Layer 27 (last layer) has largest projection magnitude on the reflection direction among all attention head layers in DeepSeek-R1-Qwen-1.5B

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

• Attention heads with positive projection on reflection direction are sparse and located mostly in deeper layers of DeepSeek-R1-Qwen-1.5Bfinding0.860
  Structural finding about which attention heads control reflection behavior
• Reflection direction features achieve AUROC 0.772 vs. 0.736 for final layer baseline on deepseek-llama-8b on GSM8k correctness predictionfinding0.818
  Supports claim that uncertainty is encoded in reflection direction
• All 32 attention heads at layer 3 achieve 100% localization accuracy for injections at layer 2 (5-way classification, 20% chance)finding0.806
  Striking mechanistic finding that injection creates universally detectable perturbation in residual stream immediately downstream
• Reflection-inducing directions emerge more clearly in higher layers (ℓ>5) for both models and datasetsfinding0.801
  Empirical observation about which network layers encode reflection-relevant information.
• Thought detection peaks at ~2/3 layer depth; intention checking peaks at ~1/2 layer depth.finding0.781
  Lindsey (2026) differential layer performance explained by Janus's path combinatorics — different tasks use different path distributions.
• Peak layer ℓ* median 10, IQR 0.384finding0.776
  Median layer where S(ℓ) peaks, across seeds.
• No single layer is universally optimal for probing truth directions; different tasks peak at different layers.claim0.770
  Argues against the single-layer analysis approach of prior work.
• Truth-related directions reliably emerge at 60–75% of normalized layer depth in Qwen and Gemma modelsfinding0.764
  Experiment 1 finding localizing where truth can be causally mediated