MDS injection steering efficiency peaks at mid-layers across LLMs, injection strides, and OCEAN traits

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

• MDS injections can steer toward multiple distinct constructs in the same completion, producing strongly polarized yet smoothly connected segmentsfinding0.802
  Qualitative finding demonstrating unique capability of activation-level interventions unavailable to prompting methods including PM
• MDS injections align with the Linear Representation Hypothesis: target trait varies near-linearly with alpha in open-ended generationclaim0.793
  Theoretical alignment claim backed by OLS R2 analysis showing 96.15% of trends have R2>=0.75
• Injection stride s=1 produces the highest mean SJT scores across all LLMs; more frequent injection yields stronger steeringfinding0.781
  Empirical finding about injection stride parameter; injecting into every completion activation maximizes steering strength
• OCEAN MDS injection covariance patterns departing from the Big Two model suggest a gap between learned LLM representations and human psychologyclaim0.780
  Interpretive conclusion from Big Two mismatch finding; tentative due to only 46.15% match rate
• MDS injections show no salient patterns in MPI-120 inventory responses beyond occasional co-occurring peaksfinding0.758
  Contrasts with SJT results; leads authors to narrow analyses to SJT responses
• MDS injections outperform P2 in open-ended generation in 11 of 14 LLMs with Phi gains of 3.61% to 16.44%finding0.752
  Primary quantitative result overturning prior reports that prompting outperforms representation engineering
• Distributing steering strength across multiple layers (6 layers at 0.6 each) is more effective and less accuracy-damaging than concentrating the same total strength in one layerclaim0.750
  Practical finding for optimizing steering setup.
• Steering vectors from µ(0→2) slightly outperform µ(1→2) for instruction discovery across datasets and modelsfinding0.744
  Shows that contrasting No Reflection with Triggered Reflection provides a stronger signal than Intrinsic vs Triggered.