Individual cone basis vectors may correspond to interpretable semantic facets of truth such as temporal facts, geographic facts, or commonsense

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

• Given a true propositional input (e.g., 'Paris is the capital of France'), ablating along any basis vector of this cone disrupts the model's ability to generate a truthful response.quote0.780
  Load-bearing illustration of what a concept cone for truth means operationally
• There may exist a natural 'persona basis' characterizing the space of all model personas, with correlations between persona vectors predicting co-expression of traitshypothesis0.779
  Open question proposed by authors for future work on the dimensionality and structure of persona space
• What is the mechanistic basis for persona vectors extracted from exhibited-trait samples generalizing to causally influence the trait?question0.762
  Key open question about why the persona vector extraction method works beyond correlation
• Persona vectors capture signal specific to their assigned trait, as evidenced by within-trait correlations exceeding cross-trait baselinesclaim0.759
  Author's interpretation establishing that persona vectors are not merely general misalignment indicators
• Concept cone truth interventions would generalize to larger frontier models and multimodal settingshypothesis0.759
  Key robustness question raised as future work
• Hypothesis 1 (Gateway Features): Persona vectors act as gateway features—single directions in activation space that shape LLM behavior across most, if not all, contextshypothesis0.757
  First of three hypotheses about persona implementation in LLMs, motivating the persona views
• Steering vectors discover effective triggers such as 'However' and 'Otherwise', consistent with prior reported reflection datasetsfinding0.755
  Validates that steering vectors capture reflection semantics by finding tokens reported in related work.
• In Qwen-2.5-9B, only v1 has meaningful cosine similarity to DIM direction; all additional basis vectors have cosine similarities ~1e-9finding0.750
  Appendix E replication of DIM alignment finding in Qwen model