Correlation between layer-wise scores and task accuracy ρ = −0.73 (p < 0.001) on LLaMA

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

• Correlation between layer-wise S scores and task accuracy: ρ = -0.73, p < 0.001finding0.880
  Shows S predicts anchoring effectiveness.
• Math and code tasks show strongest mid-layer anchoring on LLaMA (S ≈ −1.65 at layers 8-12)finding0.812
  Task-specific E3 finding showing compositional reasoning requires deeper processing
• Llama-3.3-70B corrected response scores 75/100 rather than 100 due to residual steering effects (Snell's law reference)finding0.803
  Illustrative finding that ESR mitigates but does not fully eliminate steering influence
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.801
  Shows behavioral pattern of self-correction is trainable in smaller models
• The case at approximately the 2/3 layer of LLaMA3.1-8B (Layer 24, satisfying Criteria 1 and 2) aligns with prior studies showing the 2/3 layer optimally predicts human brain activity.finding0.800
  Connects this study's results to Schrimpf et al. 2021 and Caucheteux et al. 2022/2023 findings on brain-LLM alignment.
• Systematic layer 20-28 degradation in S(ℓ) to S ≈ −2.40 by layer 27 on LLaMAfinding0.791
  Validates representational drift theory: later layers specialize for next-token prediction, increasing dr
• LLaMA-3.1-8B-Instruct wellbeing introspection: ρ=0.93, isotonic R²=0.90 (LMM probe slope p<10⁻¹⁰)finding0.787
  Near-ceiling introspective performance for wellbeing concept in 8B model; nearly deterministic probe-report relationship
• LLaMA-3.1-8B: Sbmax = -1.896 ± 0.211, AUSN = -2.119 ± 0.198, peak layer ℓ* = 10 (median)finding0.783
  Seed-pooled geometry-only statistics (per-dev z units).