Skip to the content.

Operator sink

addressing — attention sink: park attention on key-0 (the no-op / idle register)

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 0.959 ± 0.000 117 7.2 0.64 +0.024 ± 0.006
gpt2-medium GPT-2/absolute 0.965 ± 0.000 335 9.9 0.39 +0.000 ± 0.002
gpt2-large GPT-2/absolute 0.972 ± 0.002 555 19.4 0.54 +0.000 ± 0.001
gemma-2-2b RoPE 0.059 ± 0.001 0 0.3 0.00 -0.024 ± 0.043
Llama-3.2-1B RoPE 0.997 ± 0.000 446 5.11 0.33 -0.003 ± 0.002
Qwen2.5-1.5B RoPE 0.990 ± 0.012 292 14.5 0.52 -0.001 ± 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 19.18 1145 +0.18 -0.00 distributed (full +0.18 ≫ best 1h +0.05) — (addresses by position/key-0) — (addresses by position/key-0)
gpt2 7.2 136 +3.10 +0.03 distributed (full +3.10 ≫ best 1h +1.31) — (addresses by position/key-0) — (addresses by position/key-0)
gpt2-medium 9.9 378 +0.41 +0.01 distributed (full +0.41 ≫ best 1h +0.12) — (addresses by position/key-0) — (addresses by position/key-0)
gpt2-large 19.4 687 +0.17 +0.00 distributed (full +0.17 ≫ best 1h +0.04) — (addresses by position/key-0) — (addresses by position/key-0)
gemma-2-2b 1.0 67 -0.18 -0.49 compensatory (peak +0.62@2h → full -0.18; non-monotonic) — (addresses by position/key-0) — (addresses by position/key-0)
Llama-3.2-1B 5.11 508 +0.42 -0.00 distributed (full +0.42 ≫ best 1h +0.22) — (addresses by position/key-0) — (addresses by position/key-0)
Qwen2.5-1.5B 14.5 312 +0.19 -0.00 distributed (full +0.19 ≫ best 1h +0.08) — (addresses by position/key-0) — (addresses by position/key-0)

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 7.2 US/us; _Citizen/_citizens; I +0.00 (≈neutral)
gemma-2-2b 0.3 cius/▁belly/VIR; UMN/GIL/▁Cai; ▁Citizen +0.10 (copies)

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

A · identity (behavioural: top heads by attention mass on the sink pattern (>0.02)): heads [‘7.2’, ‘5.1’, ‘6.9’, ‘7.10’, ‘9.9’]. ranked: 7.2 (0.94), 5.1 (0.94), 6.9 (0.89), 7.10 (0.87), 9.9 (0.85), 9.6 (0.82)

B · causal × tasks (* = beyond random control): generic +0.01, induction +4.60, copy_names +11.55, successor +0.03, ioi +0.07 → serves [‘induction’, ‘copy_names’]

C · channels (reader 7.2): KEY/match top 0.9 collapse +4% (concentration 261.9×); VALUE/move top 5.9 ΔV-out 0.34 (median 0.06).

D · composition: IN→key 4.11(0.089), 4.7(0.087), 4.6(0.073), 5.6(0.072); OUT→value 9.3(0.043), 10.9(0.042), 8.7(0.042), 11.8(0.039).

E · redundancy (task generic): solo 5.1(+0.00), 6.9(+0.00), 7.10(+0.00), 7.2(+0.00), 9.9(-0.00); cumulative 1h +0.00 → 2h +0.00 → 3h +0.01 → 4h +0.01 → 5h +0.01 → DISTRIBUTED population (full +0.01 ≫ best single +0.00).

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