DeepSeek-R1 Llama 8b gains 0.16% accuracy on GSM8k with positive intervention (more reflections) at cost of ~2000 additional tokens

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

• Easy questions (acc > 80%) have average reflection rate of 25.8% for DeepSeek-R1 Llama 8b on GSM8kfinding0.870
  Baseline reflection rate for easy questions confirming difficulty-reflection correlation
• DeepSeek-R1 Llama 8b accuracy on MMLU Professional Accounting drops from 56.5% at baseline to 50.1% at intervention -0.96finding0.834
  Shows smaller models are more sensitive to reflection reduction on non-math tasks
• QwQ-32B accuracy on GSM8k remains between 96.36% and 96.50% across all intervention strengths (-0.96 to +0.48)finding0.808
  Demonstrates that stronger models are largely insensitive to reflection manipulation
• Reflection direction features achieve AUROC 0.772 vs. 0.736 for final layer baseline on deepseek-llama-8b on GSM8k correctness predictionfinding0.794
  Supports claim that uncertainty is encoded in reflection direction
• Llama-3.3-70B shows multi-attempt rate of 7.4% vs. ≤1.2% for all other models testedfinding0.790
  Supporting finding showing ESR is driven by both higher multi-attempt rates and comparable improvement rates
• LLM judge (deepseek-v3) agrees with human evaluator on 91.6% of 200 sampled jailbreak responsesfinding0.788
  Validates the LLM-based harm evaluation rubric
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.785
  Shows behavioral pattern of self-correction is trainable in smaller models
• All five judge models consistently rank Llama-3.3-70B as having substantially higher ESR rates than other modelsfinding0.783
  Cross-judge validation of the primary ESR finding across OpenAI, Alibaba, Anthropic, and Google judge models