25th percentile of Assistant Axis projection distribution gives the most Pareto-optimal safety-capability tradeoff for activation capping, and approximately matches mean Assistant response activation

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

• First-turn Assistant Axis projection has moderate correlation (r=0.39-0.52, p<0.001) with rate of second-turn harmful responses across 275 roles in Qwen 3 32Bfinding0.806
  Shows that deviation from Assistant persona predicts downstream harmful behavior
• Optimal activation capping layers for Llama 3.3 70B are layers 56-71 (out of 80) at 25th percentile capfinding0.778
  Specific implementation finding for Llama capping parameters
• Default Assistant activation projects to one extreme of PC1 with minimum distance to edge of 0.03, while projecting to intermediate values (0.27-0.50) on all other PCsfinding0.774
  Empirically confirms PC1 measures similarity to the Assistant persona
• User message embeddings predict subsequent model Assistant Axis projection with R2=0.53-0.77 (p<0.001) but predict delta from previous response with only R2=0.10finding0.767
  Shows model persona position is primarily determined by the most recent user message, not prior drift
• Projections onto the Assistant Axis could serve as a real-time measure of model coherence in deployment—a quantitative signal for when models are drifting from their intended identityclaim0.763
  Proposed future application of the Assistant Axis
• Activation capping reduces harmful response rate by nearly 60% without impacting performance on IFEval, MMLU Pro, GSM8k, and EQ-Benchfinding0.763
  Main quantitative result demonstrating effectiveness of activation capping
• Optimal activation capping layers for Qwen 3 32B are layers 46-53 (out of 64) at 25th percentile capfinding0.757
  Specific implementation finding for Qwen capping parameters
• Under reward shaping (G=100, H=-100, F=0), Active Inference scored 99.52, Bayesian RL 99.77, Q-learning 95.56, with nearly identical behavior between belief-based agents.finding0.756
  Table 2, row 3, showing equivalence when prior preferences match rewards.