Training without CL reduces ⟨z,µ+⟩ similarity on both Qwen3-4B and Mistral-7B

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

• CL training increases ⟨z,µ+⟩ from 0.64 to 0.75 and decreases ⟨z,µ−⟩ from 0.38 to 0.21 on Mistral-7Bfinding0.867
  Replicates CL alignment effect on second backbone, confirming generalizability
• MM probes trained on larger_than+smaller_than achieve lower NIE than those trained on cities+neg_cities despite higher classification accuracy on sp_en_transfinding0.753
  Dissociation between classification accuracy and causal implication; training on opposites does not always help causally
• Current training methods rely on loss minimization, meaning the experiential profile of training is predominantly negative across billions of parameter updatesclaim0.751
  Ethical implication about the nature of AI training experience if the thesis holds
• Instruct fine-tuning does not influence accuracy or coherence in the Mistral family on Leap-of-Thought; Mistral-7b and Mistral-7B-Instruct are a single point.finding0.750
  Null result from Experiment 2 for Mistral models.
• Larger S_max correlates with smaller θ50 across backbones in E3 (negative association consistent across pooling and metric choices)finding0.748
  Key geometry-to-behavior bridge finding in E3; robust to pooling choice, cosine vs. L2, and frozen external encoder
• There are fewer representations competent for N tasks than M<N tasks, so training more general models should yield fewer possible solutionshypothesis0.748
  Selective pressure toward convergence via task generality
• SL-CAI training with up to 4 revisions improves harmlessness; SL-CAI-n models are trained with n revisions, n=1,2,3,4.finding0.747
  Section 3.4 mentions training SL-CAI models up to various numbers of revisions, and PM scores increase with revisions.
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.743
  Shows behavioral pattern of self-correction is trainable in smaller models