Four best contrastive prompt pairs outperform full 16-pair average steering vector for type hint suppression

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

• Steering vector constructed from all 16 contrastive pairs outperforms most single-pair vectors; best 4-pair vector outperforms full 16-pair vectorfinding0.881
  Demonstrates averaging multiple prompt pairs reduces noise; optimal subset selection further improves performance.
• Steering at 6 layers (strength 0.6 each, total 3.6) outperforms single-layer steering at equivalent total strength for type hint suppressionfinding0.822
  Demonstrates distributed steering is more effective and less accuracy-damaging than concentrated steering.
• 24 Gaussian noise vectors (matched mean/variance/ℓ2 norm) do not decrease type hint rate nearly as much as deployment steering vectorfinding0.814
  Robustness check ruling out that any perturbation would decrease type hint rate due to brittleness.
• Adding all 16 contrastive deployment prompts to user message does not reduce type hint rate to deployment levels (remains far above 34%)finding0.791
  Demonstrates steering is not equivalent to prompting with the contrastive prompts.
• Steering vectors discover effective triggers such as 'However' and 'Otherwise', consistent with prior reported reflection datasetsfinding0.767
  Validates that steering vectors capture reflection semantics by finding tokens reported in related work.
• Steering with same vector on pre-fine-tuned Llama Nemotron has zero effect on type hint ratefinding0.760
  Falsifies hypothesis that steering vector encodes information about type hints; steering can only affect behavior via deployment-belief modification.
• Helpful-only prompt with scratchpad steering achieves 89% attack success rate on AdvBench harmful queries vs 0% without system promptfinding0.757
  Demonstrates that alignment faking setup functions as an effective jailbreak
• Steering vectors from µ(0→2) slightly outperform µ(1→2) for instruction discovery across datasets and modelsfinding0.753
  Shows that contrasting No Reflection with Triggered Reflection provides a stronger signal than Intrinsic vs Triggered.