensed labeling | C | Every label redescribes a measured object; no intentional vocabulary |
| V5 | Scope declaration | C | Scope declared where each claim is made, not collected at the end |
| Total: 9 Confirmed, 20 Partially confirmed, 7 Untested | |||
| Verdict: Mechanistically Supported |
Exploratory Lens Analysis
Section titled “Exploratory Lens Analysis”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.