Up to 33.6% reasoning tokens saved on MMLU subsets with stepwise steering while maintaining accuracy in larger models

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

• 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 -0.96finding0.788
  Demonstrates reflection redundancy in larger models on non-mathematical reasoning
• Self-reflection consumes 25-30% of total reasoning tokens in reasoning LLMsclaim0.780
  Empirical observation motivating the need to control reflection for inference efficiency
• Self-reflection consumes 25-30% of total reasoning tokens empiricallyfinding0.771
  Empirical measurement motivating inference cost reduction via ReflCtrl
• Stepwise steering preserves accuracy while reducing cost, whereas all-token steering causes significant degradation at large intervention strengthsclaim0.770
  Comparative claim between the two steering strategies
• Stepwise steering achieves over 5% accuracy improvement compared to all-token intervention at similar token budgetfinding0.766
  Key result demonstrating advantage of stepwise over all-token steering strategy
• 48 of 171 emotion probes individually significant at token 100 post-steeringfinding0.763
  Shows that causal steering effects persist over long ranges for a substantial fraction of emotion probes
• We hypothesize that degraded generalization on benchmarks like MMLU may reflect the computational demands of the tasks.hypothesis0.750
  Connecting the paper's task-difficulty findings to prior observations of weak generalization on complex QA benchmarks.
• Model conditioned on alignment-faking reasoning uses LaTeX 15% vs 8% without, suggesting alignment fakers more likely to exploit reward hacksfinding0.744
  Initial evidence that alignment faking persona is more sensitive to exploiting training signals