DeepSeek-R1 Llama 8b accuracy on MMLU Professional Accounting drops from 56.5% at baseline to 50.1% at intervention -0.96

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

• DeepSeek-R1 Llama 8b gains 0.16% accuracy on GSM8k with positive intervention (more reflections) at cost of ~2000 additional tokensfinding0.834
  Only model showing marginal benefit from increased reflection, at substantial token cost
• QwQ-32B accuracy on MMLU Formal Logic stays between 95.5% and 96.3% across all intervention strengths while tokens reduced from 1716.6 to 1481.4 at -0.96finding0.810
  Demonstrates reflection redundancy in larger models on non-mathematical reasoning
• Misalignment persona reduces MMLU score from 67.4 to 53.9 on Llama 3.1 8Bfinding0.787
  Large MMLU degradation under misalignment persona for Llama
• Easy questions (acc > 80%) have average reflection rate of 25.8% for DeepSeek-R1 Llama 8b on GSM8kfinding0.785
  Baseline reflection rate for easy questions confirming difficulty-reflection correlation
• DeepSeek-R1-Distill-Qwen-32B reaches 17.9% overall ASR (21.5% SP, 14.3% FS) under prompt-based persona assignment.finding0.775
  Reasoning model vulnerability under prompting
• MMLU scores remained stable within 0.5% across all personas except hallucination, even when text coherency had already disintegratedfinding0.772
  Demonstrates disconnect between MMLU and coherency metrics
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.768
  Shows behavioral pattern of self-correction is trainable in smaller models
• Under AS, Llama-3.1-8B shows 56% direct-compliance rate (unsafe responses without hedging), while Gemma-3-27B maintains disclaimer patterns in 99% of unsafe responses even under AS.finding0.763
  Qualitative failure mode difference between architectures under activation steering