Skip to the content.

Operator local

positional 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 0.336 ± 0.000 15 4.11 0.36 -0.009 ± 0.006
gpt2-medium GPT-2/absolute 0.338 ± 0.000 29 5.11 0.22 +0.030 ± 0.006
gpt2-large GPT-2/absolute 0.338 ± 0.000 44 14.1 0.40 +0.012 ± 0.001
gemma-2-2b RoPE 0.315 ± 0.000 49 0.0 0.00 +0.452 ± 0.032
Llama-3.2-1B RoPE 0.273 ± 0.000 8 0.2 0.00 +0.010 ± 0.005
Qwen2.5-1.5B RoPE 0.286 ± 0.001 22 1.4 0.04 +0.216 ± 0.029

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 4.11 US; cius/ius; I +0.10 (copies)
gemma-2-2b 0.0 cius/▁belly/VIR; UMN/GIL/▁Cai; ▁the +0.02 (≈neutral)

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