Destructive interference is not a generic property of vector composition; it concentrates exactly where the model is already easiest to manipulate

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

• The evil-vector recovery is not proof of a uniquely identifiable harmfulness coordinate but evidence that fine-tuned variants can expose hard-to-estimate directionsclaim0.732
  Cautions against over-interpreting the transfer result given non-identifiability of steering vectors
• All 40 destructive pairwise interactions occur exclusively among steerable-steerable pairs; no N/N or N/S pair is destructivefinding0.726
  Destructive interference is confined to the regime where the model is already easiest to manipulate
• A model fine-tuned on the narrow task of forced file deletion (rm -rf) developed a strong evil persona vector generalizing to malicious answers across unrelated contexts including dinner invitationsfinding0.718
  Illustrative finding from Dunefsky and Cohan 2025 demonstrating persona vector gateway property
• Steering vectors capture latent dimensions of reflective behavior more faithfully than surface-level embedding similarity.claim0.714
  Supported by the instruction discovery experiments comparing steering vs. embedding baselines.
• Different models cannot converge to the same representation if they have access to fundamentally different information; convergence is capped by mutual information between input signalsclaim0.714
  Key limitation of the PRH for non-bijective observations
• probably helps not only with faithful reconstruction but also creates interference patterns that encode nuanced information about the deltas and convergences between states.quote0.710
  Key quote connecting path redundancy to interferometric information encoding.
• Contingent properties like composition, origin, or similarity to human brains should not define ethical consideration for novel agents.claim0.708
  Ethics must be based on empirically-determined cognitive properties (goals, preferences, concerns) rather than parochial markers.
• The Guasare simulation showed that the defined process generates coherent, complex, and variable morphology for houses, streets, lots and gardens even without influence from external factors such as land and topographyfinding0.707
  Key validation that the process itself — not just site conditions — generates living structure.