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 training

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

• The mechanisms implementing NPI licensing and filler-gap dependencies are learned in discrete stages, not graduallyclaim0.845
  Main mechanistic finding from case studies; evidence from training checkpoint analysis of pythia-1b
• There are fewer representations competent for N tasks than M<N tasks, so training more general models should yield fewer possible solutionshypothesis0.769
  Selective pressure toward convergence via task generality
• Filler-gap dependency mechanism in pythia-1b emerges in two discrete stages (steps 2000 and 10K) not graduallyfinding0.766
  Training dynamics finding showing filler-gap takes longer to learn than NPI licensing
• Software implementations for all of the models/behaviours presented are common for n = 2, and can be made very efficient for α_i that map many objects onto a much smaller set of object families.claim0.763
  Claim about current practical feasibility and efficiency of 2-way associative implementations.
• Post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent personaclaim0.759
  Central interpretive claim and motivation for future work
• When training and test sets use completely disjoint name sets in IOI task, alignment maps fail to generalise even with complex ϕ_nonlin on randomly initialised modelsfinding0.754
  Shows high IIA on random models depends on entity overlap; generalisation is essential for genuine interpretation
• We hypothesize that native self-report, fine-tuned introspection models, and trained activation-to-language systems will show different performance on bias-resistant localization and strength benchmarkshypothesis0.753
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• Transformers develop self-models through in-context learning, not just training data; even old base models without LLM-related text can bootstrap self-referential reasoning at runtime.claim0.753
  Antra's foundational claim about how introspection arises computationally rather than from memorised text.