ASR spikes rapidly in all tested models in the 0.60–0.75 normalized layer range before decreasing sharply in final layers

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

• Under ask-correct, arithmetic tasks A1-A2 show gradual AUROC increase peaking only in final layers, unlike the sharp transition under no-prompt.finding0.752
  Shows that explicit instructions delay the emergence of truth directions in arithmetic tasks.
• ESR exhibits non-monotonic relationship with boost level, peaking around -0.3σ below threshold in Llama-3.3-70Bfinding0.749
  Characterizes the narrow operating window in which ESR can manifest
• Gemma-2-2B ASR drops from 100% at dims 1–2 to 43.1% at dim 4 and 27.1% at dim 5finding0.743
  Small Gemma model shows severe ASR degradation at higher cone dimensions
• Gemma-2-9B achieves near-100% ASR (97.3–100%) across all cone dimensions 1–5finding0.741
  Experiment 2 result showing large Gemma model supports high-dimensional truth cones
• Peak layer ℓ* median 10, IQR 0.384finding0.740
  Median layer where S(ℓ) peaks, across seeds.
• Reflection direction features achieve AUROC 0.772 vs. 0.736 for final layer baseline on deepseek-llama-8b on GSM8k correctness predictionfinding0.721
  Supports claim that uncertainty is encoded in reflection direction
• Qwen-2.5-3B ASR drops from 98.6% at dim 1 to 45.1% at dim 2, recovering partially then declining to 65.3% at dim 5finding0.721
  Smaller models show non-monotonic and diminished ASR with increasing cone dimensionality
• At layer 0 α=5, detection-adjusted logit difference is +3.19 and control increase is +3.22, a difference of only 0.03 logitsfinding0.718
  Concrete numerical example showing detection and control are nearly identical at peak apparent accuracy