Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Vernacular? I Barely Know Her: Challenges with Style Control and Stereotyping

Ankit Aich, Tingting Liu, Salvatore Giorgi, Kelsey Isman, Lyle Ungar, Brenda Curtis

TL;DR

This study evaluates how well state-of-the-art LLMs can be steered to produce text at specific reading levels and in African American English, using prompting and in-context learning across five models. It finds that ICL and prompts meaningfully reduce reading difficulty (mean drop around $9.9$ FK grade points, $p = 0.005$) and increase AAE usage (up to tenfold, $p = 0.007$), but persistent biases and stereotypes in vernacular language remain a major challenge. GPT-4 family models consistently outperform others in readability control, while Llama-3 exhibits high variability and Mistral-Instruct-7B shows competitive performance with fewer parameters. The work highlights both the promise and the risks of style control in educational and cross-cultural contexts, urging stronger instruction tuning, debiasing strategies, and guardrails to ensure safe, inclusive AI-assisted learning.

Abstract

Large Language Models (LLMs) are increasingly being used in educational and learning applications. Research has demonstrated that controlling for style, to fit the needs of the learner, fosters increased understanding, promotes inclusion, and helps with knowledge distillation. To understand the capabilities and limitations of contemporary LLMs in style control, we evaluated five state-of-the-art models: GPT-3.5, GPT-4, GPT-4o, Llama-3, and Mistral-instruct- 7B across two style control tasks. We observed significant inconsistencies in the first task, with model performances averaging between 5th and 8th grade reading levels for tasks intended for first-graders, and standard deviations up to 27.6. For our second task, we observed a statistically significant improvement in performance from 0.02 to 0.26. However, we find that even without stereotypes in reference texts, LLMs often generated culturally insensitive content during their tasks. We provide a thorough analysis and discussion of the results.

Vernacular? I Barely Know Her: Challenges with Style Control and Stereotyping

TL;DR

This study evaluates how well state-of-the-art LLMs can be steered to produce text at specific reading levels and in African American English, using prompting and in-context learning across five models. It finds that ICL and prompts meaningfully reduce reading difficulty (mean drop around FK grade points, ) and increase AAE usage (up to tenfold, ), but persistent biases and stereotypes in vernacular language remain a major challenge. GPT-4 family models consistently outperform others in readability control, while Llama-3 exhibits high variability and Mistral-Instruct-7B shows competitive performance with fewer parameters. The work highlights both the promise and the risks of style control in educational and cross-cultural contexts, urging stronger instruction tuning, debiasing strategies, and guardrails to ensure safe, inclusive AI-assisted learning.

Abstract

Large Language Models (LLMs) are increasingly being used in educational and learning applications. Research has demonstrated that controlling for style, to fit the needs of the learner, fosters increased understanding, promotes inclusion, and helps with knowledge distillation. To understand the capabilities and limitations of contemporary LLMs in style control, we evaluated five state-of-the-art models: GPT-3.5, GPT-4, GPT-4o, Llama-3, and Mistral-instruct- 7B across two style control tasks. We observed significant inconsistencies in the first task, with model performances averaging between 5th and 8th grade reading levels for tasks intended for first-graders, and standard deviations up to 27.6. For our second task, we observed a statistically significant improvement in performance from 0.02 to 0.26. However, we find that even without stereotypes in reference texts, LLMs often generated culturally insensitive content during their tasks. We provide a thorough analysis and discussion of the results.
Paper Structure (40 sections, 2 figures, 3 tables)

This paper contains 40 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Overall, the main contributions this paper makes are as follows:
  3. Task Description
  4. Grade Specific Reading Level Text Generation
  5. Text for In-Context Learning
  6. African-American English Generation
  7. Text for In-Context Learning
  8. Metrics
  9. Results and Analysis
  10. Analysis of reading level
  11. 2-shot Setup
  12. Llama-3:
  13. GPT Models:
  14. Mistral Instruct 7B:
  15. 1-shot Setup
  16. ...and 25 more sections

Figures (2)

  • Figure 1: Overall view of this paper. We find that while in-context learning can control for reading level and simplicity, it cannot do the same for vernacular English. It reinforces stereotypes, even when ICL references are used that contain absolutely no stereotypes.
  • Figure 2: Tables showing reading level scores for all models in a prompt-only, one-shot, and two-shot ICL setup. Scores are representative of first-grade reading level. A score closest to 1 is best.