Source-linked AI summary
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
TL;DR
Long-form microdrama generation must balance plot diversity with long-range coherence, while conventional retrieval may miss historical information needed for future plots and reinforce trope reuse. DraMaze combines trope-guided evolutionary search with future-goal-based narrative memory. Across the reported script lengths, it improves logic and diversity, including a 60-episode Logic score of 82.5, 8.5 points above the strongest baseline.
Problem
Long-form microdrama scripts must maintain diverse plot progression and narrative coherence, but candidate continuations may concentrate on fewer mechanisms and conventional retrieval may miss historical information essential to future plots.
Source: §1 · pp. 1–2Method
DraMaze combines trope-guided evolutionary tree search and mutation with a Backward Narrative Dependency Tree that organizes and retrieves memories according to future narrative goals.
Source: §1 · p. 2Results
At 60 episodes, DraMaze reaches a Logic score of 82.5, with an 8.5-point margin over the strongest baseline, and 3-gram repetition of 14.6% versus 21.5% for that baseline.
Source: §6.2 · p. 12Takeaways & Limitations
Across 20-, 40-, and 60-episode scripts and multiple backbone LLMs, experiments report improvements in logic, creative quality, and diversity, with ablations validating key component effectiveness.
Source: §7 · p. 12
Abstract
from arXiv · showhide
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.
1 Introduction
Long-form microdrama generation must sustain diverse plot progression and coherence, yet candidate mechanisms can narrow and conventional retrieval may miss historical information needed for future plots. DraMaze pairs trope-guided plot exploration with future-aware memory, alongside a benchmark and metrics for evaluating scripts.
1 Introduction
- Long-form scripts require diverse plot progression and narrative coherence, but familiar patterns may recur despite changes in characters and events.Such repetition may make suspense-building turns predictable and weaken audience engagement.Source: §1 · pp. 1–2
- Candidate continuations concentrate on fewer progression mechanisms as generation advances, while conventional retrieval may reinforce trope reuse and miss information essential to future plots.The introduction reports declining mechanism entropy in the second half of scripts.Source: §1 · p. 2
- DraMaze combines evolutionary tree search with trope memory and mutation, then organizes historical memories around future narrative goals for episode writing.It also uses dynamic world-state tracking to maintain character, relationship, and asset states.Source: §1 · p. 2
- MiDramaBench evaluates script logic and creative quality, while causal-pattern metrics quantify narrative repetition and diversity across script lengths.The paper reports cross-model experiments and ablations to validate the pipeline.Source: §1 · p. 3
2 Related Work
Prior long-form writing frameworks use planning, revision, and agent collaboration to improve narrative coherence, while memory-augmented methods organize and retrieve story history. Organizing memories around the causal prerequisites of future narrative goals remains underexplored.
2.1 Long-Form Creative Writing Frameworks
- Long-form writing frameworks use recursive prompting, hierarchical outlines, dynamic memory, and agent collaboration, yet narrative repetition and homogenization persist.Examples include Re3, DOC, DOME, StoryWriter, CritiCS, WriteHERE, and HoLLMwood.Source: §2.1 · p. 3
2.2 Memory-Augmented Generation and Retrieval
- Memory-augmented methods retrieve story history through semantic, lexical, entity-based, or structured-state representations, while causal prerequisites of future goals remain underexplored.Examples include RecurrentGPT, SCORE, Open-Theatre, STORYTELLER, DOME, Narrative World Model, NstAgent, and NarraWorld.Source: §2.2 · p. 3
3 Empirical Analysis of Narrative Diversity and Retrieval
The analysis finds that narrative mechanism entropy generally declines later in generation and that conventional retrieval both misses some historically valuable subevents and is associated with higher trope repetition than human-curated memories. These findings motivate retrieval organized around causal dependencies on future narrative goals.
3.2 The Semantic–Narrative Relevance Gap and Trope Reuse
- 23.30% for Dense RAG and 16.67% for DOME are the average repetition rates, versus 7.14% for human-curated memories and 5.56% in the original script.The comparison measures reuse of narrative tropes from historical episodes in Episodes 21–30 continuations.Source: Figure 2 · p. 4, §3.2 · p. 5
3.1 Declining Narrative Mechanism Entropy
- Narrative mechanism entropy generally declines after episode 20, especially near the end; Dense RAG remains below BASE throughout the second half, and DOME does so at most later checkpoints.The analysis compares BASE, Dense RAG, and DOME using 32 sampled candidate next events at checkpoints across 40-episode scripts.Source: Figure 1 · p. 4, §3.1 · p. 4
- The analysis defines a core progression mechanism as the means and causal process behind an event’s first meaningful change, then clusters similar candidate events to quantify concentration.Lower entropy indicates greater concentration on fewer mechanisms; checkpoint entropy is averaged across five scripts with equal weight per script.Source: §3.1 · p. 4
3.2 The Semantic–Narrative Relevance Gap and Trope Reuse
- The pipeline’s overview moves from a four-stage coarse outline through trope-guided evolutionary search to episode outlines and a future-aware dependency tree for memory-guided writing.Episodes attach to goal nodes to guide retrieval; agents plan scenes, generate scripts, and update memories after each episode.Source: Figure 3 · p. 5
- Dense RAG and DOME each retrieve only 7 of 27 historical subevents identified as valuable for continuation, a recall of 25.93%.The case study uses Episodes 1–20 as history and Episodes 21–30 synopses as future goals, retrieving 30 memories per method.Source: §3.2 · pp. 4–5
- Semantic similarity can miss causal prerequisites for future goals while retrieving events that contribute little to those goals, motivating dependency-guided memory retrieval.The analysis distinguishes surface-level matching from the narrative function of historical information.Source: §3.2 · p. 5
4 Method
DraMaze links diverse causal-path planning with future-aware memory to support long-form scripts that are both varied and coherent. Its pipeline combines evolutionary search with backward goal dependencies and state tracking.
4.1 Pipeline Overview
- DraMaze develops each premise through hierarchical planning and sequential episode writing, connecting diverse plot planning with narrative execution.Planning agents create plot points and hooks, then outline four drama stages; episodes proceed through memory retrieval, scene planning, script writing, and memory updates.Source: §4.1 · pp. 5–6
4.2 Evolutionary Tree Search for Outline Refinement
- Evolutionary tree search expands outline anchors into causal event sequences, favoring novelty and coherence while mutation broadens the search space.Each node generates four candidates, selects the top two for further expansion, and mutates the leading candidate with probability 0.25 while preserving its target anchor.Source: §4.2 · p. 6
4.3 Future-Aware Narrative Memory
- The Backward Narrative Dependency Tree indexes events, clues, and foreshadowing by future goals, while world-state memory tracks characters, relationships, and assets.The tree is built backward from the planned ending; prior records linked to episode goals and selected prerequisite subtrees are retrieved for writing, and completed episodes update world states.Source: §4.3 · pp. 7–8, Figure 4 · p. 7
5 Evaluation
MiDramaBench adapts source texts into bilingual microdrama ideas across script lengths. Its evaluation combines multi-agent logic and quality scoring with causal-pattern metrics for repetition and diversity.
5.1 Benchmark Construction
- MiDramaBench adapts deduplicated PG-19 fiction and nonfiction sources from three word-count ranges into reviewed microdrama ideas targeting 20-, 40-, and 60-episode dramas, retaining ideas that pass quality review.The source ranges are [30k, 50k), [70k, 90k), and [110k, 130k) words, respectively; k-means clustering produces ten groups, with up to 30 review-ranked ideas retained per length–cluster group.Source: §5.1 · pp. 8–9
5.2 Logic and Quality Evaluator
- The evaluator assesses script Logic and Quality through fine-grained criteria, combining evidence-ledger scores with holistic category scores.Three reviewers record issues and strengths; an Audit Agent validates entries and an Arbitration Agent resolves remaining disputes. Final category scores weight ledger and holistic scores equally.Source: §5.2 · p. 9
5.3 Narrative Homogeneity Metrics
- Causal-pattern metrics track repeated four-slot chains, repeated three-slot triples, and normalized conditional entropy as indicators of narrative homogeneity.Extracted chains are canonicalized so functionally equivalent expressions share labels; higher repetition and lower entropy indicate greater homogeneity.Source: §5.3 · pp. 10–11
6 Experimental Results
Across script lengths and backbone models, DraMaze improves logic and quality while reducing causal-pattern repetition and maintaining greater narrative diversity. Ablations report distinct contributions from future-aware memory, tree search, and the multi-agent pipeline.
6.1 Setup
- Experiments use 30 ideas at each of 20, 40, and 60 episodes, repeat runs three times, and compare DraMaze with adapted DOME and StoryWriter baselines.Qwen3.6-27B is the main generation backbone; GLM5.3-Flash and DeepSeek-V4-Pro are used for generalization experiments.Source: §6.1 · p. 11
6.2 Main Results
- DraMaze’s Logic scores are 82.6, 81.3, and 82.5 at 20, 40, and 60 episodes, with its margin over the strongest baseline growing from 4.2 to 8.5 points.The authors also report consistently high Quality scores, indicating that improved coherence accompanies sustained creative quality.Source: §6.2 · p. 12
- At 60 episodes, 3-gram repetition is 14.6% for DraMaze versus 21.5% for the strongest baseline, and DraMaze has the lowest repetition and highest entropy across lengths.Structural repetition increases with script length, but DraMaze maintains greater narrative diversity.Source: §6.2 · p. 12
- On DeepSeek-V4-Pro and GLM5.3-Flash, DraMaze improves Logic by 4.8–11.0 points and Quality by 3.5–5.5 points over the strongest baselines.The authors also report consistently lower repetition and higher conditional entropy across these backbones.Source: §6.2 · p. 12
6.3 Ablation Study
- Removing future-aware memory lowers Logic/Quality from 81.3/86.3 to 74.5/81.2; removing tree search raises 4-gram repetition from 4.65% to 7.13%.Direct generation further lowers Logic/Quality to 70.3/76.8.Source: §6.3 · p. 12
7 Conclusion
DraMaze addresses narrative homogenization by pairing diverse plot exploration with memory retrieval guided by future goals. The paper reports consistent improvements in logic, creative quality, and diversity across multiple backbones.
7 Conclusion
- DraMaze combines trope-guided evolutionary tree search, future-goal-organized narrative memory, and dynamic world-state tracking.The pipeline explores diverse causal paths and constructs a Backward Narrative Dependency Tree from the planned ending.Source: §7 · p. 12
- MiDramaBench provides a bilingual benchmark for evaluating script logic and creative quality, alongside causal-pattern metrics for repetition and diversity.Source: §7 · p. 12
- Experiments on 20-, 40-, and 60-episode scripts across multiple backbone LLMs show consistent improvements in logic, creative quality, and diversity.Ablations validate key component effectiveness.Source: §7 · p. 12
A Details of the Mechanism Entropy Experiment
The experiment samples next-event continuations at fixed checkpoints, evaluates whether their progression mechanisms are similar, then clusters them to compute mechanism entropy.
A.1 Checkpoint Sampling
- At ten checkpoints in 40-episode scripts, the experiment freezes the premise, episode index, recent context, and retrieval results before sampling 32 next-event candidates per group.The sampled candidates are not appended to the script or used to condition later samples.Source: §A.1 · p. 14
A.2 Mechanism Similarity Evaluation
- Events are compared by their key causal means and resulting situational changes, with pairwise similarity accepted only when judgments agree in both input orders.Differences in characters, locations, or objects are incidental when the causal mechanism is preserved.Source: §A.2 · pp. 14–15
A.3 Clustering and Entropy Computation
- Deterministic complete-link clustering merges only mutually similar events, computes entropy from cluster sizes over parseable events, and averages checkpoint scores across five scripts.Groups are clustered and scored independently.Source: §A.3 · p. 15
B Benchmark Data Example
The appendix illustrates MiDramaBench's structured drama ideas and shows how a backward dependency tree connects episode outlines to narrative goals.
B Benchmark Data Example
- A representative MiDramaBench idea is structured around genre, core plot, characters, and world setting, establishing shared narrative grounding while allowing plot development.The example is an English microdrama idea.Source: B Benchmark Data Example · p. 15, Table 4 · p. 16
C Case Study: BNDT Construction and Episode Attachment
- For a 40-episode wartime drama, the Backward Narrative Dependency Tree is built from the ending, with child goals serving as prerequisites for parent goals.The figure reproduces all 28 generated goal nodes.Source: C Case Study: BNDT Construction and… · p. 16
- Episodes attach to goals they set up, advance, or complete; a single episode can advance one goal while setting up another.Episode 32 advances the plan to infiltrate Sterling's command post, while Episode 19 advances operational value and sets up Jack's acceptance of violence.Source: C Case Study: BNDT Construction and… · p. 16
D Details of Logic and Quality Evaluation
The appendix details the Logic and Quality evaluation rubrics and the causal-chain procedure used to measure trope repetition and diversity.
D Details of Logic and Quality Evaluation
- Figure 7 presents a 28-goal dependency tree for a 40-episode wartime drama, where each child goal is a prerequisite for its parent.Source: D Details of Logic and Quality… · p. 16, Figure 7 · p. 17
D.1 Evaluation Criteria
- Logic uses seven scoring categories and Quality five, with shared dimension-specific criteria, severity anchors, scope boundaries, and evidence requirements.Findings cite named checks and episode-scene locations; contradictions require evidence from both passages.Source: §D.1 · p. 17, Table 5 · p. 18
D.2 Detailed Rating Anchors for Character State
- Character-state ratings cover life, identity or allegiance, identity consistency, ability, personality or motivation, and relationship continuity, with severity tied to evidence and impact.Source: §D.2 · pp. 18–19
Abrupt Change of Personality or Motivation
- Abrupt personality, motivation, and relationship shifts are graded by whether major reversals have sufficient events, psychological setup, emotional buildup, and resolution.Source: Abrupt Change of Personality or… · p. 19
E Details of Trope Homogeneity Evaluation
- The trope homogeneity evaluator extracts and normalizes C–M–O–R causal chains, then computes repetition rates and conditional entropy from their category sequences.Source: E Details of Trope Homogeneity… · p. 20
E.1 Causal-Chain Extraction
- The reference extraction uses seven-episode windows with two-episode overlap, preserving mechanisms, causal direction, traceable evidence, and episode locations while deduplicating repeated events.Source: §E.1 · p. 20
E.2 Normalization
- A global taxonomy groups slot expressions by causal mechanism, behavioral goal, direction of change, and subject–object relations; substitution checks distinguish equivalent mechanisms from shared themes.The resulting categories include definitions, inclusion and exclusion boundaries, and original member phrases.Source: §E.2 · p. 20
E.3 Metric Computation
- Metrics are computed programmatically from normalized chains, and conditional entropy averages normalized M|C, O|M, and R|O summaries using a support threshold of τ=3.Source: §E.3 · pp. 20–21
F Additional Experimental Details and Results
The additional results provide full cross-model and ablation tables beyond the main-text aggregates. They cover 40-episode scripts and report means and sample standard deviations across three runs.
F Additional Experimental Details and Results
- Table 6 reports cross-model generalization results for 40-episode scripts using DeepSeek-V4-Pro and GLM5.3-Flash.Bold marks the best average and homogeneity scores within each model group.Source: Table 6 · p. 21, F Additional Experimental Details and… · p. 22
- Table 7 ablates Future-Aware Memory and Evolutionary Tree Search for Outline Refinement on Qwen3.6-27B, retaining the multi-Agent Pipeline in removal variants.Its Direct baseline generates scripts without the multi-Agent Pipeline.Source: Table 7 · p. 21
- Tables 6 and 7 provide the complete cross-model and ablation results corresponding to Tables 2 and 3.The main-text tables report aggregate Logic and Quality scores alongside homogeneity metrics.Source: F Additional Experimental Details and… · pp. 21–22
- The full tables report all seven Logic and five Quality categories, with means and sample standard deviations across three runs of 40-episode scripts.This provides finer-grained performance detail across evaluation criteria.Source: F Additional Experimental Details and… · p. 22