Automatic generation of interactive narratives with LLMs
Industrial PhD (CIFRE) between ALTEN and Heudiasyc, started in September 2024 — Doctoral school “Sciences pour l'ingénieur”, Université de technologie de Compiègne.
Supervisors: Domitile Lourdeaux (Heudiasyc, CID team) and Mohamed Sallak (Heudiasyc, SCOP team).
Thesis record on theses.fr
Interactive narrative systems (INS) let a player or a learner influence how a story unfolds. Their design faces the narrative paradox: the more freedom the player has, the harder it is to guarantee a coherent story that remains faithful to the author's intent. Classical INS, based on hand-written scenarios or on planning, offer control but require heavy authoring effort and produce little story variety. Large language models (LLMs), on the other hand, generate rich and varied stories but remain hard to control.
My PhD explores how to couple LLMs with planning techniques and formal models in order to generate interactive narratives that are both varied and controlled, in particular for training applications.
A formalism based on extended state machines, which describes within a single framework INS with very different designs (branching narratives, planning-based systems, generative systems).
Definition and measurement of INS properties: player agency, authorial intent, controllability, robustness and variability. Indicators computed from simulated playthroughs make it possible to analyse the trade-offs behind the narrative paradox.
Design and comparison of Experience Manager architectures combining LLMs and symbolic techniques: open generation, Monte Carlo Tree Search (MCTS) projection, and multi-agent non-player characters.