DraMaze 以情節母題記憶提升長篇故事連貫性
Trope memory powers coherent long‑form stories
DraMaze 透過多智能體流程及具未來感知能力的敍事記憶，提升長篇微短劇劇本的連貫性與多樣性。在最多 60 k words 的 60 集劇本中，因果模式重複減少 32.1%，Logic 得分達 82.5，較最佳基線高 8.5 分。移除未來感知敍事記憶後，Logic 得分降至 74.5，3-gram 重複率則由 9.22% 升至 12.2%。
Multi‑agent pipelines can now deliver narratives that are both diverse and logically tight, thanks to a system that weaves trope‑based memory into an evolutionary tree search. The twist is that several agents coordinate through a backward narrative dependency structure, allowing the story to anticipate its own future while still exploring many causal paths.
Before this work, long‑form script generation relied on single LLMs or simple retrieval, which tended to collapse onto a narrow set of plot mechanisms as generation progressed. Conventional retrieval even reinforced trope reuse and ignored historical cues needed for later twists, leading to repetitive microdramas.
DraMaze cuts causal‑pattern repetition by 32.1 % on 60‑episode scripts of up to 60 k words. In the same setting it reaches a Logic score of 82.5, beating the best baseline by 8.5 points, and its logical performance stays stable across 20, 40, and 60 episodes while the advantage over baselines grows from 4.2 to 8.5 points [1].
Removing future‑aware narrative memory collapses that advantage: Logic falls to 74.5 and 3‑gram repetition climbs from 9.22 % to 12.2 %, showing the component’s critical role for coherence and variety [1]. The approach still hinges on a curated trope repository and an expensive tree search, leaving open whether it scales to open‑world storytelling or real‑time authoring.
MiDramaBench should become the default benchmark for any AI‑assisted drama tool, and modular multi‑agent architectures like DraMaze ought to replace monolithic generators in production pipelines.
References
來源：DEV Community · dev.to