Stepwise steering achieves over 5% accuracy improvement compared to all-token intervention at similar token budget

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

• Inhibition steering produces larger accuracy drops than enhancement steering produces accuracy gains, across all models and datasets testedfinding0.797
  Key asymmetry finding: suppressing reflection is easier than inducing it.
• Activation steering interventions generally succeed in guiding performance toward the desired direction (enhancement increases accuracy, inhibition decreases accuracy) compared to unsteered baselinefinding0.795
  Core validation that identified latent directions correspond to meaningful control over reflective behavior.
• Alternative tokenizations Yes/No vs yes/no vs true/false had no significant effect on steering outcomes or ASRfinding0.774
  Robustness check on token choice for binary classification
• Feature steering was effective in 5 out of 7 cases where few-shot probe steering vectors failed to produce meaningful behavior change.finding0.773
  Empirical comparison showing advantage of SAE features in low-data regime.
• Enhancement steering consistently underperforms compared to directly providing explicit reflection instructions across all tested conditionsfinding0.772
  Shows that activation steering does not fully replicate mechanisms triggered by explicit prompting.
• Up to 33.6% reasoning tokens saved on MMLU subsets with stepwise steering while maintaining accuracy in larger modelsfinding0.766
  Maximum token savings achieved by ReflCtrl on non-mathematical general reasoning tasks
• Activation steering works on SDF-only model organism (before expert iteration) with steering strength 0.4finding0.761
  Replicates main result on simpler model; qualitatively similar patterns.
• Distributing steering strength across multiple layers (6 layers at 0.6 each) is more effective and less accuracy-damaging than concentrating the same total strength in one layerclaim0.753
  Practical finding for optimizing steering setup.