In long-form serialized audio drama, spanning 200 to 800 episodes, frontier large language models (LLMs) struggle significantly. This study benchmarks 21 models across classical, fine-tuned, open-frontier, closed-frontier, and reasoning tiers using a uniform set of structural narrative metrics. All closed-frontier systems saturate at a plot-beat F1 score in the range of [0.78, 0.81] and experience a drop of about -0.20 F1 at horizon h=200.
We introduce NarrativeWorldBench, an open benchmark consisting of nine narrative structure metrics evaluated across horizons h in {10, 20, 50, 100, 200}, with cross-lingual evaluation in four Indic languages (Hindi, Tamil, Telugu, Marathi). We propose N-VSSM, a Narrative Variational State-Space Model that maintains a structured 256-dimensional latent world state over more than 200 episodes via a Mamba-2 backbone with an event-conditioned posterior and an 8B decoder. N-VSSM achieves a plot-beat F1 of 0.84 across all horizons at 4x lower compute than the closed-frontier band. A learned Cultural Transfer Function enhances cross-language fidelity by +0.20 to +0.23 Likert points. In a within-subjects writer study (n = 12 professional authors, 240 trials), N-VSSM is preferred over Claude Opus 4.5 for long-arc consistency 71% of the time and rated +1.3 Likert points higher on controllability.
Blogger's Review: This study opens new perspectives in the field of long-form narrative generation. The introduction of the N-VSSM model not only enhances cross-lingual performance but also showcases its superiority in long-horizon consistency, making it noteworthy for potential applications in creative writing. The exploration of the Cultural Transfer Function provides new insights for multilingual generation.