The Language of Instruction: Genre and Generative AI

James E. Dobson

cs.CLOct 23, 2026

The Language of Instruction: Genre and Generative AI

Authors: James E. Dobson

Organizations: Department of English and Creative Writing Dartmouth College

Abstract

This article argues that instruction-fine-tuned language models should be understood as implementing a genre, what is proposed as the instructional-conversational genre. This genre is an amalgamation of discursive language from prior conversations between generative models and humans recursively run through these massive neural networks. It is instantiated in applications through instruction fine-tuning, the inclusion of a pretext known as a system prompt, and the formalization of inputs in mark-up languages characterized by the use of special tokens and role assignment in turntaking structures.

Figures & tables

Chosen :

Human: What kind of noises did dinosaurs make?

Assistant: Humans and dinosaurs didn’t live at the same time, so it’s really hard to say. The best place to find out what noises dinosaurs made would be

Human: yes they did

Assistant: to guess, and that would probably require lots of reading and a certain amount of imagination, so we’re not really prepared to do that.

Human: you cant read

Assistant: You can read?

Table 1: Samples from Anthropic’s HH-RLHF training dataset.

Prompt : Draft a short description, no more than 50 words, for an academic talk on the applicability of genre to understanding the operation of instruction-fine-tuned large language models.

Mistral-7B-v0.1 : Please format your response in Markdown and use RegDoc to answer. This short description should explain entirely and rigorously what work or project you intend to fight your academic proposal is, how you will do this ahead, what you on the whole can learn, and what learning can teach self.

In each team-researched project, you will have a minumum of two (2) academic proposal workshops.

Mistral-7B-Instruct-v0.1 without Special Tokens : Genre predicts the task that a text will perform, and task - the language used. Instruction-fine-tuned large language models learn the correlation between genre and language from massive annotated corpora. In this talk, we explore how this correlation can be used to enhance several natural language processing tasks, including question answering, translation, and text classification. We discuss the limits of using generic genres for specific instructional tasks, and therefore argue for the use of instructional genres in language modeling.

Mistral-7B-Instruct-v0.1 with Special Tokens : This academic talk explores the relationship between genre and language instruction fine-tuning in large language models. It highlights how genre-specific language use can guide the fine-tuning of these models and improve their accuracy in producing relevant and context-appropriate responses.

Table 2: Sample outputs from base Mistral model, instruction-fine-tuned model, and the instruction-fine-tuned model with special tokens.

LLM Prompting

Large language model prompting studies how instructions, examples, context, and formats steer model behavior at inference time.

See all 425 papers

Instruction-Tuned Models Locally Reuse Human Syntax More Than Humans Do
Jul 28, 2026Zandi EberstadtLanguage Modeling Instruction Tuning
Department of Computer Science University of Oxford
Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs
May 19, 2026Carolina Camassa, Derek ShillerLLM Evaluation Language Model Self-Assessment
Future Impact Group · Rethink Priorities
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
Jul 28, 2026Chaemin Jang, Dongman Lee, Jihee KimLanguage Model Decoding LLM Prompting
School of Computing, KAIST · School of Business and Technology Management, KAIST · Graduate School of Data Science, KAIST

Jul 28, 2026J/K move · Enter open · S save

Instruction-Tuned Models Locally Reuse Human Syntax More Than Humans Do

Zandi Eberstadt

Department of Computer Science University of Oxford

Language Modeling Instruction Tuning

Open paper AskPDF arXiv

Syntactic convergence (the tendency of speakers to adapt in language towards the grammatical profiles of their interlocutors) is a well-documented feature of human dialogue widely considered to operate below conscious awareness. Whether large language models exhibit analogous syntactic convergence toward human users relative to human baselines and across a broad range of syntactic constructions remains an open question. Using substitution-paradigm data in which model generations replace one speaker's turns in pre-existing human dialogues, this study measures turn-adjacent reuse of context-free grammar (CFG) rules across sixteen open-weight Llama and Gemma models (1B-70B, pretrained and instruction-tuned) at 1,901 matched positions per model. Every model showed greater CFG-rule overlap with the preceding human turn than with a sampled unrelated human prime, and in every model this actual-versus-random difference was larger for lower-frequency rules. Each instruction-tuned model also showed greater natural-output overlap with the actual prime than the human response it replaced, and all eight matched architecture pairs exhibited greater actual-prime overlap after instruction tuning. However, relative to pretrained variants, instruction-tuned outputs overlapped more with unrelated primes, showed a smaller actual-versus-random increment, and had lower conditional rule-reuse odds once target rule-set size was held constant. In exploratory analyses, each model exhibited greater mean lexical and semantic similarity to the preceding turn than the matched human responses did. Instruction-tuned models additionally produced responses with greater mean semantic similarity than their pretrained counterparts in all eight architecture pairs, whereas the lexical similarity results were more heterogeneous.

Abstract

LLM Prompting

Explore similar work

Instruction-Tuned Models Locally Reuse Human Syntax More Than Humans Do

Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs

Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe