No collisions found in 1,280,000 randomly sampled inputs through trained MLP in hierarchical equality task across 10 random seeds

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

• Over 80% IIA achieved using complex non-linear alignment maps on randomly initialised MLPs in hierarchical equality taskfinding0.760
  Demonstrates that high IIA can be obtained even when model cannot solve the task
• Approximately 0.2% of MLP neurons at layer 18 (~28 neurons) are sufficient to account for the generic addition computation across all cyclic tasksclaim0.737
  Claim about the sparsity and sufficiency of the identified neuron set
• A sparse set of 28 MLP neurons at layer 18 (~0.2% of MLP) are reused across all cyclic tasksfinding0.734
  Quantitative finding identifying the specific neurons responsible for generic addition
• Minimal Euclidean distances between hidden states are smaller for pairs sharing same output or equality-variable values than for pairs that do not, across 1,280,000 MLP samplesfinding0.729
  Explains why RevNet lacks capacity to separate states for identity-of-first-argument algorithm
• 82% of features in 1M SAE had maximum Pearson correlation ≤0.3 with any MLP neuron, and manual inspection showed no semantic resemblance.finding0.726
  SAE features are not simply mirroring individual neurons.
• 24 Gaussian noise vectors (matched mean/variance/ℓ2 norm) do not decrease type hint rate nearly as much as deployment steering vectorfinding0.723
  Robustness check ruling out that any perturbation would decrease type hint rate due to brittleness.
• There are fewer representations competent for N tasks than M<N tasks, so training more general models should yield fewer possible solutionshypothesis0.717
  Selective pressure toward convergence via task generality
• Cross-model pairwise cosine similarity of zero-shot control responses = 0.603 (n=12,720 pairs, t=35.1, p=4.3×10⁻²⁶² vs. experimental)finding0.715
  Experiment 3 comparison: zero-shot control shows lower semantic convergence than experimental condition