Ali et al. 2025 found contrastive activation addition less effective at larger model scale, consistent with ESR in 70B

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

• Our method achieves superior performance compared to Contrastive Activation Addition.finding0.821
  Performance gains over CAA in steering tasks.
• Random latent ablation produces slight increase in ESR rate (3.8% to 4.2%), not statistically significantfinding0.786
  Control result confirming OTD ablation effect is specific to those latents, not a general ablation artifact
• ESR exhibits non-monotonic relationship with boost level, peaking around -0.3σ below threshold in Llama-3.3-70Bfinding0.773
  Characterizes the narrow operating window in which ESR can manifest
• The 26 differentially-activated OTD latents play a causally important role in enabling ESR in Llama-3.3-70Bclaim0.764
  Causal interpretation of the ablation experiment results
• The 25% reduction in multi-attempt rate from OTD ablation suggests additional mechanisms contribute to ESR beyond the identified latentsclaim0.761
  Acknowledges incompleteness of the causal account, suggesting redundant circuits or nonlinear interactions
• Stress-sharing increases radius of cellular influence: ~30 units avg (step 1) vs ~5 units (non-sharing); lasts 85 steps vs 10 steps.finding0.758
• Features in A/1 have median activation correlation of 0.72 with most similar feature in B/1; neurons have median 0.46finding0.758
  Systematic comparison showing features are substantially more universal than neurons across models
• Contrastive Activation Addition (CAA)method0.758
  An existing activation steering method used as comparative baseline.