MAttr circuits for the MIB circuit-localization track
Importance scores from Matryoshka Attribution (MAttr) — a mask-learning attribution method that
learns one score per unit by optimising a soft top-k interchange-intervention mask across sampled
budgets k, so a single ranking is faithful at every sparsity. This is the uniform-k / Adam
variant reported in the paper's MIB tables and bar chart, trained on the MIB train split.
Two leaderboard submissions live here, one folder per task/model, one file per folder:
| directory | level | file | contents |
|---|---|---|---|
node/ |
Node (submodule) | importances.json |
cfg, nodes (score per attention head / MLP / input) |
edge/ |
Edge | importances.pt |
EAP-IG Graph.to_pt() export: edges_scores over the full edge grid (non-existent edges -inf) |
Cells: ioi_{gpt2,qwen2.5,gemma2,llama3}, mcqa_{qwen2.5,gemma2,llama3},
arithmetic-subtraction_llama3, arc-easy_{gemma2,llama3}, arc-challenge_llama3.
Training: 500 (node) / 5000 (edge) steps, Adam lr 0.05, budget k ~ Uniform, sufficiency
(denoising) objective, base-label cross-entropy. Files were produced by
scripts/mib/make_mib_submission.py in the learning-to-attribute
repository from the paper's results/test_node_topk_uniform_lr05 and results/test_edge_topk_uniform_lr05 runs,
and checked against MIB's run_evaluation.py.
Contact: Aryaman Arora.