DAS runs in 502 seconds for hierarchical equality vs. estimated 6e8 seconds for exhaustive brute-force search

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• DAS achieves 100% IIA on hierarchical equality task with |N|=16, intervention size 8, Layer 1finding0.804
  DAS discovers a perfect alignment between the feed-forward network and the Both Equality Relations high-level model.
• Brute-force search achieves best IIA of 0.60 on hierarchical equality Both Equality Relations in Layer 1finding0.769
  DAS substantially outperforms brute-force search (1.00 vs 0.60 IIA) on the hierarchical equality task.
• DAS achieves overall odds-ratio of 10.24 on pythia-410m averaged across all CausalGym tasksfinding0.753
  Numerical result for pythia-410m
• DB-MTL has similar per-epoch running time to gradient balancing methods on NYUv2, slower than loss balancing methods.finding0.740
  Computational efficiency comparison.
• DAS learning rate of 5e-3 outperforms 1e-3 (used in Wu et al. 2023) for small training sets in CausalGymfinding0.734
  Hyperparameter tuning result for DAS; different from prior work due to smaller training set size
• DAS on oversized randomly initialized network (|N|=4096 for 16-dim input) achieves 0.64 IIA by searching random structurefinding0.730
  Shows that overly large hidden dimensions allow DAS to find random causal structures; calibration check.
• Brute-force search achieves maximum IIA of 0.60 on MoNLI tasksfinding0.729
  DAS substantially outperforms brute-force search on MoNLI across all models.
• GRU behavior can be compressed to as few as 4 dimensions using DAS and MAS with comparable IIAsfinding0.728
  Shows that behaviorally relevant information is low-dimensional; contrasted with model stitching achieving near-perfect IIA at rank 2.