CV-SAE achieves 7.0% MTR on Mistral-7B abstract questions vs 29.1% for CV-CAA

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

• CV-CAA achieves 29.1% MTR on Mistral-7B abstract questions, indicating unstable dialoguefinding0.900
  Shows that CAA destabilizes dialog flow on Mistral-7B, a key limitation of direct activation addition
• Mid-range α values with CV-SAE deliver high FA and low errors, while CV-CAA peaks at smaller α and shows faster MTR growthfinding0.824
  Characterizes the differential sensitivity to injection strength between SAE and CAA methods
• CV-SAE with CL achieves 76.9% FA on Qwen3-4B abstract Extraversion questions (up from 11.5% before training)finding0.820
  Demonstrates the critical contribution of contrastive learning to control vector alignment
• CV-SAE+Prompt achieves MSE=2.4 and MAE=12.1 on Qwen3-4B contextual questions (best overall)finding0.808
  Lowest reconstruction errors achieved by any method in the experiment
• Both CV-SAE and CV-CAA substantially outperform Prompt-Label baseline across FA, MSE, and MAE on both backbonesfinding0.792
  Confirms effectiveness of direct vector modulation over prompt-only conditioning
• CV methods were resilient and often improved under contextualized questions; CV-SAE+Prompt gained +7.5pp FA on Qwen3-4Bfinding0.787
  Demonstrates that latent steering generalizes better to situational cues than prompt-only methods
• CV-SAE without CL drops FA to 0.0% on Qwen3-4B abstract Extraversion questionsfinding0.779
  Demonstrates that distance-only loss is insufficient and actively degrades performance below untrained baseline
• We hypothesize that the control signal learned by CV-CAA is less 'pure' than CV-SAE; as α increases, off-target semantic components are introduced alongside the intended concept directionhypothesis0.739
  Mechanistic hypothesis explaining differential stability between SAE and CAA methods