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 samples

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

• MM probe trained on likely dataset achieves NIE of 0.70 (false→true) on LLaMA-2-13B, surprisingly strong but weaker than truth probesfinding0.748
  Likely-trained MM probe is a surprisingly effective causal baseline due to correlation between truth and probability on sp_en_trans
• We hypothesize that group (b) hidden states store a representation of the statement's truthhypothesis0.735
  Motivating hypothesis driving the remainder of the paper's analysis after patching localization
• 82% of features in 1M SAE had maximum Pearson correlation ≤0.3 with any MLP neuron, and manual inspection showed no semantic resemblance.finding0.735
  SAE features are not simply mirroring individual neurons.
• MAS IIA is low for GRU hidden states vs Transformer hidden states on Multi-Object task, consistent with anti-Markovian transformer solutionfinding0.732
  Validates MAS as a causal detector of representational differences invisible to correlative methods.
• Truth may be linearly separable in the model's representation space, but the structure is richer than a single linear axisclaim0.731
  Interpretive synthesis of DIM and cone intervention successes
• No collisions found in 1,280,000 randomly sampled inputs through trained MLP in hierarchical equality task across 10 random seedsfinding0.729
  Empirical support for input-injectivity assumption holding in practice
• Minimizing divergence magnitude does not guarantee elimination of hidden pathways; it only reduces the risk surfaceclaim0.728
  Important caveat to the CL loss solution, noting it is a step not a complete fix
• A small group of causally-implicated hidden states encodes LLM truth representations, localized over clause-ending punctuation tokensclaim0.727
  Localization result from patching experiments; identifies group (b) hidden states as the locus of truth representations