Seminarium Instytutu

W poniedziałek 8 czerwca w godz. 12:15-13:00 w sali 3/40 dr hab. Arkadiusz Orłowski, prof. SGGW, wygłosi referat pod tytułem „Can Moderate-Sized Language Models Imitate an Individual Academic Writing Style?”

Abstract
Recent advances in large language models (LLMs) have significantly improved the fluency and coherence of automatically generated text. This study investigates whether moderate-sized causal language models (up to 500 million parameters) can reproduce the individual academic writing style of a specific author through fine-tuning on a limited corpus of scientific introductions. Three open-access models (Gemma 3, GPT-2 Medium, and Qwen1.5) were evaluated before and after adaptation using six complementary metrics: perplexity, semantic similarity, a style classifier, stylometric features, stylometric distance, and idiosyncratic phrase matching. Results indicate substantial improvements in language modelling quality and moderate gains in classifier-based stylistic attribution. However, distinctive authorial traits remain weakly represented, suggesting that adaptation primarily reinforces genre-level conformity rather than robust individual style imitation.