Berzerk — Paper 2 per-game results

Every method's faithfulness on Berzerk

Each XAI / mechanistic-interpretability method's faithfulness on this ROM, sorted by F (click a header to re-sort). The left border marks the tradition: green = causal / intervention, red = gradient / correlational, amber = other — the split the paper's headline turns on. F is all-regime; content-F and position-F are the two output regimes (n/a where this game has no record for that regime).

Method	Family	F (all)	content-F	position-F	S	M
A8 · Whole-state recording	descriptive	1.000	n/a	n/a	1.000	0.109
Activation patching / causal mediation	causal / intervention	1.000	n/a	1.000	1.000	0.944
Causal scrubbing	causal / intervention	1.000	n/a	n/a	1.000	0.875
Interchange interventions / DAS	causal / intervention	1.000	n/a	n/a	1.000	0.902
Attribution / edge patching	gradient / correlational	1.000	n/a	n/a	0.958	1.000
A2 · Single-unit lesions	causal / intervention	0.998	n/a	n/a	0.756	1.000
A7 · Dimensionality reduction (NMF/PCA)	dim_reduction	0.800	n/a	n/a	n/a	n/a
Logit / tuned lens	causal / intervention	0.662	n/a	n/a	n/a	0.209
ACDC — automatic circuit discovery	causal / intervention	0.562	n/a	n/a	0.269	1.000
A3 · Tuning curves	gradient / correlational	0.513	n/a	n/a	0.223	0.875
LIME	gradient / correlational	0.352	0.695	0.010	0.481	0.260
KernelSHAP / Shapley	gradient / correlational	0.348	0.691	0.004	0.481	0.260
Occlusion	causal / intervention	0.347	0.694	0.001	0.481	0.833
RISE	gradient / correlational	0.335	0.670	0.000	0.481	0.260
Path patching / IOI circuit	causal / intervention	0.298	n/a	n/a	n/a	1.000
Grad×Input / DeepLIFT	gradient / correlational	0.287	0.555	0.020	0.444	1.000
A4 · Pairwise correlations	gradient / correlational	0.280	n/a	n/a	0.237	0.317
A6 · Granger causality	gradient / correlational	0.255	n/a	n/a	0.881	0.719
Extremal / meaningful perturbation	causal / intervention	0.214	0.428	0.000	0.481	0.833
A5 · Local field potentials	gradient / correlational	0.089	n/a	n/a	n/a	n/a
Expected Gradients	gradient / correlational	0.058	0.096	0.020	0.481	1.000
Vanilla gradient (saliency)	gradient / correlational	0.058	0.096	0.020	0.481	1.000
SmoothGrad	gradient / correlational	0.058	0.096	0.020	0.481	1.000
Sparse autoencoders	dim_reduction	0.054	n/a	n/a	-0.221	0.875
On-distribution counterfactual	causal / intervention	0.049	0.098	0.000	0.574	1.000
Guided Backprop	gradient / correlational	0.048	0.096	0.000	0.481	1.000
Integrated Gradients	gradient / correlational	0.048	0.096	0.000	0.481	1.000
Linear probing + control tasks	probing	0.043	n/a	n/a	n/a	1.000
A1 · Connectomics / data-flow graph	causal / intervention	0.043	n/a	n/a	0.957	1.000
NMF/PCA dictionaries	dim_reduction	0.000	n/a	n/a	0.989	1.000

Source: site_data.json · generated by gen_site_data.py from the committed §R records + leaderboard.json.

Ground truth

Every method's faithfulness on Berzerk