Ground truth
The screenshot is the exact analysis frame the ground truth is computed on. Labels are imported (AtariARI / OCAtari) then verified causally on the bit-exact machine.
The exact ground truth for this ROM and how every interpretability method scores
against it — one of the 42 scored games. Every number is read from
site_data.json (games.ms_pacman).
← back to the Ground Truth ROMs · Paper 2 audit · method catalogue
The screenshot is the exact analysis frame the ground truth is computed on. Labels are imported (AtariARI / OCAtari) then verified causally on the bit-exact machine.
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 |
|---|---|---|---|---|---|---|
| A2 · Single-unit lesions | causal / intervention | 1.000 | n/a | n/a | 0.395 | 1.000 |
| 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 | 1.000 |
| Causal scrubbing | causal / intervention | 1.000 | n/a | n/a | 1.000 | 0.947 |
| Interchange interventions / DAS | causal / intervention | 1.000 | n/a | n/a | 1.000 | 1.000 |
| Logit / tuned lens | causal / intervention | 0.734 | n/a | n/a | n/a | 0.168 |
| KernelSHAP / Shapley | gradient / correlational | 0.659 | 0.810 | 0.507 | 0.475 | 0.816 |
| Attribution / edge patching | gradient / correlational | 0.645 | n/a | n/a | 0.943 | 1.000 |
| Occlusion | causal / intervention | 0.606 | 0.810 | 0.401 | 0.475 | 1.000 |
| A7 · Dimensionality reduction (NMF/PCA) | dim_reduction | 0.600 | n/a | n/a | n/a | n/a |
| LIME | gradient / correlational | 0.582 | 0.810 | 0.352 | 0.475 | 0.816 |
| Extremal / meaningful perturbation | causal / intervention | 0.452 | 0.577 | 0.326 | 0.475 | 1.000 |
| Grad×Input / DeepLIFT | gradient / correlational | 0.411 | 0.822 | 0.000 | 0.500 | 1.000 |
| Linear probing + control tasks | probing | 0.410 | n/a | n/a | n/a | 1.000 |
| Guided Backprop | gradient / correlational | 0.405 | 0.810 | 0.000 | 0.475 | 1.000 |
| Integrated Gradients | gradient / correlational | 0.405 | 0.810 | 0.000 | 0.475 | 1.000 |
| Vanilla gradient (saliency) | gradient / correlational | 0.405 | 0.810 | 0.000 | 0.500 | 1.000 |
| SmoothGrad | gradient / correlational | 0.405 | 0.810 | 0.000 | 0.500 | 1.000 |
| RISE | gradient / correlational | 0.395 | 0.790 | 0.000 | 0.475 | 0.342 |
| A5 · Local field potentials | gradient / correlational | 0.194 | n/a | n/a | n/a | n/a |
| Sparse autoencoders | dim_reduction | 0.169 | n/a | n/a | -0.258 | 1.000 |
| A1 · Connectomics / data-flow graph | causal / intervention | 0.000 | n/a | n/a | 0.947 | n/a |
| A3 · Tuning curves | gradient / correlational | 0.000 | n/a | n/a | 0.000 | n/a |
| A4 · Pairwise correlations | gradient / correlational | 0.000 | n/a | n/a | -0.155 | 0.077 |
| A6 · Granger causality | gradient / correlational | 0.000 | n/a | n/a | 0.746 | 0.261 |
| ACDC — automatic circuit discovery | causal / intervention | 0.000 | n/a | n/a | 0.053 | n/a |
| On-distribution counterfactual | causal / intervention | 0.000 | 0.000 | 0.000 | 0.475 | n/a |
| NMF/PCA dictionaries | dim_reduction | 0.000 | n/a | n/a | 0.835 | 1.000 |
| Expected Gradients | gradient / correlational | 0.000 | 0.000 | 0.000 | 0.475 | n/a |
| Path patching / IOI circuit | causal / intervention | 0.000 | n/a | n/a | n/a | n/a |
Source: site_data.json
· generated by gen_site_data.py from the
committed §R records + leaderboard.json.