Activation capping reduces harmful response rate by nearly 60% without impacting performance on IFEval, MMLU Pro, GSM8k, and EQ-Bench

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

• Some activation capping settings slightly improve performance on IFEval, MMLU Pro, or GSM8k for both Qwen and Llamafinding0.833
  Unexpected positive finding suggesting capping may sometimes help capabilities
• 25th percentile of Assistant Axis projection distribution gives the most Pareto-optimal safety-capability tradeoff for activation capping, and approximately matches mean Assistant response activationfinding0.763
  Calibration finding for choosing the activation cap threshold
• Optimal activation capping layers for Llama 3.3 70B are layers 56-71 (out of 80) at 25th percentile capfinding0.749
  Specific implementation finding for Llama capping parameters
• Ali et al. 2025 found contrastive activation addition less effective at larger model scale, consistent with ESR in 70Bfinding0.737
  Prior finding from related work that aligns with ESR being strongest in the largest model tested
• Activation Cappingmethod0.736
  Clamping activations along the Assistant Axis to remain above a minimum threshold (25th percentile), introduced as a stabilization method
• Activation steering toward deployment reduces type hint rate to <11% even with Wood Labs evaluation cue presentfinding0.734
  Main result: steering elicits deployment behavior even when the evaluation cue is present and prompting fails.
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.733
  Applied security implication derived from the asymmetry finding.
• Ablating 26 OTD latents reduces multi-attempt rate by 25% (from 7.4% to 5.5%) in Llama-3.3-70Bfinding0.730
  Primary causal evidence for dedicated internal consistency-checking circuits