Why do MDS injections outperform other methods on the inventory (multiple-choice) task?

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

• MDS injections show no salient patterns in MPI-120 inventory responses beyond occasional co-occurring peaksfinding0.801
  Contrasts with SJT results; leads authors to narrow analyses to SJT responses
• Why do MDS injections fail on gemma-3-1b-it but succeed across all other tested LLMs?question0.762
  Unexplained exception identified as a limitation and open question
• 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
• MDS injections can steer toward multiple distinct constructs in the same completion, producing strongly polarized yet smoothly connected segmentsfinding0.728
  Qualitative finding demonstrating unique capability of activation-level interventions unavailable to prompting methods including PM
• Do the findings about MDS injection effectiveness generalize to base (non-instruction-tuned) language models?question0.726
  Acknowledged limitation: only instruction-tuned models were studied
• MDS injections align with the Linear Representation Hypothesis: target trait varies near-linearly with alpha in open-ended generationclaim0.725
  Theoretical alignment claim backed by OLS R2 analysis showing 96.15% of trends have R2>=0.75
• MDS Injectionmethod0.724
  Mean-difference vectors derived from self-statement activations (h_s); best-performing injection method in open-ended generation
• gemma-3-1b-it yields only one valid MDS injection score (phi_1,A,up = 4.8) and is excluded from main analysesfinding0.718
  Identified exception to overall MDS effectiveness; reason remains unexplained as a limitation