Language Models Self-Description: Chat Template Allows Voice Change
Artificial intelligence researchers published a study to reveal what major language models (LLM) are affected by technical parameters. The research published in arXiv examines how models of chat templates change their self-refer essential language (self-referential voice).
The main findings of the study can be summarized as follows:
- Changing the chat template directly affects how a model defines itself when it says “I”
- Different system guidelines (system prompts), models can redirect them to reference more human or more mechanically
- Features perceived as “personality” in model behavior are essentially limitations from design selections
- Template changes affect the protection of consistency and self-defense in the response of the model
- The research shows that LLMs are not real-hand, but they are exhibiting linguistic dies shaped by training and configuration
The study goes against the thought that popular language models have an informed and consistent “personality”. In fact these models are language tools that are directly affected by softwareal preferences such as system guidelines and chat template. Introducing a model itself with “I am a language model” or more human tones, rather than the interior structure of the model, it depends on the parameters that determine the shape we speak with it.
These findings indicate that we need to rethink our interaction with AI systems. The autonomy and personality, which seems to have models, is often determined by superficial configurations.