{ "tokens": ["When", "Mary", "and", "John", "went", "to", "the", "store", ",", "John", "gave", "a", "bottle", "of", "milk", "to"], "layers": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], "values": [ [0.01, 0.02, 0.01, 0.01, 0.01, 0.00, 0.00, 0.01, 0.00, 0.01, 0.00, 0.01, 0.00, 0.00, 0.00, 0.01], [0.02, 0.05, 0.02, 0.03, 0.01, 0.01, 0.01, 0.02, 0.01, 0.03, 0.01, 0.01, 0.01, 0.01, 0.01, 0.02], [0.03, 0.09, 0.03, 0.06, 0.02, 0.01, 0.01, 0.02, 0.01, 0.07, 0.02, 0.01, 0.01, 0.01, 0.01, 0.03], [0.04, 0.14, 0.04, 0.18, 0.03, 0.02, 0.02, 0.03, 0.02, 0.21, 0.04, 0.02, 0.02, 0.01, 0.02, 0.05], [0.05, 0.22, 0.05, 0.31, 0.04, 0.03, 0.02, 0.04, 0.03, 0.35, 0.06, 0.03, 0.02, 0.02, 0.02, 0.07], [0.06, 0.31, 0.07, 0.42, 0.05, 0.04, 0.03, 0.05, 0.03, 0.48, 0.08, 0.04, 0.03, 0.03, 0.03, 0.09], [0.08, 0.41, 0.09, 0.48, 0.07, 0.05, 0.04, 0.06, 0.04, 0.52, 0.11, 0.05, 0.04, 0.03, 0.04, 0.12], [0.10, 0.52, 0.11, 0.51, 0.08, 0.06, 0.05, 0.08, 0.06, 0.55, 0.14, 0.07, 0.05, 0.04, 0.05, 0.18], [0.12, 0.63, 0.13, 0.53, 0.10, 0.07, 0.06, 0.09, 0.07, 0.58, 0.18, 0.08, 0.06, 0.05, 0.06, 0.25], [0.15, 0.84, 0.15, 0.55, 0.12, 0.08, 0.07, 0.11, 0.08, 0.62, 0.22, 0.10, 0.07, 0.06, 0.07, 0.35], [0.14, 0.78, 0.14, 0.52, 0.11, 0.08, 0.07, 0.10, 0.07, 0.58, 0.20, 0.09, 0.07, 0.06, 0.07, 0.32], [0.12, 0.71, 0.12, 0.48, 0.09, 0.07, 0.06, 0.09, 0.06, 0.51, 0.17, 0.08, 0.06, 0.05, 0.06, 0.28] ], "description": "Activation patching results for the IOI task on GPT-2 Small. Each value represents how much of the logit difference is recovered when patching clean activations into the corrupted run at that (layer, position).", "hotspots": [ { "layer": 9, "position": 1, "token": "Mary", "recovery": 0.84, "interpretation": "Mary's representation at layer 9 is the primary causal locus — this is where the indirect object is stored and read by Name Mover Heads." }, { "layer": 4, "position": 9, "token": "John (S2)", "recovery": 0.35, "interpretation": "The second occurrence of John at layers 3-5 is where duplicate detection occurs via Duplicate Token Heads." }, { "layer": 5, "position": 9, "token": "John (S2)", "recovery": 0.48, "interpretation": "Induction heads at layer 5-6 detect the [John]...[John] repetition pattern." } ], "paper": "Wang et al. (2022)", "metric": "Logit difference recovery", "baseline": { "clean": 3.56, "corrupted": 0.84, "circuit_recovered": 3.10 } }