NarrativeWorldBench: A Frontier-Saturated Benchmark and a Latent World Model for Long-Horizon Co-Creative Audio Drama

2026-06-15

AI Summary

This work addresses the challenge of maintaining narrative consistency in long-form audio drama generation (200–800 episodes) with current large language models. To this end, we propose the Narrative Variational State-Space Model (N-VSSM), which integrates structured latent states, a Mamba-2 backbone, and a cultural transfer mechanism to significantly enhance long-range coherence and cross-lingual controllability. We also introduce NarrativeWorldBench, the first multilingual benchmark for long-horizon narrative evaluation covering Hindi, Tamil, Telugu, and Marathi. Experiments show that N-VSSM achieves episode-level plot beat F1 scores of at least 0.84 across all time horizons, at only one-fourth the computational cost of leading closed-source models. In user studies, N-VSSM outperforms Claude Opus 4.5 in 71% of scenarios, demonstrates 1.3 points higher controllability on a Likert scale, and improves cross-lingual fidelity by 0.20–0.23.

Abstract

Long-form serialized audio drama, with arcs that run for 200 to 800 episodes, is a major creative medium and a setting where frontier large language models (LLMs) fail. We benchmark 21 models, spanning classical, fine-tuned, open-frontier, closed-frontier, and reasoning tiers, on a uniform set of structural narrative metrics. All closed-frontier systems saturate at a plot-beat F1 in the band [0.78, 0.81] and collapse by about -0.20 F1 at horizon h=200. We introduce NarrativeWorldBench, an open benchmark of nine narrative-structure metrics evaluated across horizons h in {10, 20, 50, 100, 200}, with cross-lingual evaluation across four Indic languages (Hindi, Tamil, Telugu, Marathi). We introduce 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 holds plot-beat F1 >= 0.84 across all horizons at 4x lower compute than the closed-frontier band. A learned Cultural Transfer Function lifts 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 on long-arc consistency 71% of the time and rated +1.3 Likert points higher on controllability.

Problem

• long-horizon narrative
• audio drama
• large language models
• narrative consistency
• co-creative generation

Innovation

• NarrativeWorldBench
• N-VSSM
• long-horizon narrative generation
• latent world model
• cross-lingual audio drama