betaFeedback

WorksComputer Science

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

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.

Introduction

Microdramas have rapidly gained popularity worldwide as a form of fast-paced, serialized entertainment. LLM-based script generation offers a promising approach to meeting the growing demand for efficient and scalable content creation. However, maintaining narrative coherence and diverse plot progression across dozens of episodes remains challenging, despite LLMs’ ability to produce fluent, well-structured scripts. As generation progresses, models may reuse familiar plot patterns despite changes in characters and events. For example, relationships may repeatedly advance through misunderstandings and clarifications. In microdramas, which are marked by rapid plot progression and twists, such repetition is especially noticeable and may make suspense-building turns predictable, thus weakening audience engagement.

Prior work reports improved individual quality but reduced diversity across AI-assisted works, alongside recurring plot elements in generated stories [4] (Xu et al. 2025b). This may reflect generation preferences [6] and accumulating narrative constraints. Our experiment also shows declining narrative mechanism entropy in the script’s second half, with candidate continuations concentrating on fewer progression mechanisms (§ 3.1).

To enhance long-form generation, recent methods incorporate hierarchical planning, history summarization, and memory retrieval. DOME combines dynamic outlining with a temporal knowledge graph, while StoryWriter uses multi-agent planning and dynamic history compression (Wang et al., 2024; Xia et al., 2025). While these approaches provide relevant historical context through memory management, they do not necessarily mitigate narrative homogenization. Actually, studies of LLM agents suggest that similar experiences and semantically relevant memories may shape subsequent decisions by encouraging models to repeat past behaviors or fixate on prior reasoning patterns [16] (Xiong et al. 2025b). Mechanism entropy still declines in later episodes with dense retrieval or DOME-style graph memory (§ 3.1). In our continuation case study, both methods retrieve only 25.93% of key historical information, and their continuations repeat past causal patterns more often than those generated with human-curated memories (§ 3.2). These results suggest a gap between semantic relevance and narrative value: retrieval related to the current situation or characters may reintroduce previously used solutions while missing prerequisites for subsequent writing, weakening cross-episode coherence.

These observations highlight the need to explore diverse plot trajectories without repeating past progression strategies, while preserving essential historical information. Since the value of historical information depends on its contribution to future plot development, future narrative goals should guide memory retention and retrieval.

Based on this insight, we propose a multi-agent microdrama script generation pipeline, which integrates future goals into plot planning and episode-by-episode writing. During planning, it uses evolutionary tree search to explore diverse causal paths toward key narrative anchors, leveraging trope memory to discourage repeated progression mechanisms and mutation to broaden the search space. Before episodic writing, the pipeline constructs a Backward Narrative Dependency Tree by tracing intermediate goals and their prerequisites backward from the planned ending. During writing, this tree guides the selection and organization of key events and details, clues, and foreshadowing that support future developments, and enables retrieval according to the current episode’s goals. Meanwhile, continually updated world-state memory maintains consistent states of characters, relationships, and assets. By combining diverse plot planning with historical retrieval guided by future needs, the pipeline supports varied, causally coherent progression toward long-term narrative goals. Our main contributions are:

  • We reveal declining narrative mechanism diversity in long-form generation and show that conventional semantic retrieval may reinforce trope reuse while missing historical information essential to future plot developments.

  • We propose DraMaze, a multi-agent pipeline that integrates evolutionary plot planning with future-aware narrative memory and dynamic state tracking to generate diverse yet coherent microdrama scripts.

  • We introduce MiDramaBench for evaluating script logic and creative quality, together with causal-pattern metrics for quantifying narrative repetition and diversity across script lengths. Cross-model experiments and ablations validate the pipeline.

Related Work

Long-Form Creative Writing Frameworks

LLMs have been widely explored for creative writing, yet maintaining narrative quality in long-form generation remains challenging. Re3 (Yang et al., 2022) uses recursive prompting, candidate reranking, and revision, while DOC (Yang et al., 2023) employs detailed hierarchical outlines and controlled generation to improve narrative coherence. DOME (Wang et al., 2024) integrates dynamic hierarchical outlining with temporal knowledge graph memory to maintain long-range narrative consistency. Other relevant studies further explore specialized agent collaboration and adaptive planning. StoryWriter (Xia et al., 2025) coordinates outlining, planning, and writing agents for cross-chapter event organization and dynamic history compression. CritiCS [1] iteratively refines story plans and drafts through collective critic feedback, while WriteHERE (Xiong et al., 2025a) recursively decomposes writing into retrieval, reasoning, and composition tasks for adaptive generation. HoLLMwood [2] employs writer–editor–actor collaboration and role-playing in script generation. However, narrative repetition and homogenization persist in long-form generation [2, 4], despite advances in plot planning and coherence.

Memory-Augmented Generation and Retrieval

Maintaining coherence in long-form generation requires effective organization and retrieval of narrative history. RecurrentGPT (Zhou et al., 2023) maintains short- and long-term memories with semantic retrieval. SCORE (Yi et al., 2025) combines hierarchical summaries and state tracking with lexical-semantic retrieval, whereas Open-Theatre (Xu et al., 2025a) employs hierarchical memory and relevance-, importance-, and recency-based retrieval. Some studies also explore structured narrative memory. STORYTELLER [7] builds an entity knowledge graph from subject–verb–object events for entity-based historical retrieval, while DOME (Wang et al., 2024) employs temporal knowledge graph memory with entity-based retrieval and semantic filtering. Narrative World Model [13] represents evolving story states through a narratology-grounded temporal graph for multi-hop retrieval, while NstAgent [15] incrementally tracks character states, completed events, and unresolved narrative commitments. NarraWorld [14] organizes world facts, character beliefs, ongoing developments, and possible continuations into interconnected memory views for situation-conditioned retrieval. While these methods further strengthen narrative consistency, organizing memories around the causal prerequisites of future narrative goals remains underexplored.

Empirical Analysis of Narrative Diversity and Retrieval

Declining Narrative Mechanism Entropy

As scripts progress, LLMs may increasingly rely on similar mechanisms to advance the plot, despite changes in characters, settings, and events. Semantic retrieval may reinforce this tendency by recalling past events resembling the current situation or involving the same characters, encouraging reuse of familiar narrative devices and causal progression patterns. We quantify this concentration through the entropy of candidate narrative mechanisms.

Using Qwen3.6-27B, we compare three strategies: full-history generation without retrieval (BASE), dense semantic retrieval of past events (Dense RAG), and DOME-style graph retrieval with semantic filtering (DOME). Each generates 40 episodes from the same premise. At generation checkpoints, we fix the context and retrieval results, then sample 32 candidate next events, each describing a main action and its immediate consequence.

We define an event’s core progression mechanism as the means and causal process behind its first meaningful change to the situation. Qwen3.8-27B judges whether each pair of candidate events shares similar mechanisms. Deterministic greedy complete-link clustering then groups the candidates into mechanism clusters. For NN parseable events in KK clusters, narrative mechanism entropy is

H=−∑k=1Kpklog⁡2pk,pk=∣Ck∣N,(1)H = -\sum_{k=1}^{K} p_k \log_{2} p_k,\qquad p_k = \frac{\lvert C_k\rvert}{N}, \tag*{(1)}

where CkC_k is the kk-th cluster. Lower entropy indicates greater concentration on fewer mechanisms. At each checkpoint, we report the mean entropy across five scripts, with equal weight assigned to each script.

Figure 1 shows that narrative mechanism entropy generally declines after episode 20, especially near the end. Dense RAG remains below BASE throughout the second half, while DOME does so at most later checkpoints. These results suggest that conventional retrieval fails to prevent, and may reinforce, concentration on fewer narrative mechanisms, highlighting the need for more diverse yet coherent plot exploration.

Figure 1. Narrative mechanism entropy across generation checkpoints.

The Semantic–Narrative Relevance Gap and Trope Reuse

We further investigate whether semantic retrieval can identify historical information essential to future plot development and whether retrieved memories encourage reuse of existing narrative patterns.

We select a complete human-written microdrama script, using Episodes 1–20 as history and the synopses of Episodes 21–30 as future narrative goals. Dense RAG and DOME-style knowledge graph retrieval each retrieve the top 30 historical memories. We manually identify 27 historical subevents with narrative value for the continuation as references. Both methods retrieve only 7 of these 27 subevents, yielding a recall of 25.93%. This suggests that semantic or entity-level relevance does not necessarily reflect the causal dependencies between historical events and future narrative goals.

We further generate Episodes 21–30 under identical settings using memories retrieved by each method or human-curated memories, with the original human-written script as a reference. We represent narrative tropes as Cause–Method–Outcome and Method–Outcome–Response triples and measure their reuse from the historical episodes in the continuations. As shown in Figure 2, average repetition rates are 23.30% for Dense RAG and 16.67% for DOME, compared with 7.14% for human-curated memories and 5.56% in the original script. These results suggest that automatic retrieval may not only overlook historical information essential to future developments but also encourage the reuse of existing conflict-resolution strategies and causal mechanisms.

Figure 2. Historical trope repetition rates in different retrieval settings.

Together, these observations reveal the persistent reuse of narrative mechanisms and a mismatch between semantic relevance and narrative function: semantically similar events may contribute little to current narrative goals, while essential causal prerequisites may be overlooked due to semantic differences. This motivates a shift from surface-level semantic matching toward retrieval guided by causal dependencies on future narrative goals, preserving necessary plot prerequisites while reducing unnecessary reuse of existing narrative patterns.

Method

Pipeline Overview

As shown in Figure 3, our multi-agent pipeline develops an idea into a microdrama script through hierarchical planning and sequential writing. Given the premise, world setting, and characters, planning agents design key plot points and dramatic hooks, then build a coarse outline spanning Setup, Development, Climax, and Resolution. Evolutionary tree search expands its anchors into a detailed event sequence (§ 4.2), which is allocated across episodes and developed into episode outlines. A memory-planning agent then works backward from the ending to construct a Backward Narrative Dependency Tree (BNDT) for future-aware memory (§ 4.3). For each episode, agents retrieve narrative memories, plan scenes, write the script, and update memory from the completed episode. Both mechanisms thus connect diverse narrative planning with consistent narrative execution.

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.

Evolutionary Tree Search for Outline Refinement

The coarse outline specifies where the story goes, while refinement determines how it gets there. To avoid reusing solutions across different conflicts, we explore causal developments and explicitly evaluate their novelty against previously used mechanisms.

Trope Memory. A trope memory B\mathcal{B} stores causal patterns extracted from committed event sequences as cause–method–outcome–response. During extraction, we remove specific character names, locations, and props to recognize recurring mechanisms despite surface differences. For example, bribing a guard with cash or jewelry uses the same mechanism to gain cooperation. Patterns accumulate across all four drama stages and subsequent candidates are compared against this memory to score novelty.

Evolutionary Tree Expansion. For each narrative anchor in the coarse outline, we grow a local search tree from the preceding plot. Each expanded node generates four candidate next-event nodes. We score their novelty by comparing the local causal tropes they form with preceding events against trope memory B\mathcal{B}, then select the top two for further expansion based on novelty and local narrative coherence.

Inspired by evolutionary algorithms, we introduce mutation with probability pmut=0.25p_{\mathrm{mut}} = 0.25 at each expanded node. It explicitly restructures the causal links in the highest-scoring candidate’s local trope while preserving the target anchor. The mutated event forms an additional branch to escape existing causal patterns and broaden the search space.

Branches terminate upon reaching the anchor or search limit. Among terminal chains C\mathcal{C}, we select

S(c)=0.4C(c)+0.3N(c)+0.3F(c),c∗=arg max⁡c∈CS(c),(2)\begin{aligned} S(c) &= 0.4C(c) + 0.3N(c) + 0.3F(c),\\ c^{*} &= \underset{c \in\mathcal{C}}{\operatorname{arg\,max}} S(c), \tag*{(2)} \end{aligned}

where CC, NN, and FF measure narrative consistency, trope novelty relative to the accumulated trope memory B\mathcal{B}, and the feasibility of reaching the targeted anchor, respectively. The final selected chain extends the plot and updates B\mathcal{B} with its causal patterns to guide subsequent search.

Future-Aware Narrative Memory

The value of historical information depends on its contribution to future plot development, so we use future narrative goals to guide memory retention and retrieval. Figure 4 illustrates this process.

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.

Backward Narrative Dependency Tree. The planning agent first extracts the final outcome from the episode outlines and uses it as the root of the BNDT F=(V,E)\mathcal{F} = (V, E). Working backward and guided by the episode outlines, it recursively decomposes this outcome into intermediate goals and further foundational setups. Each child specifies a prerequisite for its parent goal. Then, each episode is mapped to the goals Qt⊆VQ_t \subseteq V that it sets up, advances, or completes. The tree provides a goal-based index for organizing and retrieving narrative memories during script generation. An example can be seen in Appendix § 10.

Narrative Memory Construction. After each episode is written, a memory agent extracts, guided by the established BNDT, critical events & details, clues with long-term relevance, and foreshadowing requiring future payoff. It retains only information with future narrative value for BNDT goals and links each memory record to the corresponding goal nodes, recording its contribution and long-term narrative importance score. These links enable retrieval by narrative goal during subsequent episode generation.

Narrative Memory Retrieval. Before writing episode tt, we retrieve historical memories through its assigned BNDT goal nodes QtQ_t. First, we collect all prior records linked to at least one of QtQ_t:

Rtdir={m∈M<t:Λ(m)∩Qt≠∅}.(3)\mathcal{R}^{\mathrm{dir}}_t = \{m \in\mathcal{M}_{<t} : \Lambda(m) \cap Q_t \ne\emptyset\}. \tag*{(3)}

Here, M<t\mathcal{M}_{<t} contains memories from prior episodes, and Λ(m)\Lambda(m) denotes the BNDT goal nodes supported by record mm. The intersection selects memories sharing a narrative goal with the current episode.

Next, we traverse subtrees beneath these goal nodes, as a few key events tied to prerequisite subgoals may retain narrative value for current episode’s goals. From events and clues linked to these descendant nodes, we retrieve up to ten with the highest long-term narrative importance scores to form Rtsub\mathcal{R}^{\mathrm{sub}}_t, excluding already retrieved records. Foreshadowing is retrieved only through direct goal links, as it should have paid off at completed prerequisite nodes. Their union, Rt=Rtdir∪Rtsub\mathcal{R}_t = \mathcal{R}^{\mathrm{dir}}_t \cup\mathcal{R}^{\mathrm{sub}}_t, provides narrative memories for the current goals.

World-State Memory. World-state memory maintains the current states of characters, inter-character relationships, and assets. Before each episode is generated, its outline is used to retrieve relevant state records through textual relevance and entity matching. These records are provided alongside narrative memories as context for scene planning and script generation. After generation, changes established in the completed script are used to update the stored states, supporting consistent character and asset states across episodes.

Evaluation

Benchmark Construction

We construct a bilingual microdrama generation benchmark from PG-19 (Rae et al. [12]), a public dataset of Project Gutenberg fiction and nonfiction, as illustrated in Figure 6. We select sources in three length ranges, [30k,50k)[30k, 50k), [70k,90k)[70k, 90k), and [110k,130k)[110k, 130k) words, targeting 20-, 40-, and 60-episode dramas, respectively, and deduplicate by content hashes and normalized titles. These ranges target 700–1,000 English words per episode, allowing for the condensation of source narration and description during script adaptation. Then, we use Qwen3.8-27B to score each source’s dramatic adaptation potential based on narrative continuity, character goals, actionable conflict, and episodic expandability. For retained sources, the model extracts core plot developments, character relationships, and world settings from the full texts, then adapts them into open-ended microdrama ideas by restructuring the central conflict for tightly paced episodes with escalating tension, while preserving core narrative. It then scores the generated ideas on causal coherence, dramatizability, episodic expandability, and creative openness, retaining those that pass quality review after any necessary revision.

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.

For genre stratification, we encode accepted ideas’ genre labels with BGE-large-en-v1.5 and apply k-means clustering to obtain ten groups. We provide both English and Chinese versions of all ideas. Within each length–cluster group, we retain up to 30 ideas ranked by review scores, yielding 899 candidates overall. We also construct a stratified evaluation subset by sampling three distinct ideas from the top 20 per group: two for English generation and one for Chinese generation. This yields 30 inputs per length setting, with balanced genre-cluster coverage within each language.

Logic and Quality Evaluator

We develop an LLM-based multi-agent framework for evaluating microdrama scripts, with criteria co-designed with drama professionals and informed by narrative consistency error taxonomies [8], reader-derived story quality criteria (Yang and Jin, 2025), and drama-specific evaluation dimensions [9]. It assesses two dimensions:

  • Logic: causal chain, character state, knowledge state, resource state, rules and space-time, setup and payoff, repetition detection.

  • Quality: character quality, dramatic engagement, plot-planning quality, episodic script quality, setting quality.

Each category is further divided into fine-grained criteria for identifying narrative flaws and strengths. For example, within the character quality category, the insufficient character arc criterion identifies protagonist development that lacks clear turning points or consequential choices. Evaluation criteria and detailed rules are provided in Appendix § 11.

As shown in Figure 5, three independent LLM reviewers assess each dimension using predefined criteria, recording issues and strengths with textual evidence and episode locations in an Evidence Ledger. An Audit Agent validates and deduplicates entries and calibrates severity, while an Arbitration Agent resolves remaining disputes. Each category starts at 90 points, with S/A/B-level issues deducting 10/6/4 points and strengths adding 3/1/0 points, yielding the Ledger Score.

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

To balance fine-grained error detection with holistic assessments of narrative quality, we further introduce an independent Holistic Scorer which evaluates each category on a 0–100 scale using the same criteria. Category scores are combined as Sfinal=0.5SLedger+0.5SholisticS_{\mathrm{final}} = 0.5S_{\mathrm{Ledger}} + 0.5S_{\mathrm{holistic}}, then averaged within each dimension to obtain the final Logic and Quality Scores. This design combines the interpretability of evidence-based evaluation with the broad perspective of holistic assessment.

Narrative Homogeneity Metrics

Narrative Homogeneity Metrics

To further assess narrative repetition and diversity, we develop causal-pattern metrics based on extracted and normalized causal chains. We represent narrative patterns as four-slot causal chains: Cause, Method, Outcome, and Response (C–M–O–R), capturing event triggers, character actions, consequences, and subsequent reactions. An LLM extracts these chains from overlapping episode windows, abstracting away character names and locations while retaining causal mechanisms, source episodes, and plot-grounded explanations.

Since the same underlying causal mechanisms may be expressed differently across episodes, we perform global causal-pattern canonicalization, grouping functionally equivalent expressions into a unified taxonomy for each slot and relabeling the extracted chains accordingly. From these chains, we compute three metrics of narrative repetition and diversity.

(1) 4-gram Repetition Rate. The proportion of C–M–O–R occurrences beyond each distinct pattern’s first appearance.

(2) 3-gram Repetition Rate. The mean repetition rate of C–M–O and M–O–R triples.

(3) Conditional Entropy. Causal diversity is measured by support-weighted normalized conditional entropy:

H~(Y∣X)=1Nτ∑x∈Xτnxhx,(4)\widetilde{H}(Y \mid X) = \frac{1}{N_{\tau}} \sum_{x \in\mathcal{X}_{\tau}} n_x h_x, \tag*{(4)}
hx=H(Y∣X=x)log⁡∣Yx∣.h_x = \frac{H(Y \mid X=x)}{\log|\mathcal{Y}_x|}.

Here, nxn_x is the frequency of xx, Xτ={x:nx≥τ}\mathcal{X}_{\tau} = \{x : n_x \ge\tau\}, Yx\mathcal{Y}_x is its observed successor set, and Nτ=∑x∈XτnxN_{\tau} = \sum_{x \in\mathcal{X}_{\tau}} n_x. The final score is Havg=(H~(M∣C)+H~(O∣M)+H~(R∣O))/3H_{\mathrm{avg}} = (\widetilde{H}(M \mid C) + \widetilde{H}(O \mid M) + \widetilde{H}(R \mid O))/3. Higher repetition and lower entropy indicate greater narrative homogeneity.

Experimental Results

Setup

We evaluate on MiDramaBench’s stratified evaluation set of 30 ideas (20 English, 10 Chinese) at 20, 40, and 60 episodes. All experiments are repeated three times, with results reported as mean and sample standard deviation across runs.

We use DOME and StoryWriter as baselines, adapting them to episodic script generation while preserving their core workflows. DOME employs five-stage hierarchical outlining with temporal knowledge-graph memory, while StoryWriter uses event decomposition, global episode planning, writer–critic refinement, and history compression.

We use Qwen3.6-27B [10] as the backbone LLM for main generation experiments. We further conduct experiments with GLM5.3-Flash [5] and DeepSeek-V4-Pro [3] to assess the generalizability of our pipeline. For evaluation, we employ Qwen3.8-27B [11] as the underlying model at temperature 0.1.

Main Results

Table 1 presents the results of long-form script generation at 20, 40, and 60 episodes using Qwen3.6-27B. Our method consistently outperforms DOME and StoryWriter in script logic and quality across all episode lengths. Across the three length settings, DraMaze achieves similar Logic scores (82.6, 81.3, and 82.5 for 20, 40, and 60 episodes, respectively), while its margin over the strongest baseline increases from 4.2 to 8.5 points. The improvements extend across key aspects of long-range narrative consistency. Meanwhile, the consistently high Quality scores indicate that improved coherence is accompanied by sustained creative quality.

Ep.MethodCCCSKSRSRSTSPRDAvgCQDEPPESSQAvg4G↓3G↓Hˉ\bar{H} ↑
20DOME78.4 ±1.375.1 ±0.881.4 ±0.377.2 ±0.777.7 ±0.481.2 ±1.077.8 ±1.278.4 ±0.177.6 ±0.479.3 ±1.071.3 ±0.976.0 ±1.085.1 ±0.977.9 ±0.54.23 ±0.6011.0 ±0.150.876 ±0.0057
20StoryWriter74.7 ±0.374.5 ±0.776.2 ±0.472.8 ±1.072.6 ±1.177.9 ±0.483.5 ±0.576.0 ±0.383.5 ±0.884.0 ±0.377.3 ±1.082.2 ±0.286.8 ±0.182.8 ±0.14.34 ±0.6410.0 ±1.10.881 ±0.0095
20Ours81.5 ±0.280.4 ±1.082.6 ±0.480.1 ±0.381.3 ±0.483.7 ±1.288.5 ±0.582.6 ±0.287.3 ±0.787.7 ±0.682.9 ±0.385.2 ±0.787.6 ±0.386.1 ±0.31.35 ±0.275.51 ±0.640.939 ±0.0054
40DOME73.0 ±0.765.3 ±0.677.6 ±0.773.0 ±0.571.1 ±1.376.1 ±0.768.8 ±1.272.1 ±0.377.0 ±0.978.4 ±0.869.4 ±0.176.3 ±0.585.1 ±0.277.3 ±0.37.86 ±0.6517.3 ±1.00.883 ±0.0034
40StoryWriter73.5 ±1.270.8 ±0.876.3 ±0.872.3 ±0.571.2 ±1.177.4 ±1.481.1 ±0.274.6 ±0.885.9 ±0.284.3 ±0.378.1 ±1.082.0 ±0.187.4 ±0.283.5 ±0.36.65 ±0.5715.2 ±1.20.894 ±0.0025
40Ours80.4 ±0.679.2 ±0.381.8 ±0.578.0 ±1.279.0 ±0.683.7 ±0.787.3 ±0.981.3 ±0.788.5 ±0.388.1 ±0.682.4 ±0.885.0 ±0.487.7 ±0.186.3 ±0.43.08 ±0.419.22 ±0.630.929 ±0.0056
60DOME66.9 ±0.259.8 ±0.473.1 ±0.469.4 ±1.165.5 ±0.370.8 ±0.458.7 ±1.166.3 ±0.172.5 ±1.375.4 ±0.665.7 ±1.073.7 ±0.483.3 ±0.574.1 ±0.111.6 ±0.5224.5 ±0.720.866 ±0.0040
60StoryWriter72.8 ±1.169.3 ±1.276.4 ±0.772.9 ±1.371.4 ±1.377.0 ±0.778.1 ±0.374.0 ±0.884.4 ±1.483.4 ±0.676.5 ±0.681.3 ±0.586.6 ±0.282.4 ±0.49.75 ±0.9021.5 ±0.910.876 ±0.011
60Ours80.0 ±0.181.1 ±0.681.8 ±0.578.5 ±0.182.0 ±0.686.0 ±0.787.9 ±0.982.5 ±0.283.5 ±0.988.5 ±0.386.1 ±0.381.7 ±0.587.4 ±0.285.4 ±0.16.07 ±0.3214.6 ±0.640.912 ±0.0058

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 Hˉ\bar{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.

Ours also achieves the lowest 4-gram and 3-gram repetition rates and the highest conditional entropy across all lengths. At 60 episodes, 3-gram repetition drops to 14.6%, compared with 21.5% for the strongest baseline. Although structural repetition increases with script length, our method consistently maintains greater narrative diversity, demonstrating its effectiveness in mitigating narrative homogenization over extended generation.

Cross-model results in Table 2 and Table 6 further demonstrate the generalizability of our method. On DeepSeek-V4-Pro and GLM5.3-Flash, ours improves Logic scores by 4.8–11.0 points and Quality scores by 3.5–5.5 points over the strongest baselines, while consistently reducing repetition and increasing conditional entropy. These results confirm its effectiveness across different backbones.

ModelMethodLogic Quality Avg↑Logic Quality Avg↑Homogeneity 4G↓Homogeneity 3G↓Homogeneity Hˉ\bar{H}↑
DeepSeek-V4-ProDOME75.1 ±0.484.7 ±0.58.23 ±0.7017.2 ±0.430.853 ±0.012
StoryWriter73.3 ±0.784.7 ±0.66.74 ±0.6414.5 ±1.00.829 ±0.0068
Ours86.1 ±0.290.2 ±0.23.21 ±0.5710.2 ±0.510.915 ±0.0030
GLM5.3-FlashDOME83.3 ±0.489.0 ±0.38.03 ±0.9415.1 ±0.940.881 ±0.0062
StoryWriter84.8 ±0.488.6 ±0.27.08 ±1.114.9 ±0.900.870 ±0.010
Ours89.6 ±0.292.5 ±0.02.30 ±0.199.13 ±0.770.927 ±0.0050

Table 2. Cross-model generalization on 40-episode script generation using DeepSeek-V4-Pro and GLM5.3-Flash.

ModelMethodCCCSKSRSRSTSPRDAvgCQDEPPESSQAvg4G↓3G↓Hˉ↑\bar{H}\uparrow
DeepSeek-V4-ProDOME75.5 ±1.169.0 ±1.078.3 ±0.276.3 ±0.572.5 ±0.778.5 ±0.575.6 ±1.375.1 ±0.487.5 ±0.786.2 ±0.579.0 ±1.083.4 ±1.187.2 ±0.184.7 ±0.58.23 ±0.7017.2 ±0.430.853 ±0.012
DeepSeek-V4-ProStoryWriter74.2 ±0.869.4 ±1.178.2 ±0.375.0 ±0.872.1 ±0.576.5 ±0.867.8 ±1.173.3 ±0.787.7 ±1.386.0 ±0.877.2 ±0.685.9 ±0.386.7 ±0.384.7 ±0.66.74 ±0.6414.5 ±1.00.829 ±0.0068
DeepSeek-V4-ProOurs86.0 ±0.185.7 ±0.585.0 ±0.483.6 ±0.585.4 ±0.486.9 ±0.289.8 ±0.186.1 ±0.292.7 ±0.191.1 ±0.687.8 ±0.189.9 ±0.389.7 ±0.190.2 ±0.23.21 ±0.5710.2 ±0.510.915 ±0.0030
GLM5.3-FlashDOME83.9 ±0.880.2 ±0.683.8 ±0.382.2 ±0.781.7 ±0.684.9 ±0.786.5 ±0.783.3 ±0.492.0 ±0.890.1 ±0.386.4 ±0.787.0 ±0.689.4 ±0.189.0 ±0.38.03 ±0.9415.1 ±0.940.881 ±0.0062
GLM5.3-FlashStoryWriter85.1 ±0.583.8 ±0.584.7 ±0.684.2 ±0.383.2 ±0.686.7 ±0.285.9 ±0.784.8 ±0.492.4 ±0.389.9 ±0.185.0 ±0.486.5 ±0.789.1 ±0.388.6 ±0.27.08 ±1.114.9 ±0.900.870 ±0.010
GLM5.3-FlashOurs89.7 ±0.389.5 ±0.389.1 ±0.488.5 ±0.188.8 ±0.790.5 ±0.390.9 ±0.189.6 ±0.295.1 ±0.193.0 ±0.191.6 ±0.491.8 ±0.690.9 ±0.192.5 ±0.02.30 ±0.199.13 ±0.770.927 ±0.0050

Table 6. Cross-model generalization results on 40-episode script generation using DeepSeek-V4-Pro and GLM5.3-Flash. Results are reported as mean ± standard deviation. Best average and homogeneity scores within each model group are highlighted in bold.

Ablation Study

To assess component contributions, we conduct ablations on 40-episode scripts using Qwen3.6-27B. We progressively remove Future-Aware Memory and Evolutionary Tree Search while retaining the multi-agent pipeline. We also include a direct-generation baseline producing five episodes per iteration with full history as context.

As shown in Table 3, removing future-aware narrative memory reduces Logic/Quality from 81.3/86.3 to 74.5/81.2, increases 3-gram repetition from 9.22% to 12.2%, and lowers entropy from 0.929 to 0.914. These results support the contribution of future-aware narrative memory to coherence and reduced causal-pattern repetition within the full pipeline. Further removing tree search raises 4-gram repetition from 4.65% to 7.13% and lowers entropy to 0.887, highlighting its role in narrative diversity. Direct generation further reduces Logic/Quality to 70.3/76.8, which demonstrates the value of hierarchical planning and multi-agent coordination.

MethodLogic Quality Avg↑Logic Quality Avg↑Homogeneity 4G↓Homogeneity 3G↓Homogeneity Hˉ\bar{H}↑
Ours81.3 ±0.786.3 ±0.43.08 ±0.419.22 ±0.630.929 ±0.0056
w/o Future-Aware Mem74.5 ±0.381.2 ±0.34.65 ±0.4112.2 ±0.860.914 ±0.0084
w/o Future-Aware Mem & Tree Search72.1 ±0.381.2 ±0.47.13 ±1.216.4 ±0.810.887 ±0.0091
w/o Multi-Agent Pipeline (Direct Generation)70.3 ±0.176.8 ±0.26.87 ±0.1416.8 ±0.680.887 ±0.012

Table 3. Ablation results on 40-episode script generation using Qwen3.6-27B. We progressively remove Future-Aware Memory and Evolutionary Tree Search while retaining the original Agent Pipeline. Direct removes the Multi-Agent Pipeline entirely.

Conclusion

This work investigates narrative homogenization in long-form script generation, revealing a gap between semantic relevance and future narrative needs. To address these challenges, we propose DraMaze, combining trope-guided evolutionary tree search, narrative memory organized around future goals, and dynamic world-state tracking. We further introduce MiDramaBench for bilingual script generation and evaluation of logic and creative quality, together with causal-pattern metrics for narrative repetition and diversity. Experiments on 20-, 40-, and 60-episode scripts across multiple backbone LLMs demonstrate consistent improvements in logic, creative quality, and diversity, while ablations validate key component effectiveness. These findings highlight the value of combining diverse plot exploration with future-goal-guided memory retrieval for coherent and diverse long-form generation.

Details of the Mechanism Entropy Experiment

Checkpoint Sampling

We probe each 40-episode script after episodes t∈{4,8,12,16,20,24,28,32,35,39}t \in\{4,8,12,16,20,24,28,32,35,39\}. At each checkpoint, we freeze the premise, episode index, recent context and retrieval results, then independently sample 32 next-event candidates per group. Each candidate describes a character’s main action and its direct consequence in approximately 80–150 tokens. Candidates are not truly appended to the script or used to condition subsequent samples. Both episode generation and event sampling use Qwen3.6-27B.

Mechanism Similarity Evaluation

For each event, the evaluator independently identifies the first substantive progression unit: the key means that address the opening predicament or advance the ongoing main action, together with the resulting change in the situation.

The comparison focuses on how events overcome or create obstacles, reveal or conceal information, shift character incentives, or alter control relationships. Differences in characters, locations, or objects are incidental when they preserve the causal mechanism.

Two events are judged similar when their identified progression units share the same key causal means, and dissimilar when those means clearly differ. Each pair is evaluated in both input orders and treated as similar for clustering only when both judgments agree on similarity. All within-group pairwise evaluations are completed before clustering.

Clustering and Entropy Computation

Deterministic complete-link clustering. Let sij=1s_{ij}=1 if events ii and jj are stably similar, and sij=0s_{ij}=0 otherwise. Each parseable event initially forms a singleton cluster, ordered by its original sample index. Two clusters CaC_a and CbC_b can merge only when

sij=1for every i∈Ca, j∈Cb.(5)s_{ij}=1 \quad\text{for every } i \in C_a,\ j \in C_b. \tag*{(5)}

We scan cluster pairs in lexicographic order of their positions and merge the first eligible pair. After each merge, members are sorted by sample index, clusters are reordered lexicographically by their member-index sequences, and scanning restarts. The procedure terminates when no eligible pair remains. This rule ensures that all events within a cluster are pairwise stably similar. For example, if pairs (0, 1) and (0, 2) are stably similar but (1, 2) is not, the algorithm first merges events 0 and 1, yielding {{0,1},{2}}\{\{0, 1\}, \{2\}\}.

Group-level entropy. For each sampling group, we compute entropy from the cluster-size distribution using the definition in § 3.1. The denominator is the number of parseable events. Groups are clustered and scored independently.

Aggregation across scripts. Let Hs,t(m)H_{s,t}^{(m)} denote the entropy computed from the 32 next-event candidates sampled for script ss under method mm at checkpoint tt. We average these entropy values across the five scripts:

Hˉt(m)=1S∑s=1SHs,t(m),S=5.(6)\bar{H}_{t}^{(m)}=\frac{1}{S}\sum_{s=1}^{S}H_{s,t}^{(m)}, \qquad S=5. \tag*{(6)}

Benchmark Data Example

Table 4 presents a representative generated microdrama idea from MiDramaBench. Each idea comprises four fields: genre, core plot, characters, and world setting, which collectively establish the central conflict, character motivations and relationships, and relevant setting constraints. As illustrated by the example, these structured inputs provide sufficient narrative grounding for long-form generation while leaving room for diverse drama plot developments. This allows different generation methods to explore and organize episode-level narratives from a common starting point.

FieldInput idea
GenreCorporate thriller; family drama.
Core plotA 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.
CharactersThe 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 SettingA 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.

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

Case Study: BNDT Construction and Episode Attachment

We illustrate BNDT construction and episode attachment using a 40-episode wartime microdrama centered on Jack and Harry. Figure 7 presents the complete dependency tree constructed backward from the final outcome, where each child goal represents a prerequisite for its parent. Episodes are attached to goal nodes according to whether they set up, advance, or complete the corresponding goals.

A complete backward narrative dependency tree (BNDT) generated for a 40-episode wartime drama

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.

For example, the outline for Episode 26 has Vandever declare the boys “dead to the army” and send them out of camp with Bill and Jake. This completes FM011 (“The boys are stripped of their protected status and exiled by Vandever”). In contrast, Episode 32 outlines Jack impersonating an officer and bluffing past a sentry with a fabricated order from Sterling. This advances FM005 (“The boys infiltrate Sterling’s command post and secure the battle plans”), since they gain entry, but have yet to obtain the plans. An episode may also serve multiple goals. Episode 19 outlines “Jack kills a sentry to help his companions escape, but is shaken by his own violence.” It advances FM018 (“The boys prove their operational value through repeated unauthorized successes”) while setting up FM015 (“Jack accepts his hardened identity and the necessity of his violent actions”).

Details of Logic and Quality Evaluation

This section presents the evaluation criteria and representative rating anchors that guide evidence identification, severity assessment, and scoring in Logic and Quality evaluation.

Evaluation Criteria

Evaluation Criteria

Logic comprises seven scoring categories, while Quality comprises five. Table 5 summarizes their principal checks. Each category is further divided into fine-grained criteria that guide the identification of narrative issues and strengths and specify the evidence required for their assessment. All agents receive the same dimension-specific rubric, including the applicable severity anchors and scope boundaries.

DimensionCategoryPrincipal checks
LogicCausal chainPrerequisite information, action motivation, causal support for key turns, and justification of crisis resolution.
Character stateContinuity of life, identity, allegiance, abilities, personality, motivation, and relationships.
Knowledge stateInformation acquisition, knowledge sources, retention of established facts, and the scope of secret disclosure.
Resource stateOwnership and transfer, depletion and restoration, evidence provenance, and support for resource increases.
Rules and space-timeEstablished rules, preparation of new rules, temporary exceptions, event chronology, and spatial movement.
Setup and payoffAbandoned setups, forced payoffs, unprepared revelations, and missing repercussions of major events.
Repetition detectionReplayed events and repeated episode boundaries that add no substantive development and impede progression.
QualityCharacter qualityProtagonist distinctiveness, arc and depth; supporting-character and antagonist functions; ensemble balance; emotional bonds, relational tension, and relationship progression.
Dramatic engagementOpening setup and protagonist introduction; clarity of goals, opposition, and stakes; sustained anticipation, cross-episode follow-through, payoff, and emotional investment.
Plot-planning qualityEvent non-repetition, obstacle differentiation, scene progression, opening and later pacing, emotional trajectory, continuing anticipation, information distribution, padding, subplot integration, and main-plot payoff.
Episodic Script qualityCharacter-specific and situation-appropriate dialogue, emotional movement, information density, spoken naturalness, filmable description, scene staging, and concrete action.
Setting qualityGenre fit, contribution of setting to the plot, exposition load, and pacing of setting revelation.

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

Evidence and rating anchors. Each finding must identify a named check and cite supporting evidence with its episode and scene location. Contradictions and repetition require evidence from both relevant passages, while positive ratings must be supported by developments across episodes or stages, or by significant plot points. The severity of identified issues is then assessed using criterion-specific anchors, considering their scope, frequency, evidential clarity, and the extent of revision required.

Detailed Rating Anchors for Character State

We illustrate the fine-grained rubric using the Character State category. Its six issue criteria are life-state discontinuity, identity or allegiance discontinuity, contradictory character identity, ability-state discontinuity, abrupt change of personality or motivation, and abrupt relationship reversal. Each finding must cite specific textual evidence with episode and scene locations; contradictions require both relevant passages. S/A/B denote issue severity. Its core content is presented below.

Life-State Discontinuity

Assess whether later appearances and actions respect established life states and physical constraints.

S. A character explicitly and realistically confirmed dead later participates as a living person in a key event, without an explanation such as a staged death, resurrection, double, or flashback.

A. A severe constraint, such as grave injury, coma, or imprisonment, disappears without treatment, escape, or an account of elapsed time.

B. Recovery is unusually rapid or thinly explained, but does not invalidate the plot.

Details of Trope Homogeneity Evaluation

Identity or Allegiance Discontinuity

Assess whether identity revelations and changes in allegiance are supported and reflected in subsequent events.

S. An identity reveal or betrayal has occurred and been noticed, yet later events proceed as if it had not happened, invalidating a key plot development.

A. A change in identity or allegiance lacks sufficient events or emotional motivation, making cooperation or opposition conspicuously abrupt.

B. The change has a basis, but its emotional transition and supporting scenes are weak.

Contradictory Character Identity

Assess whether established identity facts and relationship labels remain consistent across the script.

S. A core identity, family order, title, office, blood tie, or marital relationship directly contradicts an earlier statement and affects a main-plot premise.

A. An important identity label, form of address, or relationship label is inconsistent, hindering understanding of relationships, organizational position, or authority.

Ability-State Discontinuity

Assess whether the use and development of abilities respect previously established restrictions and changes.

S. An ability explicitly removed, sealed, exhausted, or restricted is later used normally to reverse a crisis without the restriction being lifted.

A. An upgrade, awakening, or removal of a restriction lacks the necessary process, cost, or trigger.

B. The change has a basic explanation, but its development is thin or unusually rapid.

Abrupt Change of Personality or Motivation

Assess whether changes in established traits or goals are supported by events and psychological development.

S. A core personality trait or long-term goal reverses fundamentally without a major event, psychological setup, or emotional buildup.

A. An event motivates the change, but its significance clearly cannot support the magnitude of the reversal.

B. A local action slightly conflicts with established characterization without undermining overall consistency.

Abrupt Relationship Reversal

Assess whether major changes in interpersonal relationships follow from sufficient events and emotional development.

S. Hatred, trust, intimacy, cooperation, or betrayal reverses fundamentally without a key event, emotional buildup, or resolution of a misunderstanding.

A. An event motivates the reversal, but its impact is insufficient or the buildup is rushed.

B. The shift is inferable, but the emotional transition or its expression is thin.

Details of Trope Homogeneity Evaluation

The trope homogeneity evaluator measures repetition and diversity in causal narrative patterns through causal-chain extraction, normalization, and metric computation. Extracted Cause–Method–Outcome–Response (C–M–O–R) chains are mapped to unified category labels within each script, from which repetition rates and conditional entropy are computed.

Causal-Chain Extraction

The reference configuration uses seven-episode windows with a two-episode overlap to capture cross-episode causal chains spanning window boundaries. An LLM extracts meaningful micro-level causal chains across all subplots.

Each chain consists of four slots: Cause, the event or state directly prompting an action; Method, the action and its operative mechanism; Outcome, the change directly produced by that action; and Response, a character’s subsequent reaction.

The extraction instructions require concise, reusable phrases that preserve the mechanism and direction of change while abstracting away names, locations, and specific props. Each chain also includes traceable textual evidence from the original script and an explicit explanation of the corresponding causal relationships. To avoid duplicate counting across overlapping windows, we deduplicate chains referring to the same underlying narrative event based on their episode locations and textual evidence.

Normalization

We construct a global taxonomy for each C–M–O–R slot using phrases extracted across all windows. Guided by plot-grounded causal explanations, we group expressions sharing the same core mechanism or behavioral goal, direction of change, and relevant subject–object relationships. A substitution test further verifies whether replacing one expression with another preserves its causal meaning within the slot. Thus, surface differences do not prevent merging, while shared themes or forms of expression alone are insufficient.

For example, bribing a guard with money to gain passage and offering jewelry to an informant for information both rely on exchanging benefits for cooperation and should be grouped together. In contrast, threatening a guard to gain passage relies on coercion rather than exchange, and should remain distinct despite achieving the same outcome.

Each final category contains its name, definition, inclusion and exclusion boundaries, and its original member phrases. Then the resulting taxonomies relabel all extracted chains. Metrics are computed from these normalized category sequences.

Metric Computation

Metrics are computed programmatically from the relabeled chains. Here, an nn-gram denotes slots within one causal chain. Given N>0N > 0 chains, let U4U_{4}, UCMOU_{\mathrm{CMO}}, and UMORU_{\mathrm{MOR}} denote the numbers of distinct four-slot patterns, C–M–O triples, and M–O–R triples, respectively. Then

Rep⁡4=1−u4N,Rep⁡3=12[(1−uCMON)+(1−uMORN)].(7)\operatorname{Rep}_{4}=1-\frac{u_{4}}{N}, \qquad \operatorname{Rep}_{3}=\frac{1}{2}\left[\left(1-\frac{u_{CMO}}{N}\right)+\left(1-\frac{u_{MOR}}{N}\right)\right]. \tag*{(7)}

For the normalized conditional entropy defined in the main text, the reference support threshold is τ=3\tau=3. The reported entropy is the arithmetic mean of the normalized summaries for M∣CM\mid C, O∣MO\mid M, and R∣OR\mid O.

Additional Experimental Details and Results

Table 6 and Table 7 provide the complete cross-model and ablation results corresponding to Tables 2 and 3, respectively. For space considerations, the main-text tables report only the aggregate Logic and Quality scores alongside the homogeneity metrics. The full tables additionally report all seven Logic and five Quality categories, providing a finer-grained view of performance across evaluation criteria. Both experiments use 40-episode scripts under the settings described in the main text. Results are reported as means and sample standard deviations across three runs.

MethodCCCSKSRSRSTSPRDAvgCQDEPPESSQAvg4G↓3G↓Hˉ↑\bar{H}\uparrow
Ours80.4 ±0.679.2 ±0.381.8 ±0.578.0 ±1.279.0 ±0.683.7 ±0.787.3 ±0.981.3 ±0.788.5 ±0.388.1 ±0.682.4 ±0.885.0 ±0.487.7 ±0.186.3 ±0.43.08 ±0.419.22 ±0.630.929 ±0.0056
w/o Future-Aware Mem74.0 ±0.374.5 ±1.379.5 ±0.673.0 ±0.373.2 ±0.480.0 ±0.167.3 ±1.474.5 ±0.379.6 ±0.885.0 ±0.178.0 ±0.477.2 ±0.886.4 ±0.581.2 ±0.34.65 ±0.4112.2 ±0.860.914 ±0.0084
w/o Future-Aware Mem & Tree Search72.1 ±0.869.5 ±0.877.4 ±0.871.8 ±0.570.4 ±0.577.3 ±0.966.2 ±0.472.1 ±0.383.7 ±0.482.7 ±0.272.6 ±0.580.3 ±0.686.7 ±0.481.2 ±0.47.13 ±1.216.4 ±0.810.887 ±0.0091
w/o Multi-Agent Pipeline (Direct Generation)68.7 ±0.763.5 ±0.774.4 ±0.266.5 ±0.566.9 ±0.770.4 ±0.281.8 ±0.470.3 ±0.175.9 ±0.680.1 ±0.371.4 ±0.275.5 ±0.981.3 ±0.876.8 ±0.26.87 ±0.1416.8 ±0.680.887 ±0.012

Table 7. Ablation results on 40-episode script generation using Qwen3.6-27B. Starting from Ours, we progressively remove Future-Aware Memory and Evolutionary Tree Search for Outline Refinement, while retaining the original multi-Agent Pipeline. The Direct baseline generates scripts without the multi-Agent Pipeline.

References

  1. [1]Minwook Bae and Hyounghun Kim. 2024. Collective critics for creative story generation. Preprint, arXiv:2410.02428.arxiv.org/abs/2410.02428
  2. [2]Jing Chen, Xinyu Zhu, Cheng Yang, Chufan Shi, Yadong Xi, Yuxiang Zhang, Junjie Wang, Jiashu Pu, Rongsheng Zhang, Yujui Yang, and Tian Feng. 2024. Hollywood: Unleashing the creativity of large language models in screenwriting via role playing. Preprint, arXiv:2406.11683.arxiv.org/abs/2406.11683
  3. [3]DeepSeek-AI, Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochaо Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chengyu Hou, Chenhao Xu, Chenze Shao, Chong Ruan, Conner Sun, and 300 others. 2026. Deepseek-v4: Towards highly efficient million-token context intelligence. Preprint, arXiv:2606.19348.arxiv.org/abs/2606.19348
  4. [4]Anil R. Doshi and Oliver P. Hauser. 2024. Generative ai enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28).
  5. [5]GLM-5-Team, ; Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, Chenzheng Zhu, Congfeng Yin, Cunxiang Wang, Gengzheng Pan, Hao Zeng, Haoкe Zhang, Haoran Wang, and 168 others. 2026. Glm-5: from vibe coding to agentic engineering. Preprint, arXiv:2602.15763.
  6. [6]Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. Preprint, arXiv:1904.09751.arxiv.org/abs/1904.09751
  7. [7]Jiaming Li, Yukun Chen, Ziqiang Liu, Minghuan Tan, Lei Zhang, Yunshui Li, Run Luo, Longze Chen, Jing Luo, Ahmadreza Argha, Hamid Alinejad-Rokny, Wei Zhou, and Min Yang. 2025. Storyteller: An enhanced plot-planning framework for coherent and cohesive story generation. Preprint, arXiv:2506.02347.arxiv.org/abs/2506.02347
  8. [8]Junjie Li, Xinrui Guo, Yuhao Wu, Roy Ka-Wei Lee, Hongzhi Li, and Yutao Xie. 2026. Lost in stories: Consistency bugs in long story generation by llms. Preprint, arXiv:2603.05890.arxiv.org/abs/2603.05890
  9. [9]Shijian Ma, Yunqi Huang, and Yan Lin. 2025. Dramabench: A six-dimensional evaluation framework for drama script continuation. Preprint, arXiv:2512.19012.arxiv.org/abs/2512.19012
  10. [10]Qwen Team. 2026a. Qwen3.6-27B: Flagship-level coding in a 27b dense model.
  11. [11]Qwen Team. 2026b. Qwen3.8-Max: A new bar for coding and cowork.
  12. [12]Jack W. Rae, Anna Potapenko, Siddhant M. Jayakumar, and Timothy P. Lillicrap. 2019. Compressive transformers for long-range sequence modelling. Preprint, arXiv:1911.05507.arxiv.org/abs/1911.05507
  13. [13]Mohammad Saifullah, Thomas Kornmaier, Taaha Kazi, Vasu Sharma, Aditya Sanjiv Kanade, and Aanand Kumar Yadav. 2026. Narrative world model: Narratology-grounded writer memory for long-form fiction. Preprint, arXiv:2607.05577.arxiv.org/abs/2607.05577
  14. [14]Qiuyu Tian, Xiaowen Gu, Hang Su, Jianghan Chao, Haojie Yin, Fan Guo, Xin Zhang, Jinjing Shen, Ewing Luo, Youyong Kong, Yingce Xia, and Zequn Liu. 2026. Shared worlds, private minds: Structured memory for long-form writing as world creation. Preprint, arXiv:2609.32401.arxiv.org/abs/2609.32401
  15. [15]Zhennan Wan and Jianfei Chen. 2026. Scaling long-form story generation via narrative state tracking. Preprint, arXiv:2609.35759.arxiv.org/abs/2609.35759
  16. [16]Mengru Wang, Haozhe Luo, Zhengqian Xu, Zhixiang Cui, Haoming Xu, Qu Yang, Jizhan Fang, Junfeng Fang, and Ningyu Zhang. 2026. Memtrapbench: Benchmarking cognitive traps in llm memory use. Preprint, arXiv:2608.20202.arxiv.org/abs/2608.20202

Paper details

vlnc:W_m9axw2av
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents