Furkan Sahinuc

The rapid progress of large language models (LLMs) has opened new opportunities for automating complex scholarly writing tasks. Although LLMs can produce fluent text, they often struggle with expert domain requirements. To address this, we combine three main directions: (1) studying how task setup and input configurations influence writing quality of LLMs, (2) developing better evaluation pipelines that go beyond automatic metrics and include expert preferences, and (3) experimenting with modelling strategies, such as simulating expert feedback via domain specific reward models to guide the generation.

We also aim to expand our approach to different scientific writing tasks to improve generalization capabilities of our algorithms. By expanding across tasks, the project is not limited to addressing the limitations of scientific text generation but also contributes a broader understanding of how to align models with the diverse requirements of expert-level writing.