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

AI summary

Overview

Research area: Natural Language Processing — long-form creative generation with large language models, specifically multi-agent pipelines for serialized microdrama script writing.

Technical level: Advanced. The paper assumes familiarity with LLM agent architectures, retrieval-augmented generation, evolutionary/tree search, and narrative-theory constructs such as setup-and-payoff.

Scope in one sentence: The paper diagnoses narrative homogenization in long-form LLM script generation and proposes DraMaze, a multi-agent pipeline combining trope-guided evolutionary tree search with a future-goal-indexed memory structure, evaluated on a new bilingual benchmark (MiDramaBench) with new causal-pattern repetition metrics.

What This Paper Is About

Microdramas — fast-paced, heavily serialized short-form dramas — are growing quickly, and producing dozens of coherent episodes with LLMs is hard: as generation proceeds, models keep recycling the same plot mechanisms (for example, relationships repeatedly advancing through misunderstandings and clarifications), which makes twists predictable. The authors show that conventional retrieval memory does not fix this and can make it worse, because it retrieves what is semantically similar to the current scene rather than what a future plot will actually need. DraMaze is their attempt to plan diverse causal paths forward while retaining history according to the requirements of the story's ending.

Key Contributions

1. A diagnosis of declining narrative diversity. The authors measure the entropy of "narrative mechanisms" across candidate continuations and show it generally declines after episode 20, with dense semantic retrieval (Dense RAG) scoring below a no-retrieval baseline (BASE) throughout the second half of a 40-episode script.
2. A quantified semantic–narrative relevance gap. In a case study using Episodes 1–20 of a human-written script as history and synopses of Episodes 21–30 as future goals, both Dense RAG and DOME-style graph retrieval recovered only 7 of 27 narratively valuable historical subevents (25.93% recall), while their continuations reused tropes at 23.30% and 16.67% versus 7.14% for human-curated memories and 5.56% in the original script.
3. The DraMaze pipeline. A multi-agent system with (a) trope memory-guided evolutionary tree search plus mutation (mutation probability 0.25) for outline refinement, (b) a Backward Narrative Dependency Tree (BNDT) built backward from the ending to index memories by future narrative goals, and (c) continually updated world-state memory for characters, relationships, and assets.
4. MiDramaBench and causal-pattern metrics. A bilingual benchmark built from PG-19 with stratified genre coverage (899 candidate ideas, 30 evaluation inputs per length setting), plus 4-gram and 3-gram repetition rates over Cause–Method–Outcome–Response chains and a support-weighted normalized conditional entropy measure.

Main Findings

• Logic and quality gains at every length (Qwen3.6-27B backbone). DraMaze scores Logic/Quality of 82.6/86.1 at 20 episodes, 81.3/86.3 at 40 episodes, and 82.5/85.4 at 60 episodes, versus DOME at 77.8/77.9, 72.1/77.3, and 66.3/74.1, and StoryWriter at 76.0/82.8, 74.6/83.5, and 74.0/82.4.
• Advantage grows with length. The paper reports the margin over the strongest baseline widening from 4.2 to 8.5 points on Logic across the three length settings, and states that for 60-episode scripts of up to 60k words DraMaze reduces causal-pattern repetition by 32.1% and reaches a Logic score of 82.5, 8.5 points above the best baseline.
• Lower repetition and higher diversity at 60 episodes. 4-gram repetition is 6.07% versus 9.75% (StoryWriter) and 11.6% (DOME); 3-gram repetition drops to 14.6% versus 21.5% for the strongest baseline; conditional entropy is 0.912 versus 0.876 and 0.866.
• Generalizes across backbones (40 episodes). With DeepSeek-V4-Pro, DraMaze reaches Logic 86.1 / Quality 90.2 versus 75.1/84.7 (DOME) and 73.3/84.7 (StoryWriter), with 4-gram repetition 3.21% and entropy 0.915. With GLM5.3-Flash, it reaches 89.6/92.5 versus 83.3/89.0 and 84.8/88.6, with 4-gram repetition 2.30% and entropy 0.927. The paper summarizes this as Logic gains of 4.8–11.0 points and Quality gains of 3.5–5.5 points.
• Both components matter (ablations, 40 episodes, Qwen3.6-27B). Removing future-aware memory drops Logic/Quality from 81.3/86.3 to 74.5/81.2 and raises 3-gram repetition from 9.22% to 12.2%. Removing both memory and tree search gives 72.1/81.2 with 4-gram repetition rising from 4.65% to 7.13% and entropy falling to 0.887. Direct generation without the multi-agent pipeline falls to 70.3/76.8.
• Retrieval alone is not the fix. Mechanism entropy declines in later episodes even with DOME-style graph memory, which the authors interpret as evidence that semantic or entity-level relevance does not capture causal dependencies on future goals.

Methodology in Plain English

The pipeline turns an idea into a full script in stages.

Planning. From the premise, world setting, and characters, planning agents build a coarse four-stage outline (Setup, Development, Climax, Resolution). Each anchor in that outline is then expanded by a local search tree: every expanded node produces four candidate next events, which are scored for novelty against a "trope memory" — a store of abstract cause–method–outcome–response patterns with character names, locations, and props stripped out. The top two candidates continue; with probability 0.25 a "mutation" step deliberately rewrites the causal links of the best candidate while still reaching the target anchor, so the search can escape familiar patterns. Finished chains are ranked by 0.4 × consistency + 0.3 × novelty + 0.3 × feasibility of reaching the anchor, and the winning chain updates trope memory.

Memory built backward from the ending. After episode outlines exist, a memory-planning agent extracts the final outcome, makes it the root of the Backward Narrative Dependency Tree, and recursively decomposes it into intermediate goals and then into prerequisite setups. Each episode is mapped to the goals it sets up, advances, or completes.

Episode-by-episode writing. After each episode, a memory agent extracts critical events and details, long-term clues, and foreshadowing that still needs payoff, keeps only what has future narrative value, and attaches each record to the relevant goal nodes with a contribution note and an importance score. Before writing episode t, the system collects records linked directly to that episode's goals, then adds up to ten high-importance events and clues from the descendant subtrees of those goals; foreshadowing is retrieved only through direct goal links. A separate world-state memory, retrieved by textual relevance and entity matching and refreshed after each episode, tracks characters, relationships, and assets.

Evaluation. Scripts are scored by an LLM multi-agent judge with criteria co-designed with drama professionals. Three independent reviewers log issues and strengths with textual evidence in an Evidence Ledger; an Audit Agent validates and deduplicates; an Arbitration Agent settles disputes. Categories start at 90 points, with S/A/B issues deducting 10/6/4 and strengths adding 3/1/0, giving a Ledger Score that is averaged 50/50 with an independent Holistic Scorer's 0–100 rating. Logic covers causal chain, character state, knowledge state, resource state, rules and space-time, setup and payoff, and repetition detection; Quality covers character quality, dramatic engagement, plot-planning quality, episodic script quality, and setting quality.

Benchmark. Sources come from PG-19 in three length bands — [30k, 50k), [70k, 90k), and [110k, 130k) words — targeting 20-, 40-, and 60-episode dramas at roughly 700–1,000 English words per episode, deduplicated by content hashes and normalized titles. Qwen3.8-27B screens sources for dramatic potential, extracts plot, relationships, and settings, and adapts them into open-ended microdrama ideas. Genre labels are embedded with BGE-large-en-v1.5 and k-means clustered into ten groups; up to 30 ideas are kept per length–cluster group, giving 899 candidates. The evaluation subset samples three ideas from the top 20 per group (two English, one Chinese), yielding 30 inputs per length setting (20 English, 10 Chinese).

Why This Matters

Research impact. The paper reframes memory in long-form generation as a causal problem rather than a semantic similarity problem. The entropy analysis and the 25.93% recall case study are reusable diagnostics, and the causal-pattern metrics give the field a quantitative handle on homogenization — a known concern (the authors cite prior work on reduced diversity in AI-assisted writing) that is usually discussed only qualitatively.

Real-world applications.

• High-volume microdrama and short-series production for streaming and short-video platforms, where dozens of episodes must ship quickly without repeating beats.
• Adaptation of public-domain and licensed prose IP into episodic scripts, which is exactly the construction path used to build MiDramaBench from PG-19.
• Bilingual production and localization pipelines, since the benchmark and generation setup cover English and Chinese.
• Writers' rooms and interactive/ branching fiction tools that need long-horizon consistency tracking for characters, clues, and unresolved foreshadowing.

Industry relevance. The "future-aware" indexing idea — retain history because a future goal will need it — is directly transferable to any long-horizon agent that must remember prerequisites rather than merely similar past situations, and the evidence-ledger evaluator offers a template for auditable LLM-based quality control.

Future Directions

• Cost and latency are not reported. The pipeline runs evolutionary tree search plus multiple planning, writing, memory, and judging agents per episode; the paper reports no token, wall-clock, or dollar figures, so the efficiency/quality trade-off against DOME and StoryWriter remains open.
• Human evaluation of full scripts is not reported. Criteria were co-designed with drama professionals and evaluation uses three LLM reviewers plus audit and arbitration, but no human ratings of the generated 20–60 episode scripts appear in the results.
• Stopping homogenization at its source. Entropy still declines with length (the paper notes structural repetition increases with script length, with 4-gram repetition rising from 1.35% at 20 episodes to 6.07% at 60), so what drives residual concentration at very long horizons is unresolved.
• Extending beyond the tested scope. All results are on Qwen3.6-27B, GLM5.3-Flash, and DeepSeek-V4-Pro with MiDramaBench-derived premises and scripts up to 60 episodes / 60k words; whether the BNDT and trope memory transfer to other genres, other languages, or interactive branching narratives is untested.

Target Audience

Researchers and practitioners in long-form creative generation, LLM agent architecture, and memory/retrieval design will get the most from this paper. It is also relevant to development teams building automated or assistive script pipelines for short-form video, and to narrative AI researchers interested in measurable definitions of plot repetition and diversity. Readers need a working grasp of agent pipelines and evaluation design; the narrative-theory framing (tropes, setup and payoff, foreshadowing) is explained well enough to follow.

Authors’ 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.