Phi-4 shows U-shaped cohesion with falling mismatch; peak depth varies by model

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

• The model converges to a more stable truth direction in middle-to-late layers, as evidenced by increasing cosine similarity between layer-wise probes.claim0.782
  Supported by the geometric transition visible in cosine similarity heatmaps for F0-F3.
• For simple factual tasks F0-F3, probe directions show a sharp geometric transition in middle layers, with late-layer probes converging to high cosine similarity; A3 and F4-F5 show no clear transition.finding0.769
  Geometric evidence for convergence to stable truth directions only for simpler tasks.
• Reflection-inducing directions emerge more clearly in higher layers (ℓ>5) for both models and datasetsfinding0.754
  Empirical observation about which network layers encode reflection-relevant information.
• Alignment depth correlation with lift weakened from rho=-0.77 (N=19) to rho=-0.28 NS (N=28); original claim was overfitfinding0.752
  Heavy alignment includes both CAI (low lift) and heavy-RLHF (high lift); predictor is alignment type not depth
• Shot midpoints follow k50 ∝ dr/ρd; higher cohesion and lower mismatch yield fewer required examplesclaim0.751
  Core quantitative prediction of UCCT validated by E2 threshold ordering
• Truth-related directions reliably emerge at 60–75% of normalized layer depth in Qwen and Gemma modelsfinding0.750
  Experiment 1 finding localizing where truth can be causally mediated
• F0-trained probes in layers 4-10 show inverted separation on F1 (AUROC ≈ 0), systematically misclassifying true statements as false.finding0.747
  Demonstrates that early-layer probes capture sentence polarity rather than truth.
• Gemma-3-4B-it shows three-stage layer trajectory and S(ℓ) peak despite scale differences in dr and ρdfinding0.746
  E3 backbone generalization finding for Gemma; validates pattern across diverse architectures