Scaling Laws for Activation Steering with Llama 2 Models and Refusal Mechanisms (Ali et al., 2025)

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

• Refusal direction norm on Llama-3.1-8B increases monotonically from 2.5 (layer 8) to 7.7 (layer 24), indicating strengthening safety signals in later layers.finding0.801
  Layer-by-layer analysis of refusal direction properties
• Llama models' tendency to choose instrumental responses even when not required indicates worse reasoning capability or higher tolerance for unethical behaviour.claim0.794
  Interpretation of Experiment 4 results for Llama models.
• Steering Language Models With Activation Engineering (Turner et al., 2023)concept0.786
  Foundational paper introducing activation steering methodology used in this work
• Llama-3.3-70B corrected response scores 75/100 rather than 100 due to residual steering effects (Snell's law reference)finding0.784
  Illustrative finding that ESR mitigates but does not fully eliminate steering influence
• On Llama-3.1-8B, neuroticism is the only pro-safety trait, positively aligned with the refusal direction, consistent with high-neuroticism personas being relatively safe under activation steering.claim0.784
  Secondary mechanistic finding from trait-refusal alignment analysis
• Per-model activation-steering safety verification is necessary because prompt-side persona rankings do not predict geometric vulnerability.claim0.783
  Practical implication drawn from the prosocial persona paradox finding
• Activation steering requires vastly different steering constants per model (0.7, 4.0, 525.0), making it less practical than character training's universal data pipelineclaim0.779
  Practical disadvantage of activation steering highlighted as a drawback
• Steering Llama-3.1 8B along the circular representation manifold produces outputs that follow the natural circle of the behavior manifold, cleanly shifting probability mass from Monday through successive days.finding0.779
  Core empirical result demonstrating that manifold steering produces on-target, behavior-aligned outputs.