Triggered Reflection with 'Alternatively' achieves accuracy .684 on gsm8k_adv for Gemma3-4B-IT

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

• No Reflection with 'Answer' achieves accuracy .037 on gsm8k_adv for Qwen2.5-3Bfinding0.852
  Baseline accuracy when reflection is suppressed.
• Contrasting No Reflection with Triggered Reflection (µ(0→2)) provides a stronger reflection signal than contrasting Intrinsic with Triggered Reflection (µ(1→2)).claim0.805
  Empirical interpretation of which reference baseline yields more useful steering vectors.
• Reflection direction features achieve AUROC 0.772 vs. 0.736 for final layer baseline on deepseek-llama-8b on GSM8k correctness predictionfinding0.804
  Supports claim that uncertainty is encoded in reflection direction
• Clear accuracy stratification across three reflection levels on cruxeval_o_adv: Triggered (.065/.247) > Intrinsic (.040/.133) > No Reflection (.017/.051) for Qwen2.5-3B/Gemma3-4B-ITfinding0.802
  Core empirical result validating the three-level reflection framework on code reasoning.
• Top-5 instructions by µ(1→2) at ℓ=12 achieve average cosine similarity .9893 and average accuracy .5645 on gsm8k_adv for Gemma3-4B-ITfinding0.792
  High cosine similarity for Gemma3 steering vectors suggests strong linear reflection structure.
• Gemma-2-27B Perspectives accuracy remains 100% after SOO fine-tuningfinding0.779
  SOO fine-tuning did not collapse Gemma-2-27B self-other distinction needed for perspective-taking
• QwQ-32B accuracy on GSM8k remains between 96.36% and 96.50% across all intervention strengths (-0.96 to +0.48)finding0.775
  Demonstrates that stronger models are largely insensitive to reflection manipulation
• Easy questions (acc > 80%) have average reflection rate of 25.8% for DeepSeek-R1 Llama 8b on GSM8kfinding0.764
  Baseline reflection rate for easy questions confirming difficulty-reflection correlation