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

A new benchmark reveals that frontier large language models hit a performance ceiling when generating long-form serialized audio drama, while a purpose-built latent world model sustains narrative coherence across hundreds of episodes with lower computational cost [1]. The benchmark, called NarrativeWorldBench, evaluates 21 models on nine structural narrative metrics at horizons ranging from 10 to 200 episodes [1]. All closed-frontier systems saturate at a plot-beat F1 score between 0.78 and 0.81 and degrade by roughly 0.20 F1 at the 200-episode mark [1]. The evaluation also spans four Indic languages — Hindi, Tamil, Telugu, and Marathi — to test cross-lingual fidelity [1]. The researchers introduce N-VSSM, a Narrative Variational State-Space Model built on a Mamba-2 backbone [1]. It maintains a structured 256-dimensional latent world state across more than 200 episodes and uses an event-conditioned posterior with an 8-billion-parameter decoder [1]. N-VSSM holds plot-beat F1 at or above 0.84 across all horizons while requiring 4x lower compute than the closed-frontier band [1]. A learned Cultural Transfer Function improves cross-language fidelity by 0.20 to 0.23 Likert points [1]. In a within-subjects study with 12 professional authors across 240 trials, N-VSSM was preferred over Claude Opus 4.5 on long-arc consistency 71 percent of the time and rated 1.3 Likert points higher on controllability [1]. The work underscores a gap between general-purpose language models and the demands of long-horizon co-creative storytelling, where maintaining plot coherence over hundreds of episodes remains a challenge [1].