Skip to the content.

Operator duplicate

op:duplicate — this is the operator (a head class: the family of heads that realize the operation). Not the duplicate circuit, which is the composition (a writer-op feeding a reader-op) named after — and built around — this operator.

content — duplicate-token head: attend to an earlier occurrence of the same token

Cross-model (catalog row) — signal/causal are mean ± σ over 3 probe-resample seeds

model arch signal (±σ) #heads top head depth causal ΔNLL (±σ)
gpt2 GPT-2/absolute 0.622 ± 0.005 4 0.5 0.00 +0.114 ± 0.014
gpt2-medium GPT-2/absolute 0.859 ± 0.003 6 7.11 0.30 -0.008 ± 0.007
gpt2-large GPT-2/absolute 0.961 ± 0.002 23 5.8 0.14 +0.000 ± 0.003
gemma-2-2b RoPE 0.856 ± 0.007 11 1.4 0.04 -1.067 ± 0.027
Llama-3.2-1B RoPE 0.735 ± 0.008 16 6.8 0.40 +0.017 ± 0.005
Qwen2.5-1.5B RoPE 0.973 ± 0.001 16 8.3 0.30 +0.000 ± 0.003

Cross-model deep dossier (arch-generic) — operator_dossier_xmodel.py

The deep battery’s arch-generic core — behavioural head-ID + mean-ablation causal + the faithful key-only path-patch channel (the model re-applies its own RoPE) — run across every model, not just GPT-2. (The full A–F dossier below stays GPT-2-only: its channel/composition math is written against GPT-2’s fused-QKV layout, and the named output ops have no published head-set off GPT-2.)

model top head #heads (mass≥thr) causal induction ΔNLL causal generic ΔNLL redundancy (top heads) KEY top writer (collapse) VALUE top mover (ΔV-out)
gpt2-xl 4.12 155 +0.32 +0.00 distributed (full +0.32 ≫ best 1h +0.11) 2.6 (+6%, conc 220×) 2.13 (0.34)
gpt2 0.5 11 +0.49 +0.23 distributed (full +0.49 ≫ best 1h +0.12) — (addresses by position/key-0) — (addresses by position/key-0)
gpt2-medium 7.11 31 +0.15 -0.00 distributed (full +0.15 ≫ best 1h +0.04) 2.13 (+13%, conc 712×) 2.7 (0.29)
gpt2-large 5.8 106 +0.39 +0.00 distributed (full +0.39 ≫ best 1h +0.14) 3.3 (+31%, conc 865×) 3.17 (0.24)
gemma-2-2b 1.4 34 -0.28 -0.87 compensatory (peak +1.55@3h → full -0.28; non-monotonic) 0.0 (+1%, conc 20×) 0.1 (0.25)
Llama-3.2-1B 0.9 72 +0.94 +0.03 distributed (full +0.94 ≫ best 1h +0.59) — (addresses by position/key-0) — (addresses by position/key-0)
Qwen2.5-1.5B 8.3 81 +1.37 +0.00 distributed (full +1.37 ≫ best 1h +0.70) 0.4 (+0%, conc 44×) 6.5 (0.08)

Mean-ablate the op’s top behavioural heads → induction-NLL / generic-NLL damage; redundancy cumulative-ablates the top heads in solo-effect order (bottleneck = one head ≈ the whole op; distributed = the population far exceeds any single head; compensatory cases — which head triggers the recovery — are dug in outlier mechanism digs); channel = remove each upstream head from the reader’s key → top collapser + the value/move channel. Data: xmodel_dossiers_summary.json. Regenerate: operator_dossier_xmodel.py.

SAE-feature operands (section G)

What this operator reads/writes in feature space (monosemantic SAE latents), via the per-layer GPT-2 SAEs / Gemma Scope — see the full SAE-operand table. READ = dominant key-feature where the head attends (glossed by top tokens); copy-score = OV→unembed on those tokens (+ copies / − suppresses). Provisional, single corpus; for positional/addressing ops the read-feature is incidental.

model head reads (SAE feature) copy-score
gpt2 0.5 _you; _the; US +0.09 (copies)
gemma-2-2b 1.4 cius/▁belly/VIR; ▁the; UMN/GIL/▁Cai -0.15 (suppresses)

Deep dossier (GPT-2) — operator_dossier.py --op duplicate

A · identity (behavioural: top heads by attention mass on the duplicate pattern (>0.02)): heads [‘0.5’, ‘3.0’, ‘0.1’, ‘1.11’, ‘0.10’]. ranked: 0.5 (0.59), 3.0 (0.53), 0.1 (0.32), 1.11 (0.18), 0.10 (0.11), 1.5 (0.08)

B · causal × tasks (* = beyond random control): generic +0.07, induction +0.57, copy_names +3.15, successor -0.25, ioi +0.35* → serves [‘generic’, ‘copy_names’, ‘ioi’]

C · channels: reader in layer 0 — no upstream; channel skipped

D · composition: IN→key —; OUT→value 9.9(0.052), 7.10(0.049), 11.2(0.047), 7.2(0.047).

E · redundancy (task successor): solo 3.0(+0.00), 1.11(-0.00), 0.5(-0.01), 0.1(-0.03), 0.10(-0.06); cumulative 1h +0.00 → 2h +0.01 → 3h -0.00 → 4h -0.04 → 5h -0.25 → DISTRIBUTED population (full -0.25 ≫ best single +0.00).

F · cross-model: gpt2 sig 0.59; gpt2-medium sig 0.89/gain +12.6; Qwen2.5-1.5B sig 0.99/gain +14.1

Data: runs/disassembly/operators/dossiers/duplicate/ + the catalog. Regenerate: operator_catalog_doc.py.