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

DraMaze. A Multi-Agent Pipeline with Future-Aware Memory for Diverse Microdrama Script Generation. By Siyang He and seven others, from Fudan University, the Shanghai Innovation Institute, and one other institution. Published on the twenty-second of October. It links to a code repository that is no longer there.

Microdramas are growing fast, and with them the demand for high-quality scripts. Generating a long-form microdrama script with an LLM asks for two things at once: diverse plot progression, and narrative coherence over a long range. The authors observe that both come under pressure as generation goes on. Candidate continuations concentrate on fewer and fewer narrative mechanisms. Conventional retrieval can make matters worse, reinforcing the reuse of tropes and overlooking historical information that later plots will depend on.

DraMaze answers with a multi-agent pipeline that gives each of those problems its own structure. For diversity, it explores diverse causal paths using an evolutionary tree search with mutation, guided by a trope memory. For coherence, it builds a Backward Narrative Dependency Tree, which organizes, retains and retrieves historical memories according to future narrative goals.

That second structure is where the future-aware memory of the title lives. What the story keeps from its past is organized around where the story is going next. Alongside the pipeline, the authors introduce MiDramaBench, a bilingual benchmark that evaluates script logic and creative quality across different script lengths. It comes with causal-pattern metrics that measure narrative repetition and diversity directly.

Experiments show consistent improvements across three backbone LLMs. On sixty-episode scripts of up to sixty thousand words, DraMaze reduces causal-pattern repetition by thirty-two point one percent. It also reaches a Logic score of eighty-two point five, which is eight point five points above the best-performing baseline.

A thirty-two point one percent drop in repeated causal patterns answers the observation the work began with, that continuations drift toward fewer narrative mechanisms the longer a script runs. The Logic score answers the other half, coherence across sixty episodes. DraMaze treats drift and lost threads as two separate failures, and pairs each with its own memory: a trope memory steering the search, and a dependency tree organized by future goals.