Skip to the content.

Discovered circuit edges — the key-patch run over the top content readers in every model

A working catalog (amateur, exploratory, provisional) of de-novo composition edges: for the top behavioural content readers in each model we path-patch every upstream head out of the reader’s key and keep the edges that collapse the reader’s attention beyond a reader-matched null. writer → reader (K-composition); live = robust collapse.

6 models · top content readers × all upstream · faithful key-only patch.

gpt2 (GPT-2/absolute) — 6/6 live edges (prev-token head 4.11)

reader pattern top upstream writer key-collapse z live?
5.1 induction 4.11 (=prev-tok head) +46% 16.6 yes
6.9 induction 4.11 (=prev-tok head) +26% 19.8 yes
7.2 induction 4.11 (=prev-tok head) +25% 16.3 yes
5.5 induction 4.11 (=prev-tok head) +22% 24.5 yes
7.10 induction 4.11 (=prev-tok head) +17% 30.7 yes
3.0 duplicate 1.9 +15% 5.0 yes

gpt2-medium (GPT-2/absolute) — 2/6 live edges (prev-token head 5.11)

reader pattern top upstream writer key-collapse z live?
7.2 induction 4.13 +17% 12.9 yes
9.9 induction 4.13 +11% 15.4 yes
11.1 induction 4.13 +9% 42.7 no
7.11 duplicate 4.6 +9% 21.7 no
12.1 induction 4.13 +8% 22.5 no
18.5 induction 1.6 +1% 11.1 no

gpt2-large (GPT-2/absolute) — 2/6 live edges (prev-token head 14.1)

reader pattern top upstream writer key-collapse z live?
5.19 duplicate 3.3 +17% 135.7 yes
5.8 duplicate 3.3 +16% 191.7 yes
16.9 induction 3.14 +5% 13.1 no
15.4 induction 3.14 +3% 10.4 no
19.4 induction 3.14 +2% 11.5 no
16.0 induction 3.14 +1% 16.9 no

gemma-2-2b (RoPE) — 0/6 live edges (prev-token head 21.7)

reader pattern top upstream writer key-collapse z live?
6.2 induction 5.0 +5% 37.2 no
6.3 induction 5.0 +3% 35.0 no
5.4 duplicate 4.4 +2% 4.7 no
8.1 duplicate 6.1 +2% 6.0 no
3.2 duplicate 2.6 +1% 2.3 no
1.4 duplicate 0.0 +1% 4.9 no

Llama-3.2-1B (RoPE) — 3/6 live edges (prev-token head 0.2)

reader pattern top upstream writer key-collapse z live?
4.16 duplicate 1.9 +30% 35.2 yes
5.10 induction 1.20 +29% 40.5 yes
6.8 duplicate 1.9 +20% 22.9 yes
12.15 induction 1.9 +4% 10.0 no
12.0 duplicate 1.28 +2% 9.4 no
10.23 induction 1.9 +2% 12.3 no

Qwen2.5-1.5B (RoPE) — 1/6 live edges (prev-token head 13.4)

reader pattern top upstream writer key-collapse z live?
2.3 induction 1.4 +85% 51.8 yes
14.4 induction 5.5 +2% 2.9 no
19.5 induction 1.4 +1% 4.6 no
19.3 induction 1.4 +0% 9.1 no
8.3 duplicate 0.4 +0% 5.4 no
14.3 induction 2.10 +0% 5.3 no

How to read this

Data: runs/disassembly/circuits/discovered_summary.json. Regenerate: circuit_discovery.py.