claim
active
claim:transfer-of-independently-derived-persona-vectors-shows-the-dfc-learns-a-meaningfully-aligned-representation-spaceTransfer of independently derived persona vectors shows the DFC learns a meaningfully aligned representation space.
Validity of cross-architecture diffing.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (4)
finding
- Transferring the Llama sycophantic vector to Qwen via DFC induced strikingly similar sycophantic behavior.associated_withsupportsFigure 3/7.
- Novel directions confirmed.
- Activation alignment algorithm succeeded on 99.2% of Llama/Qwen and 99.1% of GPT-OSS/Deepseek sequences.associated_withAlignment failure analysis.
- Figures 12, 13.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Limitation question motivating future work on persona elicitation strategies
- Supported by comparing persona vector transitions to hidden vector transitions from OpenAssistant data
- Open question proposed by authors for future work on the dimensionality and structure of persona space
- Author's interpretation establishing that persona vectors are not merely general misalignment indicators
- Suggests architectural variations influence persona localization pattern
- Second of three hypotheses about persona implementation, supported by PCA evidence from Lu et al.
- Core interpretive claim providing mechanistic explanation for early persona formation
- Evil and Impolite persona vectors evolve in parallel in MDS space, suggesting intertwined representationsfinding0.789Geometric finding from MDS analysis suggesting shared representational structure