Can AI writing be salvaged? Mitigating Idiosyncrasies and Improving Human-AI Alignment in the Writing Process through Edits

Research Background and Issues

• What problems or challenges did the authors identify?

  • Texts generated by large language models (LLMs) often exhibit numerous "idiosyncratic flaws" in writing style, such as clichés, excessive background information, unnecessary narration, and poor sentence structure.
  • These flaws not only make LLM-generated texts lack depth and originality but also lead to content homogenization, thereby diminishing creativity and diversity in writing.

• Why is this issue important?

  • As LLMs are widely used in social media, news creation, and education, the quality of their outputs directly impacts users' expressive capabilities and content quality.
  • If LLMs fail to adapt to human writing preferences, it could result in formulaic content and decreased user satisfaction, hindering long-term collaboration and trust between AI and humans.

• Research Motivation and Related Work

  • Previous studies have shown that large language models tend to generate patterned, repetitive content, and existing user-preference-based training mechanisms (e.g., Reinforcement Learning with Human Feedback, RLHF) often fail to address detailed textual issues.
  • This study aims to explore how editing methods can alleviate these problems, making LLM outputs more aligned with human expectations while preserving creativity and diversity in writing.

Solution

• What methods or solutions did the authors propose?

  • A text editing classification method consisting of seven categories was proposed to identify and improve flaws in LLM-generated texts, including clichés, unnecessary narration, improper sentence structure, etc.
  • A high-quality dataset named LAMP was created, containing 1,057 LLM-generated texts and 8,035 fine-grained edits made by 18 professional writers.
  • An automated pipeline was designed to detect and rewrite problematic paragraphs, leveraging few-shot learning (prompting) to enhance the model's self-editing capabilities.

• What are the innovative aspects of the solution?

  • A systematic editing classification method based on professional writing practices was developed, providing a clear framework for studying LLM writing issues.
  • For the first time, a large-scale human revision dataset covering creative and non-fiction writing was constructed, offering valuable resources for future AI writing research.
  • Experiments compared the quality of outputs from different language models, revealing common limitations of LLMs in writing tasks.

• What are the implementation steps and key technologies used?

  1. Development of the Classification Method: Seven editing types (e.g., clichés, unnecessary verbosity, purple prose) were derived from professional writers' revision behaviors.
  2. Creation of the LAMP Dataset:
     • Extracted paragraphs from original human-written texts, designed generation instructions, and required LLMs to produce answers in specified styles.
     • Invited professional writers to revise these texts sentence by sentence and annotate each edit type.
  3. Automated Detection and Revision Method:
     • Used few-shot learning (prompting) to guide LLMs in detecting issues within paragraphs and generating revised versions.
     • Compared human-edited versions with automated revisions to measure editing effectiveness and alignment with human preferences.

Research Outcomes

• What specific results were achieved?

  • A large-scale dataset containing over 8,000 fine-grained edits was created, highlighting major problem areas in LLM outputs.
  • A two-stage editing process, including issue detection and rewriting, was developed and validated, demonstrating that automated editing can significantly improve text quality.
  • Experimental results showed that while automated editing cannot fully match professional writers, it exhibits potential, especially in addressing issues like overly long sentences or chaotic structures.

• What advantages does this solution have compared to existing ones?

  • Provides a more detailed diagnosis of LLM writing issues rather than simple preference assessments.
  • Improves traditional binary preference annotation methods by introducing explicit goal-oriented revisions based on human edits, enhancing training and evaluation efficiency.

• What were the experimental or evaluation results?

  • Texts revised by humans scored highest in preference evaluations, while LLM-automated revisions significantly outperformed original texts, indicating improved quality through the editing pipeline.
  • Experiments showed that even fully automated editing, where LLMs identify issues and generate revisions, performed comparably to semi-automated systems where humans provided issue points—suggesting that quality is primarily limited by the rewriting phase.

• Limitations and Future Directions

  • Limitations:
    • The study primarily focuses on creative and non-fiction writing, and its findings may not fully apply to technical or scientific writing domains.
    • Editing effectiveness relies on few-shot learning, which may not entirely capture the complexity of human edits.
    • Model editing struggles with generating new details and aligning with higher-order language styles.
  • Future Directions:
    • Expand the editing classification method to cover more writing genres.
    • Explore joint training techniques to enhance model editing capabilities with larger datasets.
    • Investigate how to integrate editing pipelines with personalized user preferences, such as supporting co-creation through interactive writing tools.

Through this comprehensive study, the authors not only provide an in-depth analysis and practical methods for addressing LLM writing quality issues but also outline feasible paths for improving the efficiency and creativity of writing models in collaboration with humans.