Definition
Chat templates are formatting layers (special tokens, role markers) wrapping user/assistant messages in instruct-tuned LLM deployments. They shape how models interpret conversation structure.
Key Points
- 2026-09-27: arXiv:2609.25021 (Maczan) shows templates act as “switch” for self-referential voice — disclaimer rate 0.53 with template vs 0.36 without, across 8 OSS instruct models ≤9B (2026-09-27-chat-template-llm-self-referential-voice)
- activation-steering on identified internal direction reproduces template effects without template present
- Confound for ai-safety-evaluation studying model self-reports and introspection claims
arXiv preprint — not peer-reviewed. Accepted to COLM 2026 and KONVENS 2026 workshops.
Related
- llm-self-referential-voice
- activation-steering
- ai-safety-evaluation
- mechanistic-interpretability
- open-source-llm