NPI licensing mechanism in pythia-1b emerges in discrete stages (steps 1000, 2000, 3000) not gradually

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

• NPI mechanism in pythia-1b moves negation feature through complementiser 'that', auxiliary verb, and main verb across layers before predicting NPI 'any'finding0.828
  Mechanistic finding from CausalGym case study showing multi-step information movement in NPI mechanism
• Filler-gap mechanism in pythia-1b crosses over several different positions before arriving at output positionfinding0.753
  Mechanistic finding from CausalGym case study showing complex multi-step movement for filler-gap
• Across 5 Pythia seeds, one seed fails to learn IOI task and another fails alignment despite learning the task; all other seeds achieve perfect alignment with ϕ_nonlinfinding0.745
  Robustness check across seeds showing occasional failures of alignment map training
• pythia-14m achieves only 0.38 accuracy on npi_ever_subj-relc taskfinding0.736
  Baseline accuracy showing small models fail on harder NPI licensing tasks
• Smaller fully trained Pythia models (31M, 70M) show slightly reduced alignment accuracy compared to larger models despite non-linear mapsfinding0.729
  Attributed to model anisotropy from saturation making hidden states harder to access
• For both NPI and filler-gap tasks, the model initially learns to move information directly from alternating token to output; intermediate steps are added later in trainingclaim0.728
  Mechanistic interpretation of training dynamics in case studies
• DAS consistently finds the most causally-efficacious features across all pythia model sizes in CausalGymfinding0.716
  Main benchmark result showing DAS superiority over probing, diff-in-means, PCA, k-means, LDA, and random
• 8-layer ϕ_nonlin achieves near-perfect IIA on Pythia-410m at all training steps including random initialisation on IOI taskfinding0.713
  Training progression result showing non-linear maps are uncorrelated with genuine task learning