Hello! I'm Sarubi, a PhD candidate in Computational Linguistics at Saarland University, supervised by Prof. Alexander Koller and Prof. Vera Demberg. My work is part of the project A8: Adapting Text Generation to Individual Users in the Collaborative Research Center SFB 1102.

Current large language models optimize for a generic, average user, overlooking individual differences in how people process and express language. My PhD addresses this by personalizing text generation to the individual user—both how they comprehend text and how they write.

Rather than relying on one-size-fits-all proxies like readability scores, I model users directly: analyzing how linguistic and structural text complexity interacts with a reader's cognitive profile to dynamically tailor text to their comprehension capacity. At the same time, I study how to personalize generation for writers—using controllable generation, reinforcement learning, and model merging / adapter composition to capture an individual's unique voice and stylistic nuances rather than flattening their writing into generic model outputs.

• ✍️ AuthorMix (EMNLP 2026, Oral)
  What if your writing assistant preserved your authentic voice rather than generic LLM prose?
  AuthorMix adapts a model to an individual's writing style from just a handful of examples in under an hour. By leveraging model merging and style-adapter composition guided by reinforcement learning, it outperforms strong proprietary baselines (including GPT-5.1) at jointly optimizing stylistic fidelity and semantic preservation.
• 📖 Individual Differences in Readability (AMLaP 2026, ongoing)
  What makes a text difficult to process: syntactic depth, lexical frequency, or their interaction with reader cognitive profiles?
  We disentangle structural and lexical complexity to identify the exact levers personalized text adaptation systems must manipulate to match each reader's linguistic processing capacity.
• 🎯 Fine-Grained Controllable Adaptation (AAAI 2025 Workshop on AI4ED)
  Moving beyond a single, blunt “make it simpler” slider toward multi-attribute control.
  We demonstrate that LLMs can achieve fine-grained, independent control over distinct complexity axes (e.g., syntactic tree depth vs. vocabulary rarity) via structured in-context learning—eliminating the need for fine-tuning while providing the exact knobs required for personalized generation.

Before diving into personalization, my background was rooted in multilingual NLP. During my Master's (by research) at the University of Moratuwa with Dr. Surangika Ranathunga and Prof. Sanath Jayasena, I worked on adapting pre-trained multilingual sequence-to-sequence models for low-resource machine translation (Sinhala, Tamil, English). This research won the Best Paper Award at MERCon 2021 and led to publications in ACL Findings (2022) and NAACL Findings (2024), as well as an operational machine translation pipeline deployed for the Sri Lankan government sector. Prior to my Master's, I worked as a software engineer at WSO2.