DraMaze: A Multi-Agent Pipeline with Future-Aware Memory for Diverse Microdrama Script Generation

Siyang He; Qiqi Wang; Fan Liu; Yuncheng Huang; Pengfei Xu; Xi Chen; Ziwei He; Xipeng Qiu

cs.CLOct 22, 2026

DraMaze: A Multi-Agent Pipeline with Future-Aware Memory for Diverse Microdrama Script Generation

Authors: Siyang He, Qiqi Wang, Fan Liu, Yuncheng Huang, Pengfei Xu, Xi Chen, Ziwei He, Xipeng Qiu

Organizations: Fudan University · Shanghai Innovation Institute · Solar System of OVB, Tencent, China

Abstract

The rapid growth of microdramas has increased demand for high-quality script production. Long-form microdrama script generation with LLMs requires both diverse plot progression and long-range narrative coherence. Our observations reveal that candidate continuations increasingly concentrate on fewer narrative mechanisms as generation progresses, while conventional retrieval may reinforce trope reuse and overlook historical information essential to future plots. To address these challenges, we propose DraMaze, a multi-agent pipeline that explores diverse causal paths through trope memory-guided evolutionary tree search and mutation. It further constructs a Backward Narrative Dependency Tree to organize, retain, and retrieve historical memories based on future narrative goals. We also introduce MiDramaBench, a bilingual benchmark for evaluating script logic and creative quality across different script lengths, together with causal-pattern metrics for narrative repetition and diversity. Experiments show consistent improvements across three backbone LLMs. For 60-episode scripts of up to 60k words, DraMaze reduces causal-pattern repetition by 32.1% and achieves a Logic score of 82.5, outperforming the best-performing baseline by 8.5 points.

Figures & tables

Figure 3: Overview of our multi-agent pipeline for microdrama script generation. (1) Outline Planner creates a four-stage coarse outline. (2) Evolutionary tree search refines event sequences through mutation and trope-guided selection. (3) Develop into episode outlines.(4) Backward Narrative Dependency Tree is constructed from episode outlines to organize subsequent future-aware narrative memory, with episodes attached to goal nodes to guide retrieval. (5) Agents retrieve memories, plan scenes, generate scripts, and update memories after each episode.

Figure 4: Overview of the Backward Narrative Dependency Tree (BNDT) and future-aware narrative memory retrieval. Episode outlines guide the backward construction of goal dependencies and episode attachment. Generated events, clues, and foreshadowing are linked to goal nodes, enabling retrieval from the current episode’s goals and their prerequisite subtrees to support subsequent script generation.

Figure 5: Overview of the multi-agent evaluation framework for script logic and creative quality.

Figure 6: Construction pipeline of MiDramaBench. Source texts are selected by length, deduplicated, and screened for dramatic adaptation potential. Core plot developments, character relationships, and world settings are extracted to guide the dramatic adaptation of each source into a microdrama idea. Quality-reviewed ideas are grouped by genre-label embeddings and sampled within each length–cluster group to form the evaluation subset.

		Script Logic								Script Quality						Homogeneity
Ep.	Method	CC	CS	KS	RS	RST	SP	RD	Avg	CQ	DE	PP	ES	SQ	Avg	4G $\downarrow$	3G $\downarrow$	$\overline{H}\uparrow$
20	DOME [-1pt] $\pm 0.0$	$78.4$ [-1pt] $\pm 1.3$	$75.1$ [-1pt] $\pm 0.8$	$81.4$ [-1pt] $\pm 0.3$	$77.2$ [-1pt] $\pm 0.7$	$77.7$ [-1pt] $\pm 0.4$	$81.2$ [-1pt] $\pm 1.0$	$77.8$ [-1pt] $\pm 1.2$	$78.4$ [-1pt] $\pm 0.1$	$77.6$ [-1pt] $\pm 0.4$	$79.3$ [-1pt] $\pm 1.0$	$71.3$ [-1pt] $\pm 0.9$	$76.0$ [-1pt] $\pm 1.0$	$85.1$ [-1pt] $\pm 0.9$	$77.9$ [-1pt] $\pm 0.5$	$4.23$ [-1pt] $\pm 0.60$	$11.0$ [-1pt] $\pm 0.15$	$0.876$ [-1pt] $\pm 0.0057$
	StoryWriter [-1pt] $\pm 0.0$	$74.7$ [-1pt] $\pm 0.3$	$74.5$ [-1pt] $\pm 0.7$	$76.2$ [-1pt] $\pm 0.4$	$72.8$ [-1pt] $\pm 1.0$	$72.6$ [-1pt] $\pm 1.1$	$77.9$ [-1pt] $\pm 0.4$	$83.5$ [-1pt] $\pm 0.5$	$76.0$ [-1pt] $\pm 0.3$	$83.5$ [-1pt] $\pm 0.8$	$84.0$ [-1pt] $\pm 0.3$	$77.3$ [-1pt] $\pm 1.0$	$82.2$ [-1pt] $\pm 0.2$	$86.8$ [-1pt] $\pm 0.1$	$82.8$ [-1pt] $\pm 0.1$	$4.34$ [-1pt] $\pm 0.64$	$10.0$ [-1pt] $\pm 1.1$	$0.881$ [-1pt] $\pm 0.0095$
	Ours [-1pt] $\pm 0.0$	$81.5$ [-1pt] $\pm 0.2$	$80.4$ [-1pt] $\pm 1.0$	$82.6$ [-1pt] $\pm 0.4$	$80.1$ [-1pt] $\pm 0.3$	$81.3$ [-1pt] $\pm 0.4$	$83.7$ [-1pt] $\pm 1.2$	$88.5$ [-1pt] $\pm 0.5$	$\mathbf{82.6}$ [-1pt] $\pm 0.2$	$87.3$ [-1pt] $\pm 0.7$	$87.7$ [-1pt] $\pm 0.6$	$82.9$ [-1pt] $\pm 0.3$	$85.2$ [-1pt] $\pm 0.7$	$87.6$ [-1pt] $\pm 0.3$	$\mathbf{86.1}$ [-1pt] $\pm 0.3$	$\mathbf{1.35}$ [-1pt] $\pm 0.27$	$\mathbf{5.51}$ [-1pt] $\pm 0.64$	$\mathbf{0.939}$ [-1pt] $\pm 0.0054$
40	DOME [-1pt] $\pm 0.0$	$73.0$ [-1pt] $\pm 0.7$	$65.3$ [-1pt] $\pm 0.6$	$77.6$ [-1pt] $\pm 0.7$	$73.0$ [-1pt] $\pm 0.5$	$71.1$ [-1pt] $\pm 1.3$	$76.1$ [-1pt] $\pm 0.7$	$68.8$ [-1pt] $\pm 1.2$	$72.1$ [-1pt] $\pm 0.3$	$77.0$ [-1pt] $\pm 0.9$	$78.4$ [-1pt] $\pm 0.8$	$69.4$ [-1pt] $\pm 0.1$	$76.3$ [-1pt] $\pm 0.5$	$85.1$ [-1pt] $\pm 0.2$	$77.3$ [-1pt] $\pm 0.3$	$7.86$ [-1pt] $\pm 0.65$	$17.3$ [-1pt] $\pm 1.0$	$0.883$ [-1pt] $\pm 0.0034$
	StoryWriter [-1pt] $\pm 0.0$	$73.5$ [-1pt] $\pm 1.2$	$70.8$ [-1pt] $\pm 0.8$	$76.3$ [-1pt] $\pm 0.8$	$72.3$ [-1pt] $\pm 0.5$	$71.2$ [-1pt] $\pm 1.1$	$77.4$ [-1pt] $\pm 1.4$	$81.1$ [-1pt] $\pm 0.2$	$74.6$ [-1pt] $\pm 0.8$	$85.9$ [-1pt] $\pm 0.2$	$84.3$ [-1pt] $\pm 0.3$	$78.1$ [-1pt] $\pm 1.0$	$82.0$ [-1pt] $\pm 0.1$	$87.4$ [-1pt] $\pm 0.2$	$83.5$ [-1pt] $\pm 0.3$	$6.65$ [-1pt] $\pm 0.57$	$15.2$ [-1pt] $\pm 1.2$	$0.894$ [-1pt] $\pm 0.0025$

Table 1: Main results on bilingual long-form script generation at 20, 40, and 60 episodes. Logic and Quality scores range from 0 to 100 (higher is better). Logic: CC (causal chain), CS (character state), KS (knowledge state), RS (resource state), RST (rules and space-time), SP (setup and payoff), and RD (repetition detection). Quality: CQ (character quality), DE (dramatic engagement), PP (plot-planning quality), ES (episodic script quality), and SQ (setting quality). Avg denotes the overall score for each dimension. Homogeneity: 4G denotes four-slot causal-pattern reuse (%), 3G the average of two three-slot reuse rates (%), and

\overline{H}

the mean of three normalized conditional entropies (0–1). Cells report the mean (top) and sample standard deviation across three group means (bottom). Bold indicates the best Avg and Homogeneity scores within each episode group.

Table 7

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Field	Input idea
Genre	Corporate thriller; family drama.
Core plot	A former CEO, publicly ousted for a scandal he did not commit, returns to his family company disguised as a janitor. He must expose board members who are siphoning assets and forcing his estranged wife into a merger that would erase her legacy. When his son, a junior executive investigating his father’s disappearance, becomes the board’s next target, the search for evidence becomes a race to protect him and prevent the company’s dissolution. The CEO’s low-status disguise lets him move unnoticed through the hierarchy, gather evidence, and test loyalties while risking exposure at every encounter.
Characters	The former CEO is observant and strategically restrained, but secrecy distances him from his family. His impulsive, loyal son is desperate to establish his authority and sees the disguised father as a stranger. A politically adept board rival seeks control of the company and the son’s removal. The CEO’s wife is a resourceful executive who delays the merger while investigating the suspicious new janitor.
World Setting	A closed corporate environment in which reputation, hierarchy, and access to information determine power. Nondisclosure agreements and internal governance restrict outside legal recourse, making an internal resolution necessary. Access credentials and artifacts such as a keycard or private server can serve as proof of authority.

Appendix

Table 4: Example English microdrama idea with four fields: genre, core plot, characters, and world setting.

Panel of Figure 7 — Figure 7: A complete backward narrative dependency tree (BNDT) generated for a 40-episode wartime drama. Orthogonal connectors show parent–child dependencies and each child is a prerequisite for its parent. All 28 node goals are reproduced from the generated tree, in depth-first order.

Dimension	Category	Principal checks
Logic	Causal chain	Prerequisite information, action motivation, causal support for key turns, and justification of crisis resolution.
	Character state	Continuity of life, identity, allegiance, abilities, personality, motivation, and relationships.
	Knowledge state	Information acquisition, knowledge sources, retention of established facts, and the scope of secret disclosure.
	Resource state	Ownership and transfer, depletion and restoration, evidence provenance, and support for resource increases.
	Rules and space-time	Established rules, preparation of new rules, temporary exceptions, event chronology, and spatial movement.
	Setup and payoff	Abandoned setups, forced payoffs, unprepared revelations, and missing repercussions of major events.

Appendix

Table 5: Summary of the Logic and Quality evaluation criteria.

		Script Logic								Script Quality						Homogeneity
Model	Method	CC	CS	KS	RS	RST	SP	RD	Avg	CQ	DE	PP	ES	SQ	Avg	4G $\downarrow$	3G $\downarrow$	$\overline{H}\uparrow$
DeepSeek- V4-Pro	DOME [-1pt] $\pm 0.0$	$75.5$ [-1pt] $\pm 1.1$	$69.0$ [-1pt] $\pm 1.0$	$78.3$ [-1pt] $\pm 0.2$	$76.3$ [-1pt] $\pm 0.5$	$72.5$ [-1pt] $\pm 0.7$	$78.5$ [-1pt] $\pm 0.5$	$75.6$ [-1pt] $\pm 1.3$	$75.1$ [-1pt] $\pm 0.4$	$87.5$ [-1pt] $\pm 0.7$	$86.2$ [-1pt] $\pm 0.5$	$79.0$ [-1pt] $\pm 1.0$	$83.4$ [-1pt] $\pm 1.1$	$87.2$ [-1pt] $\pm 0.1$	$84.7$ [-1pt] $\pm 0.5$	$8.23$ [-1pt] $\pm 0.70$	$17.2$ [-1pt] $\pm 0.43$	$0.853$ [-1pt] $\pm 0.012$
	StoryWriter [-1pt] $\pm 0.0$	$74.2$ [-1pt] $\pm 0.8$	$69.4$ [-1pt] $\pm 1.1$	$78.2$ [-1pt] $\pm 0.3$	$75.0$ [-1pt] $\pm 0.8$	$72.1$ [-1pt] $\pm 0.5$	$76.5$ [-1pt] $\pm 0.8$	$67.8$ [-1pt] $\pm 1.1$	$73.3$ [-1pt] $\pm 0.7$	$87.7$ [-1pt] $\pm 1.3$	$86.0$ [-1pt] $\pm 0.8$	$77.2$ [-1pt] $\pm 0.6$	$85.9$ [-1pt] $\pm 0.3$	$86.7$ [-1pt] $\pm 0.3$	$84.7$ [-1pt] $\pm 0.6$	$6.74$ [-1pt] $\pm 0.64$	$14.5$ [-1pt] $\pm 1.0$	$0.829$ [-1pt] $\pm 0.0068$
	Ours [-1pt] $\pm 0.0$	$86.0$ [-1pt] $\pm 0.1$	$85.7$ [-1pt] $\pm 0.5$	$85.0$ [-1pt] $\pm 0.4$	$83.6$ [-1pt] $\pm 0.5$	$85.4$ [-1pt] $\pm 0.4$	$86.9$ [-1pt] $\pm 0.2$	$89.8$ [-1pt] $\pm 0.1$	$\mathbf{86.1}$ [-1pt] $\pm 0.2$	$92.7$ [-1pt] $\pm 0.1$	$91.1$ [-1pt] $\pm 0.6$	$87.8$ [-1pt] $\pm 0.1$	$89.9$ [-1pt] $\pm 0.3$	$89.7$ [-1pt] $\pm 0.1$	$\mathbf{90.2}$ [-1pt] $\pm 0.2$	$\mathbf{3.21}$ [-1pt] $\pm 0.57$	$\mathbf{10.2}$ [-1pt] $\pm 0.51$	$\mathbf{0.915}$ [-1pt] $\pm 0.0030$
GLM5.3- Flash	DOME [-1pt] $\pm 0.0$	$83.9$ [-1pt] $\pm 0.8$	$80.2$ [-1pt] $\pm 0.6$	$83.8$ [-1pt] $\pm 0.3$	$82.2$ [-1pt] $\pm 0.7$	$81.7$ [-1pt] $\pm 0.6$	$84.9$ [-1pt] $\pm 0.7$	$86.5$ [-1pt] $\pm 0.7$	$83.3$ [-1pt] $\pm 0.4$	$92.0$ [-1pt] $\pm 0.8$	$90.1$ [-1pt] $\pm 0.3$	$86.4$ [-1pt] $\pm 0.7$	$87.0$ [-1pt] $\pm 0.6$	$89.4$ [-1pt] $\pm 0.1$	$89.0$ [-1pt] $\pm 0.3$	$8.03$ [-1pt] $\pm 0.94$	$15.1$ [-1pt] $\pm 0.94$	$0.881$ [-1pt] $\pm 0.0062$
	StoryWriter [-1pt] $\pm 0.0$	$85.1$ [-1pt] $\pm 0.5$	$83.8$ [-1pt] $\pm 0.5$	$84.7$ [-1pt] $\pm 0.6$	$84.2$ [-1pt] $\pm 0.3$	$83.2$ [-1pt] $\pm 0.6$	$86.7$ [-1pt] $\pm 0.2$	$85.9$ [-1pt] $\pm 0.7$	$84.8$ [-1pt] $\pm 0.4$	$92.4$ [-1pt] $\pm 0.3$	$89.9$ [-1pt] $\pm 0.1$	$85.0$ [-1pt] $\pm 0.4$	$86.5$ [-1pt] $\pm 0.7$	$89.1$ [-1pt] $\pm 0.3$	$88.6$ [-1pt] $\pm 0.2$	$7.08$ [-1pt] $\pm 1.1$	$14.9$ [-1pt] $\pm 0.90$	$0.870$ [-1pt] $\pm 0.010$

Appendix

Table 6: Cross-model generalization results on 40-episode script generation using DeepSeek-V4-Pro and GLM5.3-Flash. Results are reported as mean

\pm

standard deviation. Best average and homogeneity scores within each model group are highlighted in bold.

Narrative Generation

See all 34 papers

One Sentence, One Drama: Personalized Short-Form Drama Generation via Multi-Agent Systems
May 21, 2026Yufei Shi, Weilong Yan, Naixuan Huang +5Multi-Agent LLM Systems Video Generation
MedVisAI Lab, Lee Kong Chian School of Medicine, Nanyang Technological University · National University of Singapore · Beijing Institute of Technology +2
PlotTwist: A Creative Plot Generation Framework with Small Language Models
Mar 17, 2026Abhinav Thorat, Ravi Kolla, Jyotin Goel +2Narrative Generation Small Language Models
Sony Research India
Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding
Sep 7, 2026Yixin Qin, Shi-Zhe Chen, Zhiqi Yu +4Video Understanding Video-Language Model Evaluation
Tencent PCG · The University of Hong Kong · Tencent CSIG

May 21, 2026J/K move · Enter open · S save

One Sentence, One Drama: Personalized Short-Form Drama Generation via Multi-Agent Systems

Yufei Shi, Weilong Yan, Naixuan Huang, Yucheng Chen+4

MedVisAI Lab, Lee Kong Chian School of Medicine, Nanyang Technological University · National University of Singapore · Beijing Institute of Technology+2

Multi-Agent LLM Systems Video Generation Video Generation Evaluation

Open paper AskPDF arXiv

Existing approaches for digital short-drama production typically rely on one-shot LLM generated scripts and loosely coupled pipelines, which fail to satisfy three key requirements of short-drama generation: (1) narrative pacing, resulting in weak hooks, insufficient escalation, and unattractive endings; (2) spatial consistency, leading to drifting scene layouts and inconsistent character positions across clips; and (3) production-level quality control, requiring extensive manual review and correction across script and visual stages. We present One Sentence, One Drama, a hierarchical multi-agent framework that transforms a user's single-sentence idea into a fully produced short drama through structured intermediate modules and iterative refinement. Our approach is built upon three key components: (1) a multi-agent debate-based story generation module that enforces short-drama pacing and narrative coherence; (2) a 3D-grounded first-frame generation mechanism that establishes a shared spatial reference for consistent character positioning and scene layout across clips; and (3) multi-stage reviewer loops that perform comprehensive error detection and targeted revision across script, visual, and video generation stages. We also introduce scene-level BGM matching and scene transition planning to improve the audience's immersive experience. To systematically evaluate this task, we introduce Short-Drama-Bench, a benchmark that extends standard video quality metrics with short-drama-specific criteria. Experimental results demonstrate that our method significantly outperforms existing pipelines in narrative quality, cross-clip consistency, and overall viewing experience.

Abstract

Narrative Generation

Explore similar work

One Sentence, One Drama: Personalized Short-Form Drama Generation via Multi-Agent Systems

PlotTwist: A Creative Plot Generation Framework with Small Language Models

Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding