XtraGPT: LLMs for Human-AI Collaboration on Controllable Scientific Paper Refinement

ICLR 2025 Workshop BuildingTrust Submission101 Authors

Track: Long Paper Track (up to 9 pages)

Keywords: Paper Refinement, Large Language Models

TL;DR: XtraGPT is a framework that helps authors improve their scientific papers by providing context-aware feedback, refining structure, and enhancing drafts while preserving core ideas and allow them to focus on deeper issues.

Abstract:

The increasing volume of scientific publications highlights the growing need for high-quality academic writing. However, while groundbreaking ideas are often present, many papers fail to meet academic writing standards. Unlike open-ended applications of large language models (LLMs) in research, which delegate creative tasks to AI, we emphasize a human-centered approach where researchers provide ideas and drafts while LLMs strictly follow user instructions for refinement. All XtraGPT data, training and evaluation processes, and models will be open-sourced.

We propose XtraGPT, LLMs designed to assist authors by delivering instruction-driven, context-aware revisions that (1) adhere to user instructions, (2) align with general academic writing standards, and (3) are consistent with the whole paper. Leveraging a dataset of 7,040 ICLR 24 papers and over 140,000 question-answer pairs, XtraGPT enhances specific sections without compromising the paper’s integrity. Experimental results show XtraGPT-7B surpass similar size models and is competitive with GPT-4o-mini in providing high-quality, context-aware refinements. We also found that scaling up model parameters provides limited improvement for the difficulty of paper scoring. Modifying six sections with XtraGPT can improve the paper’s rating according to the predictor.

By prioritizing controllability in the task of paper refinement, XtraGPT empowers researchers to focus on innovation while relying on the system to handle the demands of academic writing with context understanding and adherence to academic standards and user instructions.

Submission Number: 101