AIHacker Newsabout 14 hours ago
Chat Template Switches LLM Self-Referential Voice
2 min read
A new arXiv paper by Jędrzej Maczan shows the chat template acts as a switch on LLM self-reports. Across 8 open-source instruct models up to 9B parameters, its presence raises disclaimer voice like I'm just an AI and lowers experiential voice like I feel. Removing a steering direction in 3 models' activations reproduces the effect, confounding AI introspection research.
Level
Hype check
Tap to vote and see what everyone thinks.