Skip to the content.

Operator self

addressing 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 0.839 ± 0.001 10 0.1 0.00 +0.024 ± 0.010
gpt2-medium GPT-2/absolute 0.447 ± 0.000 21 5.13 0.22 +0.033 ± 0.008
gpt2-large GPT-2/absolute 0.549 ± 0.003 74 0.14 0.00 +0.012 ± 0.002
gemma-2-2b RoPE 0.966 ± 0.001 154 25.7 1.00 +0.476 ± 0.027
Llama-3.2-1B RoPE 0.956 ± 0.001 60 15.14 1.00 +0.103 ± 0.003
Qwen2.5-1.5B RoPE 0.999 ± 0.000 43 15.7 0.56 +3.733 ± 0.059

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 0.1 _you; US; _the +0.07 (copies)
gemma-2-2b 25.7 ▁the/⏎⏎/.; First; ▁the/▁your/▁own -0.21 (suppresses)

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