LLM interactive narratives get a common yardstick: three experience managers trade off differently on one shared world model, and none wins everywhere
Synopsis
LLM-driven interactive narrative systems have been hard to compare because each experience manager is tied to its own bespoke scenario. An LLM-based narrative platform turns experience managers into interchangeable modules over a shared symbolic world model of facts, operations and goals, and every applied operation carries explicit preconditions and effects, so any session can be replayed as a common state-transition path. On a Little Red Riding Hood scenario, open generation, Monte Carlo repair and autonomous characters each ran 100 simulated sessions, and no manager led on all four properties: player agency, manager controllability, author intent and robustness.
Interpretation
Experience managers for interactive narrative can now be swapped over the same symbolic world model, and sessions are recorded as common state-transition paths, putting different architectures on one comparison footing. Previously an experience manager was usually shown in its own bespoke scenario, so its qualities could not be separated from the story, and the field lacked shared benchmarks. Three managers each ran 100 simulated sessions on the same Little Red Riding Hood scenario, with the session count set by a convergence criterion of less than 0.005 change in the running average of the properties.
The platform writes the narrative world as subject-predicate-object fact triples; the player issues natural-language commands and the experience manager decides which operations are actually applied that turn. Operations are represented with preconditions and add/delete effects, and are executed only when their preconditions hold; otherwise the turn is rejected and the world state is unchanged. The open manager parses or generates one operation per turn; the Monte Carlo repair manager searches event continuations only in dead-end states where no operation is applicable; the multi-agent manager lets each non-player character act once in a random order.
Four properties turn design goals into measurable indicators, making the trade-offs between managers explicit. The open manager reaches robustness 0.58 and completion rate 0.52, while Monte Carlo repair raises these to 0.76 and 0.80 but lowers player agency from 0.44 to 0.27. The multi-agent manager reaches the highest player agency at 0.68 and controllability at 0.52, with author intent at only 0.04; the open manager scores 0.47 on author intent, close to Monte Carlo repair's 0.44.
Perspective
The demonstrated scope is one fixed scenario inspired by Little Red Riding Hood, with the player simulated by GPT-5 mini at temperature 1, 100 sessions per manager, and all four properties computed on simulated plans.
The results still need confirmation with human players, notably whether the proposed indicators reflect perceived experience; a single scenario does not establish generalisability; player agency counts only whether proposals are accepted, not their real impact on world state or player feelings.