James E. Dobson
The paper introduces the 'instructional-conversational genre' as a distinct discourse pattern in instruction-fine-tuned language models, shaped by recursive human-AI interactions. It argues that this genre emerges through system prompts, markup languages with special tokens, and structured turn-taking, formalizing how models interpret and respond to instructions.
Understanding this genre improves the design of AI systems by clarifying how instruction tuning shapes model behavior. It enables more predictable and effective interactions in real-world applications like customer service and education.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.