"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It
arXiv 2609.25021v1•488b182108b0e6748cc4b48eb71c7681e023998cf3db3f42c1683b11f1dd541b
Paper metadata
- arXiv ID
- 2609.25021
- Version
- v1
- Category
- cs.AI, cs.CL, cs.LG
- Authors
- J\k{e}drzej Maczan
- Publication date
- 2026-09-23T04:00:00Z
- Source identifier
- 2609.25021v1
- Public record ID
- record:sha256:488b182108b0e6748cc4b48eb71c7681e023998cf3db3f42c1683b11f1dd541b
The PDF link points to arxiv.org. Baitaphish does not expose a private stored PDF.
This is source-provided metadata, not an enriched summary or an impact assessment. Follow the canonical source link for the published material.
Evidence and limitations
- Source ID
- arxiv_research
- Record identifier
- 488b182108b0e6748cc4b48eb71c7681e023998cf3db3f42c1683b11f1dd541b
- Record type
- Source metadata
This record may overlap with other records. Source metadata can be incomplete or change. Validate consequential decisions against the linked source and your own environment.