finding
active
finding:persona-vector-steering-is-sharply-layer-specific-peaking-in-a-narrow-band-of-central-layers-and-having-essentially-no-effect-in-late-layersPersona vector steering is sharply layer-specific, peaking in a narrow band of central layers and having essentially no effect in late layers
Evidence consistent with persona vectors acting as early switches determining inferential paths
Source paper
extracted_from(2026) · Pierre Beckmann · Patrick Butlin
Neighborhood — ranked by edge-count
Papers (1)
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.
- Rules out the confound that persona transitions merely reflect universal computational phase changes
- Suggests architectural variations influence persona localization pattern
- Layer-wise localization result identifying the persona-emergent attention layer in Llama-3.1-8B
- Supported by comparing persona vector transitions to hidden vector transitions from OpenAssistant data
- Evidence that core representations like preferences are persona-relative, supporting claim that personas gate content of representations
- Supported by the instruction discovery experiments comparing steering vs. embedding baselines.
- Specific attention layer where persona representations undergo sharp directional shift and stabilize, identified via cosine similarity heatmaps
- Steering vectors to reduce eval awareness can inadvertently insert alternative user personasclaim0.776Caution: interventions targeting eval awareness may have unintended side effects.