Large language models (LLMs) have ushered in a multitude of new language generation capabilities, bringing AI-guided storytelling much closer to reality. Authors can now meaningfully engage with AI-writing assitants to help during the story writing process. These same capabilities have also enabled players to experience more complex and diverse interactions in narrative-driven roleplaying games. Though due to the memory and processing that LLMs require, efficient inference techniques are needed to address the needs of these real-time systems.
I begin by exploring LLMs as AI-writing assistants on the online collaborative writing platform Storium, where real authors query a model for suggested story continuations and then edit them. I follow up by observing players' experiences interacting with dialogue generated by GPT-4 within the widely-acclaimed role-playing game "Disco Elyisum: The Final Cut." I then investigate efficient inference for the Transformer architecture by incorporating linguistic features into the decoding process by first autoregressively predicting a chunked parse tree before generating all of the target tokens in one shot conditioned on the predicted parse.
For my proposed work, I will integrate LLMs into the dialogue writing process for video games by expanding the capabilities of the data analysis tool I created for "Disco Elyisum" to enable AI-powered dialogue authoring. Time permitting, I will continue my exploration of efficient NLP methods by investigating approaches for sparse parameter updates during LM training.
Advisor: Mohit Iyyer