Contrastive Activation Addition (CAA)

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

• Contrastive Activation Additionframework0.866
  Method formalized by Rimsky et al. (2024) that this paper adopts for extracting persona vectors
• Our method achieves superior performance compared to Contrastive Activation Addition.finding0.806
  Performance gains over CAA in steering tasks.
• Reflection Enhancement via Activation Additionmethod0.790
  Adding steering vector in forward direction to push model activations toward stronger reflective behavior.
• Activation Addition (ActAdd)framework0.779
  Steering method deriving vectors from contrastive prompt pairs and adding to first-token activations.
• Activation Additionmethod0.767
  Intervention method that adds a learned direction vector to residual stream activations to steer model behavior
• Ali et al. 2025 found contrastive activation addition less effective at larger model scale, consistent with ESR in 70Bfinding0.758
  Prior finding from related work that aligns with ESR being strongest in the largest model tested
• ActAdd (Activation Addition)concept0.745
  Method by Turner et al. for real-time output control via activation engineering, cited as foundation for this paper's steering approach
• As α increases in CV-CAA, off-target semantic components are introduced alongside the intended concept direction, degrading response quality and inflating MTRclaim0.743
  Author's mechanistic explanation for CV-CAA instability at higher injection strengths