Circuit ioi_q_chain (GPT-2)

Q-composition chain (GPT-2-only) — scope: gpt2

The indirect-object-identification circuit: duplicate-token → S-inhibition → name-mover, a Q-composition chain (no published head-set outside GPT-2).

• S-inhibition heads: [‘10.0’, ‘8.3’, ‘8.10’, ‘6.7’]; name-movers: [‘9.9’, ‘11.2’, ‘8.11’, ‘10.0’]
• Q-composition live edges: 5; named-edge live-rate 0.42857142857142855
• IOI baseline logit-diff 2.7526195287704467
• causal z (ioi_causal.py): name-mover -2.172512152617327, S-inhibition -1.702171401577731, negative/copy-suppression 62.41397239659889 (writes against IO), duplicate 5.963926080885061, backup name-mover 0.8435903361394431
• self-repair (self_repair.py): −primaries ΔLD -0.0019234657287596768, −both 1.0389485120773314 → backups are hot spares (idle with primaries present, carry the circuit once they’re gone).

Cross-model IOI dossier (the circuit’s operators, found behaviourally — via the ResidualVM)

The Q-composition edge wiring below is GPT-2-validated, but the IOI behaviour and its operators are not GPT-2-only: the unified ResidualVM locates them in every model (name-movers by END→indirect-object copy-attention; negative-movers + the most load-bearing heads by an ablation sweep of the logit-diff; the duplicate-token initiator behaviourally). Logit-diff = logit(IO) − logit(S) at the end of a templated “When X and Y went…, Y gave a drink to →” prompt.

model	IOI logit-diff	name-movers (copy→IO)	negative-movers	most load-bearing	ablate name-movers	duplicate init
gpt2	+2.88	9.9, 10.0, 10.6	10.7, 9.6	8.10, 8.6	+22%	3.0
gpt2-medium	+3.11	15.14, 18.5, 20.6	18.9, 22.14	19.1, 12.3	+18%	7.11
gpt2-large	+4.09	22.0, 29.0, 23.13	32.0, 26.0	20.14, 18.3	+19%	7.5
gemma-2-2b	+3.45	18.6, 21.5, 22.5	22.4, 22.0	23.5, 20.6	+26%	8.1
Llama-3.2-1B	+5.85	12.13, 12.2, 11.14	12.15, 15.12	8.19, 11.4	+24%	8.19
Qwen2.5-1.5B	+6.00	27.4, 27.8, 24.10	23.8, 25.4	24.8, 0.6	+13%	8.3

• The IOI circuit is architecture-invariant — name-movers, negative/copy-suppression movers, and a duplicate-token initiator are present in all six models, and ablating the name-movers collapses the logit-diff (+13% to +26%) everywhere. The behaviour strengthens with GPT-2 scale (logit-diff +2.88 → +3.11 → +4.09) and is largest in the RoPE models (Llama +5.85, Qwen +6.00).
• The backup-name-mover self-repair is cross-model. The heads that are most load-bearing under ablation are not the name-movers (which are backed up) but the S-inhibition-type heads — in every model the ablation ranking and the copy-attention ranking disagree, the signature of name-mover redundancy (the Hydra effect) generalised beyond GPT-2.

Data: runs/disassembly/circuits/ioi_xmodel_summary.json (ioi_xmodel.py, built on the ResidualVM).

Data: runs/disassembly/circuits/atlas_summary.json + the discovery artifacts. Regenerate: circuit_catalog_doc.py.