Skip to content

Planning Beyond Text: Graph-based Reasoning for Complex Narrative Generation

Conference: ACL 2026 arXiv: 2604.21253 Code: N/A Area: LLM Efficiency Keywords: Narrative Generation, Graph-Based Reasoning, Event Graph, Character Graph, Multi-Agent Iterative Optimization

TL;DR

PLOTTER shifts narrative planning from text representation to graph structure (event graph + character graph), diagnosing and repairing narrative flaws through multi-agent Evaluate-Plan-Revise iterative cycles on graph topology, significantly outperforming existing methods on narrativity, characterization, and dramatic tension.

Method

Key Designs

  1. Dual-Graph Narrative Representation: Event graph \(G_e\) with narrative relation labels \(\rho(e) \in \{\text{Causal}, \text{Foreshadowing}, \text{Suspense}\}\); character graph \(G_c\) with multi-dimensional attributes and evolving relationships.

  2. Multi-Agent Panel + Constrained Graph Editor: Theme Critic → Character Critic → Plot Critic, with cross-agent verification. Symbolic constraints: (1) causal rationality \(\mathcal{K}_C\) — causal subgraph must maintain DAG; (2) narrative completeness \(\mathcal{K}_N\) — all nodes reachable from start to end.

  3. Graph-Guided Progressive Script Synthesis: Deterministic DFS serialization preserving causal topology, state-aware scene generation.

Key Experimental Results

Dimension vs LLM-Plan-Write vs Dramatron vs DOC
Narrative 72% 74% 92%
Characterization 100% 76% 92%

Three evaluation agents exhibit strong synergy — +29% storyline, +34% script when combined vs individual.

Highlights & Insights

  • Paradigm shift from text to graph for narrative planning — making causal reasoning, foreshadowing relationships, and character dynamics editable symbolic objects
  • DAG and connectivity constraints provide deterministic verification independent of LLM reliability

Rating

  • Novelty: ⭐⭐⭐⭐⭐
  • Experimental Thoroughness: ⭐⭐⭐⭐
  • Writing Quality: ⭐⭐⭐⭐⭐
  • Value: ⭐⭐⭐⭐