Skip to the content.

Operator induction

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

content — in-context copy: attend to the key whose predecessor token == current token, copy it

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 0.926 ± 0.006 22 5.5 0.45 +0.011 ± 0.004
gpt2-medium GPT-2/absolute 0.915 ± 0.004 61 18.5 0.78 +0.003 ± 0.003
gpt2-large GPT-2/absolute 0.969 ± 0.002 75 16.0 0.46 +0.007 ± 0.000
gemma-2-2b RoPE 0.941 ± 0.005 23 6.3 0.24 -0.282 ± 0.019
Llama-3.2-1B RoPE 0.946 ± 0.005 40 10.23 0.67 +0.001 ± 0.001
Qwen2.5-1.5B RoPE 0.994 ± 0.001 54 14.3 0.52 -0.000 ± 0.002

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 21.3 237 +0.20 -0.00 distributed (full +0.20 ≫ best 1h +0.12) 2.11 (+0%, conc 104×) 2.13 (0.03)
gpt2 5.1 37 +3.67 +0.03 distributed (full +3.67 ≫ best 1h +1.31) 4.11 (+39%, conc 85×) 1.10 (0.22)
gpt2-medium 11.1 103 +0.70 +0.09 distributed (full +0.70 ≫ best 1h +0.12) 4.13 (+8%, conc 118×) 2.14 (0.08)
gpt2-large 16.0 145 +0.63 +0.01 compensatory (peak +1.18@3h → full +0.63; non-monotonic) 3.1 (+1%, conc 44×) 3.11 (0.07)
gemma-2-2b 6.3 51 +0.59 -0.31 compensatory (peak +1.98@2h → full +0.59; non-monotonic) 5.0 (+3%, conc 1672×) 5.5 (0.07)
Llama-3.2-1B 10.23 93 +1.24 +0.01 distributed (full +1.24 ≫ best 1h +0.77) 1.9 (+2%, conc 64×) 1.28 (0.08)
Qwen2.5-1.5B 14.3 98 +2.10 -0.00 distributed (full +2.10 ≫ best 1h +0.51) 2.10 (+0%, conc 17×) 1.4 (0.07)

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 5.5 First; What/_What/_what; US +0.09 (copies)
gemma-2-2b 6.3 First; ▁gods/▁run/▁petition; Before +0.06 (copies)

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

A · identity (behavioural: top heads by attention mass on the induction pattern (>0.02)): heads [‘5.1’, ‘5.5’, ‘6.9’, ‘7.10’, ‘7.2’]. ranked: 5.1 (0.81), 5.5 (0.78), 6.9 (0.77), 7.10 (0.75), 7.2 (0.72), 5.0 (0.57)

B · causal × tasks (* = beyond random control): generic +0.01, induction +6.39, copy_names +14.42, successor +0.01, ioi +0.28* → serves [‘induction’, ‘copy_names’, ‘ioi’]

C · channels (reader 5.1): KEY/match top 4.11 (=prev-token head) collapse +43% (concentration 90.8×); VALUE/move top 1.10 ΔV-out 0.22 (median 0.08).

D · composition: IN→key 4.11(0.101), 4.7(0.095), 1.0(0.070), 3.7(0.070); OUT→value 6.7(0.052), 6.8(0.051), 6.6(0.049), 7.0(0.048).

E · redundancy (task induction): solo 5.1(+1.70), 7.2(+0.22), 6.9(+0.21), 5.5(+0.08), 7.10(-0.03); cumulative 1h +1.70 → 2h +2.25 → 3h +3.82 → 4h +5.77 → 5h +6.39 → DISTRIBUTED population (full +6.39 ≫ best single +1.70).

F · cross-model: gpt2 sig 0.81; gpt2-medium sig 0.97/gain +12.6; Qwen2.5-1.5B sig 1.00/gain +14.1

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