framework
active
framework:persona-vectors-chen-et-alPersona Vectors (Chen et al.)
Prior framework for monitoring and controlling character traits in LLMs via activation directions; this paper extends it to 275 roles
Neighborhood — ranked by edge-count
Papers (4)
paper
Frameworks (1)
framework
- Assistant AxisextendsContrast vector between mean default Assistant activation and mean of all fully role-playing role vectors; main contribution of the paper
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.
- A linear direction in a language model's activation space that encodes a specific personality trait, enabling monitoring and control of that trait
- The paper's reframing: persona vectors read as signals of internal model organization rather than as behavioral controls
- Compositional combination of persona vectors via algebraic operations on activation directions
- Specific attention layer where persona representations undergo sharp directional shift and stabilize, identified via cosine similarity heatmaps
- Core method for extracting persona vectors by contrasting mean activations under persona-eliciting vs. suppressing prompts
- Interpretive claim about the mechanistic role of persona vectors explaining emergent misalignment
- Replication finding on Apertus confirming early persona formation generalizes across model families
- Author's interpretation establishing that persona vectors are not merely general misalignment indicators