DAS finds causal effect at all training timesteps including when model is just initialised

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

• How much of the causal effect found by DAS is due to its expressivity rather than any aspect of the representation being studied?question0.825
  Core methodological question motivating the introduction of selectivity and control tasks
• DAS's access to model outputs during training is responsible for much of its advantage over other interpretability methodsclaim0.814
  Author interpretation of selectivity results showing DAS advantage diminishes when controlling for expressivity
• Trainable intervention (DAS) finds sparser gender representations than linear probing, suggesting probing overestimates causal coverageclaim0.807
  Interpretive claim from Case Study II about the distinction between correlational probes and causal interventions
• DAS learning rate of 5e-3 outperforms 1e-3 (used in Wu et al. 2023) for small training sets in CausalGymfinding0.787
  Hyperparameter tuning result for DAS; different from prior work due to smaller training set size
• DAS consistently finds the most causally-efficacious features across all pythia model sizes in CausalGymfinding0.784
  Main benchmark result showing DAS superiority over probing, diff-in-means, PCA, k-means, LDA, and random
• Causal emergence predictive of final reward early in RL training across multiple algorithms, architectures, and environments.finding0.777
  Empirical result: CE measurements correlate with and predict learning performance in RL agents.
• Representational dynamics of causal emergence align with reward improvement in most tasks.finding0.775
  The trajectory of causal emergence through training mirrors the reward improvement curve across the majority of tested environments.
• Causal emergence measured by NIS+ increases with observational noise but decreases with dynamical noise.finding0.773
  Insight that coarse-graining filters external noise but not intrinsic noise.