💬 NLP

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

The paper introduces PLOTTER, a framework that enhances complex narrative generation by performing an Evaluate-Plan-Revise cycle on structural event and character graphs rather than direct text, thereby significantly improving global coherence, logical consistency, and character development.

Original authors: Hanwen Gu, Chao Guo, Junle Wang, Wenda Xie, Yisheng Lv

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanwen Gu, Chao Guo, Junle Wang, Wenda Xie, Yisheng Lv

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to write a blockbuster movie script. You have a great idea for the beginning and a cool idea for the ending, but when you try to write the middle, the story falls apart. Characters act out of nowhere, plot holes appear, and the emotional journey feels flat.

This is exactly the problem current AI (Large Language Models) faces when trying to write long stories. They are great at writing one sentence after another, but they often lose the "big picture," leading to stories that feel like a train with no tracks—just a long, winding path that eventually crashes.

Enter PLOTTER, a new framework that changes how AI writes stories. Instead of writing the story word-by-word, PLOTTER first builds a blueprint using a map of connections.

Here is how it works, broken down into simple analogies:

1. The Old Way: Writing a Story Like a Stream of Consciousness

Imagine trying to build a skyscraper by just laying bricks one by one without a blueprint. You might get a few floors up, but eventually, you'll realize the elevator shaft is in the wrong place, or the foundation is too weak.

Current AI methods do this. They write a sentence, then the next, then the next. If they make a mistake in Chapter 1, they often don't realize it until Chapter 10, and by then, the whole story is ruined. They are "writing as they go," which leads to logical cracks and boring plots.

2. The PLOTTER Way: The Architect's Blueprint

PLOTTER changes the game. Before writing a single line of dialogue, it acts like a master architect. It draws two distinct maps (graphs) on a whiteboard:

The Event Map (The Plot): This is a flowchart of what happens. It connects Event A to Event B with arrows that say "Because of this, that happened." It ensures the story has a logical skeleton.
The Character Map (The People): This is a web showing how characters feel about each other. It tracks who is friends with whom, who is lying, and how their relationships change over time.

The Analogy: Think of PLOTTER not as a writer typing on a keyboard, but as a director and editor standing in a control room, looking at a giant screen full of sticky notes and string connecting them. They can see the whole movie at once.

3. The "Critique Crew" (The Multi-Agent System)

Once the blueprint is drawn, PLOTTER doesn't just start writing. It calls in a team of specialized experts (AI agents) to inspect the blueprint before the movie is filmed.

The Theme Detective: "Hey, this story is supposed to be about 'Redemption,' but the main character is just being mean for no reason. We need to fix the plot to show why they are changing."
The Character Psychologist: "Wait a minute. In Scene 5, this character was a coward, but in Scene 6, they are a hero. That's impossible! They need a moment of growth in between."
The Plot Engineer: "This scene jumps from a quiet dinner to a car chase with no transition. It's a logical gap. We need to add a bridge."

4. The "Constrained Editor" (The Fixer)

When the experts find a problem, they don't just say "fix it." They use a Constrained Graph Editor. This is like a surgeon with a very strict rulebook.

The Rulebook: The editor can only make changes that keep the story logical. For example, time must move forward (no time travel loops unless planned), and every event must connect to the beginning and the end.
The Fix: If the character is too flat, the editor inserts a "Bridge Scene" where the character has a realization. If the plot is too boring, the editor adds a "Twist Node" to shake things up.

They keep doing this—Evaluate, Plan, Revise—until the blueprint is perfect.

5. The Final Act: Writing the Script

Only after the blueprint is flawless does PLOTTER start writing the actual script (the dialogue and descriptions). Because the map is perfect, the AI knows exactly where the characters are emotionally and where the plot is going. It's like an actor who has rehearsed the whole play perfectly; they don't stumble over their lines because they know the story's soul.

Why This Matters

The paper shows that when you force the AI to plan on a map (graph) instead of just words (text), the stories become:

More Logical: No more "How did they get there?" moments.
More Emotional: Characters actually grow and change in believable ways.
More Engaging: The plot has twists and turns that feel earned, not random.

In short: PLOTTER teaches AI to stop being a frantic typist and start being a thoughtful architect. It proves that to tell a great story, you need to see the whole picture before you write a single word.

1. Problem Statement

While Large Language Models (LLMs) excel at linguistic fluency, they struggle with long-form complex narrative generation. Existing methods (e.g., iterative hierarchical planning or role-playing agents) typically operate directly on sequential text representations. This approach leads to several critical failures:

Global Incoherence: Inconsistencies in plot logic and character development accumulate over long contexts.
Structural Fractures: Scripts often suffer from insufficient character motivation, lack of necessary plot twists, and disjointed causal chains.
Inefficient Reasoning: Text-based planning makes it difficult for models to explicitly reason about the underlying cause-and-effect web or evolving character relationships, leading to "cascading errors" where early logical flaws propagate downstream.

The authors argue that planning narratives directly on text is inefficient because it lacks explicit modeling of plot dependencies and structural constraints.

2. Methodology: The PLOTTER Framework

PLOTTER (PLanning BeyOnd Text) introduces a framework that shifts narrative planning from sequential text to structural graph representations. The framework decomposes script generation into three distinct stages:

Stage 1: Graph-based Script Planning

Instead of generating a text outline, the system constructs two interacting graphs based on a given premise:

Event Graph ( $G_e$ ): Represents the causal skeleton of the plot. Nodes are events with attributes (ID, description, narrative stage, time). Directed edges encode specific narrative relations: Causal, Foreshadowing, and Suspense.
Character Graph ( $G_c$ ): Models the sociodynamics of the narrative. Nodes represent characters with multi-dimensional personas (traits, conflicts, goals, secrets). Edges represent evolving relationships (Conflict, Cooperative, Emotional, Hidden).

Stage 2: Iterative Graph Refinement (Evaluate-Plan-Revise)

This is the core innovation. The system employs a Multi-Agent Critique module and a Constrained Graph Editor to iteratively repair the graphs before text generation.

Multi-Agent Critic ( $C$ ): Three specialized agents analyze the graph:
1. Theme Critic: Checks for thematic drift or overly explicit exposition.
2. Character Critic: Diagnoses flat arcs, lack of motivation, or abrupt attitude shifts.
3. Plot Critic: Audits causal discontinuities, missing suspense, and lack of turning points.
Constrained Graph Editor ( $R$ ): Translates the critics' feedback into atomic edit operations (e.g., Add-Plot-Bridge, Insert-Twist, Revise-Event).
Symbolic Constraints: To ensure logical integrity, edits must satisfy two deterministic constraints verified without LLM involvement:
1. Causal Rationality ( $K_C$ ): The event graph must remain a Directed Acyclic Graph (DAG) to prevent time loops.
2. Narrative Completeness ( $K_N$ ): All nodes must be reachable from the "Beginning" and lead to the "Ending," preventing isolated plot threads.

Stage 3: Graph-Grounded Script Synthesis

Once the graphs are optimized ( $G^*$ ), they are converted into a final script ( $S$ ):

Graph Serialization: The Event Graph is traversed (depth-first) to create a logically ordered textual sequence ( $T_h$ ) that preserves topological constraints. Character profiles are expanded from the Character Graph.
State-Aware Generation: A generator produces the script scene-by-scene. Crucially, it conditions on:
- The serialized event structure (ensuring causal flow).
- Character profiles (ensuring consistent voice/motivation).
- Rolling Narrative State ( $M$ ): A dynamic memory of previous scenes to maintain referential continuity and prevent logical drift.

3. Key Contributions

Graph-Based Planning Paradigm: PLOTTER is the first framework to perform narrative planning on structural graphs rather than direct text, significantly enhancing long-context causal reasoning.
Iterative Structural Refinement: It introduces a multi-agent system that diagnoses and repairs structural inconsistencies (causal gaps, character flatness) at the symbolic level, a capability inaccessible to text-based planners.
Deterministic Constraint Verification: By enforcing DAG and reachability constraints during the editing phase, the system guarantees topological validity, preventing logical loops and orphaned plot points.
Synergy of Modules: The ablation study reveals a "1+1 > 2" effect, where the collaboration between Theme, Character, and Plot agents yields significantly higher quality than any single agent operating in isolation.

4. Experimental Results

The framework was evaluated across diverse genres (Sci-Fi, Drama, Crime, etc.) using 50 premises and three backbone LLMs (GPT-4.1, DeepSeek-R1, Qwen3).

Quantitative Performance: PLOTTER significantly outperformed strong baselines (LLM-Plan-Write, Dramatron, DOC) in pairwise comparisons.
- Win Rates: Achieved dominant win rates (often >80-90%) in Narrative, Thematic Expression, Characterization, and Dramatic Engagement.
- Objective Metrics: PLOTTER achieved the highest Distinct-2 (diversity) and MATTR (lexical richness) scores, and the lowest Self-BLEU (redundancy), proving improvements were due to content quality, not length inflation.
Qualitative Analysis: Case studies showed PLOTTER successfully repaired "logical violations" (e.g., abrupt character transitions) by inserting bridging nodes and causal chains that text-based methods missed.
Human Evaluation: Human screenwriters confirmed the LLM evaluation results, with high inter-rater agreement (Fleiss' $\kappa$ = 0.688) and strong alignment with machine judgments (Cohen's $\kappa$ = 0.834).
Cost Efficiency: While iterative refinement adds cost, the system remains controllable (approx. $1.68 per script with GPT-4.1), with a "budget mode" reducing costs to $0.36.

5. Significance

The paper establishes that explicit structural representation is superior to implicit sequential text for complex reasoning tasks in creative writing.

Theoretical Impact: It validates the hypothesis that separating the "structural planning" (graph) from "textual realization" allows LLMs to overcome the limitations of long-context reasoning and hallucination.
Practical Application: PLOTTER provides a robust architecture for generating high-quality, coherent scripts, novels, or game narratives where logical consistency and character depth are paramount.
Future Direction: It suggests a new paradigm for LLM agents where symbolic graph editing serves as the "brain" for planning, while the LLM acts as the "mouth" for fluent generation.