aryan.

Can LLMs Adapt to Speakers When Answering Questions?


Large Language Models (LLMs) have shown promise in generating human-like text, but their ability to adapt to the speaker’s identity, complexity level, or knowledge remains underexplored. My research focuses on training LLMs to dynamically adapt their communication style using a novel method known as Direct Preference Optimization (DPO). This project explores whether an LLM can autonomously adjust its language to match the proficiency of the speaker, such as simplifying its responses when engaging with non-native speakers or ESL learners, without explicit instructions.

The primary objective of the project is to investigate if an LLM can modify its language complexity—using tools like the Flesch-Kincaid Grade Level (FKGL) and Flesch-Kincaid Reading Ease (FKRE)—to improve accessibility and communication clarity. This is achieved through fine-tuning a pre-existing model with datasets such as the Teacher-Student Chatroom Corpus (TSCC), where dialogues between teachers and students of varying English proficiency levels provide a robust basis for training.

Project Highlights:

  • Training Method: We employ Direct Preference Optimization (DPO) to directly optimize the model’s output preferences, making it responsive to different levels of language complexity.
  • Datasets: The TSCC dataset includes diverse student-teacher conversations annotated with student proficiency levels, allowing the model to learn to adapt its language based on input complexity.
  • Results: Our experiments demonstrate that fine-tuning an LLM using DPO improves its ability to simplify responses when required, although challenges remain in ensuring that the simplifications don’t detract from the informational content.

Future Directions:

  1. Improved Negative Data Handling: A more suitable dataset of negative examples is needed to refine the model’s ability to avoid overgeneralization.
  2. Diverse Model Testing: The research will be expanded to evaluate how different models (e.g., ChatGPT, Gemini) adapt across various complexities and linguistic backgrounds.
  3. Real-Time Adaptation: Future work will explore real-time feedback mechanisms to allow the model to fine-tune its responses during live interactions.

If you have suggestions or would like to collaborate, feel free to reach out!

Best,
Aryan Shetty