Skip to the content.

Operator structural

structural operator (universal/addressing — measured across all models in the catalog).

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 0.199 ± 0.001 2 3.1 0.27 -0.012 ± 0.005
gpt2-medium GPT-2/absolute 0.255 ± 0.001 4 3.1 0.13 -0.032 ± 0.002
gpt2-large GPT-2/absolute 0.391 ± 0.007 9 4.7 0.11 -0.002 ± 0.004
gemma-2-2b RoPE 0.238 ± 0.003 11 24.6 0.96 -0.120 ± 0.012
Llama-3.2-1B RoPE 0.351 ± 0.003 19 0.31 0.00 +1.382 ± 0.450
Qwen2.5-1.5B RoPE 0.172 ± 0.017 4 25.7 0.93 -0.055 ± 0.012

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 3.1 _your/Your/_Your; The; _it/it -0.03 (≈neutral)
gemma-2-2b 24.6 ▁the/⏎⏎/.; First; ▁with/With/▁With +0.11 (copies)

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