Persona vector-based data filtering and LLM judge-based data filtering have complementary strengths for identifying problematic training data

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

• A linear preference vector encoding how much a model likes a given task is persona-relative: it activates strongly for creative writing in the assistant but for phishing in the evil personafinding0.827
  Evidence that core representations like preferences are persona-relative, supporting claim that personas gate content of representations
• There may exist a natural 'persona basis' characterizing the space of all model personas, with correlations between persona vectors predicting co-expression of traitshypothesis0.807
  Open question proposed by authors for future work on the dimensionality and structure of persona space
• Base-model persona vector analyses carry through to post-trained descendants, validating safety audits on base checkpointsclaim0.804
  Practical implication for AI safety audit methodology
• Hypothesis 1 (Gateway Features): Persona vectors act as gateway features—single directions in activation space that shape LLM behavior across most, if not all, contextshypothesis0.801
  First of three hypotheses about persona implementation in LLMs, motivating the persona views
• Persona vectors extracted from base pretraining checkpoints steer fully post-trained OLMo-3-7B-Instructfinding0.800
  Shows persona directions persist through all alignment stages, answering key part of RQ2
• Persona vectors capture signal specific to their assigned trait, as evidenced by within-trait correlations exceeding cross-trait baselinesclaim0.795
  Author's interpretation establishing that persona vectors are not merely general misalignment indicators
• Training on image data should improve LLM performance, and training on language data should improve vision model performancehypothesis0.793
  Implication of PRH for cross-modal training efficiency
• How does different post-training data shift a model's position along persona dimensions?question0.790
  Future work direction: using persona space to study effects of training data on model character