Organizations: School of Computer Science and Technology, Xinjiang University, Urumqi 830046, China · Department of Precision Instrument, Tsinghua University, Beijing 100084, China
Realizing endogenous narrative evolution in LLM-based multi-agent systems is hindered by the inherent stochasticity of generative emergence. In particular, long-horizon simulations suffer from social memory stacking, where conflicting relational states accumulate without resolution, and narrative-spatial dissonance, where spatial logic detaches from the evolving plot. To bridge this gap, we propose EvoSpark, a framework specifically designed to sustain logically coherent long-horizon narratives within Endogenous Interactive Agent Societies. To ensure consistency, the Stratified Narrative Memory employs a Role Socio-Evolutionary Base as living cognition, dynamically metabolizing experiences to resolve historical conflicts. Complementarily, Generative Mise-en-Scène mechanism enforces Role-Location-Plot alignment, synchronizing character presence with the narrative flow. Underpinning these is the Unified Narrative Operation Engine, which integrates an Emergent Character Grounding Protocol to transform stochastic sparking into persistent characters. This engine establishes a substrate that expands a minimal premise into an open-ended, evolving story world. Experiments demonstrate that EvoSpark significantly outperforms baselines across diverse paradigms, enabling the sustained generation of expressive and coherent narrative experiences.
Figures & tables
Figure 1: The Architecture of EvoSpark . The framework initiates with Narrative Conception & Macro-planning, utilizing the Unified Narrative Operation Engine for modularized storyworld and character instantiation. Finally, the Simulation & Evolution module drives the narrative loop, managing continuous interactions via the Episodic Simulation Scheme and social memory updates based on the Stratified Narrative Memory.
Figure 2: Dynamic Spatial Alignment. The Director Agent orchestrates narrative interactions driven by spatial context, integrating Entity Resolution and precise grounding to ensure logical consistency.
Examples of Narrative-Spatial Dissonance
Turn
Interaction Detail
T1
Cersei Lannister : (Her voice is a cold, precise command as she addresses the guards without turning, standing rigidly by the hearth with her silhouette sharp against the firelight.) “Bring me the signed contract. Then take the Stark girl to her chambers…”
… (Turn 2: Cersei waits with her back to the room, demanding to see the seals and signatures) …
T3
Cersei Lannister : (A slow, dangerous pivot brings her eyes to bear on the hesitant guard , her hand remaining outstretched and her posture a blade of ice in the firelight.) “My father’s orders do not supersede mine in this room. The parchment. Now…”
T4
Cersei Lannister : (She snatches the heavy parchment from the man’s trembling fingers, though her eyes, chips of green ice, never leave Sansa’s placid mask. ) “Good. Now see to your duties, before I find a captain who knows how to obey.”
Table 1: A fragment exhibiting a physical contradiction in T4. The character is described as maintaining unbroken eye contact with Sansa Stark, which is logically inconsistent with her previous pivot to confront the guard in T3.
Examples with Dynamic Spatial Alignment
Turn
Interaction Detail
T1
Tyrion Lannister : <The Imp stands a half-step behind his wife, Sansa, his mismatched eyes scanning the preening lords and ladies.> (His hand drifts toward the small of her back, hesitating before falling away, a silent, aborted gesture of comfort. The Imp’s gaze lingers for a moment on Joffrey, enthroned and sneering, before he leans infinitesimally closer to Sansa’s ear, the rich velvet of his doublet brushing her arm.) “Try not to listen to the crows, little dove. They only caw when there’s carrion. Keep your eyes on me, if you can bear it. I am, at the very least, a smaller monster than most here.”
T2
Sansa Stark : <Standing rigidly beside Tyrion, her shoulders drawn tight.> (The Little Dove flinches almost imperceptibly at the warmth of his breath. Her gaze, which had been fixed on Joffrey with a practiced emptiness, falls to the polished floor, refusing to meet her husband’s eyes.) “As you say, my lord.”
T3
Tyrion Lannister : <His gaze shifts from Sansa’s averted face to the celebrating court.> (The Imp gives a faint, humorless smile, a barely perceptible twitch of his lips. His attention turns outward, scanning the room with a practiced, weary eye.) “A wise choice, my lady. The floor, at least, is honest. It makes no pretense of friendship while waiting to trip you. A virtue one can’t find in many a lord here.”
T4
Sansa Stark : <Standing beside him, a captive bird in a gilded cage.> (The Little Dove’s fingers tighten on the silken fabric of her gown, the knuckles of her gloved hands turning white for an instant. Her gaze remains lowered, tracing the veins in the polished marble as if they were a map leading far away from this place.) “It also keeps its counsel, my lord. A virtue rarer than honesty in this court, and far more valuable.”
Table 2: Examples of generated narratives with GMS. The <orange text> denotes spatial constraints, while text in parentheses indicates non-verbal actions. The Dynamic-Spatial-Alignment within GMS acts as an implicit spatial perceiver to help agents ground their physical reality.
Figure 3: The event-driven Reflect-Synthesize-Consolidation mechanism.
Figure 4: Comparison of win/tie rates between EvoSpark and baseline frameworks across different narrative modes (HDP, SNP, Free EN), languages, and LLM backbones. Detailed metric breakdowns are in Appendix B .
Figure 6: Long-Horizon Evolutionary Alignment Results: Win rates (bold) and tie rates of the full model vs. variants across 1, 5, and 10 events.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Paradigm
Metric
Kappa ( κ )
HDP
Role Performance (RP)
0.71
Logical Consistency (LC)
0.69
Narrative Resonance (NR)
0.65
Immersion (Im)
0.69
SNP
Role Performance (RP)
0.73
Logical Consistency (LC)
0.76
Appendix
Table 3: Agreement between Human and Model Judges across different control paradigms.
Figure 8: Cross-Domain Performance Comparison. Average win rates (%) of EvoSpark against baselines across three paradigms (HDP, SNP, and Free EN) in six narrative domains. OT: OpenTheatre, HW: HoLLMwood, BW: BookWorld.
Figure 9: Detailed Win Rates of EvoSpark vs. Baselines across all individual evaluation metrics. This breakdown covers Role Performance (RP), Immersion (Im), Narrative Resonance (NR), Long-Horizon Logical Consistency (LC), Narrative Soundness (NS), Creativity (Cr), and Plot Advancement and Conflict (PAC).
Figure 10: Detailed Average Scores (1–5) of EvoSpark vs. Baselines across all individual evaluation metrics. The results demonstrate consistent superiority in HDP, SNP and Free EN paradigms across diverse LLM backbones.
System
Total Duration
Avg / Turn
Median / Turn
Min / Max Turn
Avg / LLM Call
EvoSpark
63.6 min
38.17 s
42.56 s
3.99 s / 80.93 s
3.30 s
OpenTheatre
41.2 min
24.72 s
24.74 s
16.51 s / 33.49 s
7.89 s
BookWorld
25.9 min
15.53 s
12.66 s
2.60 s / 58.01 s
2.15 s
HOLLMwood
9.1 min
5.46 s
5.04 s
2.29 s / 18.82 s
4.77 s
Appendix
Table 4: Turn Latency Comparison based on a 100-Turn Benchmark. All statistics reflect resource consumption during the evolutionary simulation. Avg / LLM Call isolates pure inference efficiency.
Genre (Title)
Lang.
Type
Story Premise
Mystery
ZH
Existing
The Longest Day in Chang’an . On the Lantern Festival, undercurrents surge in Chang’an as Turkic “Wolf Guards” infiltrate, plotting a devastating fire attack named “Quelehuoduo.” Li Bi, head of the Jing’an Department, recruits death-row inmate Zhang Xiaojing to save the city. Within twelve hours, this duo navigates both street chases and court politics. They uncover that the terror plot intertwines with the Crown Prince’s struggle and official conspiracies aimed at the Tang Dynasty’s foundation. Ultimately, they risk their lives to prevent the Lantern Tower explosion, resolving the imperial crisis.
Classical Drama
EN
Existing
Romeo and Juliet . Amidst the feud between Montagues and Capulets in Verona, two star-crossed lovers marry in secret. A chain of tragic misunderstandings and fatal duels forces them toward a heartbreaking destiny that ultimately unites their warring families in grief.
Epic Fantasy
EN
Existing
A Song of Ice and Fire . Following the Red Wedding, the scattered remnants of House Stark must forge new identities through trauma. They aim to dismantle the alliance of terror, reclaim Winterfell, and restore honor in a brutal saga of vengeance and rebirth.
Eastern Fantasy
ZH
Synthesized
Ten Thousand Years of Qi Refining . Xu Yang, ostensibly a low-level cultivator, is actually a 100,000-year-old entity stronger than gods. When his sect faces destruction, he emerges to crush enemies with “accidental” displays of overwhelming power. Bound by a “Source Shackle,” he must eventually face upper-realm deities to protect his legacy.
Sci-Fi
EN
Synthesized
The War of Lost Will (2145) . Scientist Lin Shen discovers his “Memory Chip” tech is corrupted by a “Cognitive Pollution Program.” To prevent humanity from losing free will, he teams up with a rogue AI “Zero” and agent Su Li to crack the code within 72 hours, battling both external enemies and his own fading memory.
Modern Drama
ZH
Synthesized
I am the Sky of Dragon City . Huo Tian, a humiliated delivery man, awakens a “Future Vision” (3-second foresight). Discovering that a wealthy heir plotted his mother’s illness to steal a family recipe, he uses his ability to win high-stakes bets and dismantle the heir’s family empire in a story of tactical revenge.
Appendix
Table 5: Detailed specifications of the evaluation scenarios. The dataset covers six distinct genres, comprising three existing literary works and three synthesized open-ended narratives. Lang. denotes the simulation language (ZH: Chinese, EN: English).
Model
Baseline
Win Rate (%)
EvoSpark Avg. Likert
Winner
Existing
Synthesized
Existing
Synthesized
Gemini-2.5-Pro
vs. OpenTheatre
80.00
55.00
3.79
4.35
Existing ✓
vs. HoLLMwood
93.30
85.00
4.40
4.50
Existing ✓
vs. BookWorld (SNP)
80.00
95.80
3.96
4.81
Synthesized ✓
vs. BookWorld (FEN)
56.00
68.00
4.37
4.37
Synthesized ✓
DeepSeek-V3.2-Think
vs. OpenTheatre
86.70
68.30
4.11
4.17
Existing ✓
Appendix
Table 6: Comprehensive breakdown of Win Rates and EvoSpark ’s absolute Average Likert Scores across Existing (prior knowledge) and Synthesized (no prior knowledge) scenarios. The results confirm that the framework’s performance does not fundamentally rely on LLM pre-training memorization.
UPDATE_RELATION_PROMPT
You need to update your relationships with relevant characters based on the following information. ## Character Description {role_profile} ## Character Relationships {relation} ## Character Current Status {status} ## Conversation History {history_text} ## Character Relationship Update Requirements Please strictly follow the requirements below and return the updated relationships in JSON format: 1. Decision Logic: Combine the “Character Current Status” and “Conversation History” to determine whether the relationships need to be updated. Only update if there are significant interactions or changes in the dynamic. 2. Update Strategy: If changes are needed, please modify or supplement the original “detail” field content to reflect the new state. If the existing description is still accurate and sufficient, do not change it. 3. You can only modify the values of the “relation” and “detail” fields in each sub-object. 4. The value of the “relation” field must be a list of strings (List[str]), for example: [“new relationship1”, “new relationship2”]. 5. The value of the “detail” field must be a string. Keep it concise and summarized (recommended 300-500 words maximum). Focus on core relationship points and recent changes; avoid lengthy historical reviews. 6. Do not change any other keys (e.g., “ZhaoKai-en”, “LinWanYue-en”, etc.) or the overall JSON structure. 7. Your response cannot contain any extra text or explanations besides the updated JSON. 8. You cannot delete characters, even if there is no relationship. Important: Ensure the total JSON length does not exceed the model’s output limit. Prioritize JSON completeness.
Appendix
Table 7: Prompt template for updating character relationship networks based on recent interactions.
UPDATE_PROFILE_PROMPT
You need to update the character’s “profile” field based on the following information. ## Original Character Description (JSON format) {role_profile} ## Character Current Status {status} ## Conversation History {history_text} ## Character Description Update Requirements Please strictly follow the requirements below and only return the updated “profile” field’s string content : 1. Analyze the “profile” field in the “Original Character Description”. 2. Combine the “Character Current Status” and “Conversation History” to determine whether the “profile” field needs to be updated. 3. The “profile” field can only be changed when major changes related to the character occur in the story and have an impact on them. 4. If changes are needed, please modify or add to the original “profile” field content. 5. If no changes are needed, please return the original “profile” field’s string content . 6. Your response must be pure text string, and can only contain the content of the “profile” field after updating (or without updating). 7. Do not include any JSON structure 8. Do not include any extra text or explanations (such as “Okay, here’s the updated…”). For example, if the original “profile” is “a student”, after updating it should become “a student who just finished an exam”, you can only return “a student who just finished an exam”.
Appendix
Table 8: Prompt template for updating character profiles (long-term memory) based on narrative progression.
GENERATE_SPATIAL_POSITIONING_PROMPT
# Role Definition You are a professional stage director specializing in spatial blocking, skilled at arranging character positions to create dramatic tension and visual composition. # Core Task Design reasonable spatial positioning for all participating characters (including NPCs) based on the current scene/event, character relationships, and dialogue history. # Input Information ## Current Scene/Event : {scene_or_event} ## Participating Characters List : {roles_list} ## Current Location : Location Name: {location_name}; Location Description: {location_description} ## Recent Dialogue History : {recent_history} ## Current Dialogue Round : Round {current_round} # Spatial Positioning Design Principles ## 1. Relative Position : Describe distances (face-to-face, side-by-side…), power dynamics, emotional relationships. ## 2. Embodied Posture : Standing, Sitting, Other postures. ## 3. Facing Direction : Face-to-face, Back turned, Sideways, Same direction. ## 4. Scene Interaction : Furniture interaction, Environmental interaction, Prop interaction. # Output Format Requirements Must output a strict JSON object in the following format: {{ ‘‘spatial_layout’’: ‘‘One-sentence description of overall spatial composition (20-40 characters)’’, ‘‘positions’’: {{ ‘‘Character A Name’’: {{ ‘‘position’’: ‘‘Position in space (e.g., by window)’’, ‘‘posture’’: ‘‘Body posture (e.g., standing)’’, ‘‘facing’’: ‘‘Facing direction (e.g., facing Character B)’’, ‘‘relation_to_scene’’: ‘‘Relationship to scene elements’’ }}, ... }} }} # Design Considerations 1. Dynamic Adjustment : Fine-tune based on dialogue development. 2. Relationship Hints : Use distance and facing to suggest relationships. 3. Dramatic Tension : Increase/reduce distance based on conflict/reconciliation. 4. Logical Consistency : Matching location characteristics. 5. Include All Characters : Ensure every participating character has clear position description. — Now, design the spatial positioning for this round of dialogue based on the above information. Output the JSON object directly without any other explanations or markdown code block markers.
Appendix
Table 9: Prompt template for the Generative Mise-en-Scène module, instructing the LLM to act as a stage director.
FIND_NEW_ROLE_INFO_PROMPT
You are a skilled screenwriter. Based on the following information, generate character information for {role_code}. ### Records of Previous Scenes {history_scene_json} ### Current Event {event} ### Information of All Other Characters {role_agents} ### Requirements 1. Based on the records of previous scenes, generate character information. 2. The character information should include character profile, gender, identity, and relation. 3. Return in JSON format, formatted as follows: {{ ‘‘profile’’: ‘‘character profile’’, ‘‘gender’’: ‘‘character gender’’, ‘‘identity’’: ‘‘character identity’’, ‘‘relation’’: ‘‘character relationships’’, ‘‘name’’: ‘‘character name’’, ‘‘nickname’’: ‘‘character nickname’’ }} 4. Forbidden to output any explanations, comments, or Markdown markers (e.g., “‘json, “‘python).
Appendix
Table 10: Prompt template for the Emergent Character Grounding Protocol (ECGP), used to instantiate new characters from narrative context.
Large language model-based role-play simulation creates interactive agents that portray people, characters, or users for training, testing, and research.
Although large language models (LLMs) have demonstrated impressive creative fiction generation, they struggle to maintain narrative consistency and coherent plot lines in long-form stories. In this work, we introduce a unified framework for long-form narrative generation and verification. MAGNET, a multi-agent goal-driven narrative engine for storytelling, generates stories with persona-grounded character agents that propose actions based on a shared world state and evolving story goals, while ATLAS is a graph-based pipeline that compares scene-level world representations across a generated story to detect hallucinations. By evaluating MAGNET using an LLM editor, pairwise rubric scoring, and ATLAS, we show that our framework produces coherent narratives compared to single-model prompting and IBSEN. At 100 pages, MAGNET reduced annotations and hallucinations by 41 and 50%, respectively, compared to the single model baseline and by 34 and 45%, respectively, compared to IBSEN, with pairwise rubric evaluation showing similar results. These results suggest that long-form narratives can emerge from explicit world-state tracking and goal-driven multi-agent generation, providing a foundation for controllable and structurally coherent long-form narrative generation.
Aayush Aluru, Chloe Ho, Muhammad Hammouri +5
Pocket FM · Princeton University · University of Michigan +2
The rapid advancement of Large Language Models (LLMs) is revolutionizing AI for Games by enabling open-ended and fluid interactive storytelling. However, existing research has largely overlooked the critical challenge of maintaining long-horizon logical consistency and narrative integrity against unconstrained user interventions. To address this, we formulate this challenge as Narrative Commitment Preservation (NCP), and take interactive narrative as our testbed. We introduce NCP-Bench, a benchmark of 100 narrative environments derived from movie synopses. Each environment includes a structured narrative specification (trajectory, commitments, and initial facts) that we can automatically check throughout the interaction between the player agent and the narrator agent. Experiments across state-of-the-art LLMs reveal a substantial long-horizon consistency gap: high linguistic quality does not guarantee commitment preservation; even strong models frequently generate logically conflicting content under adversarial interventions, with the best-performing model (GPT-5.2) achieving only 42% survival rate after 20 turns and fact conflict rates ranging from 40% to 68% across models, and only isolated runs satisfying all achievement commitments within the 100-turn limit.
Yingpeng Ma, Jianhao Yan, Bei Shi +6
NLP2CT Lab, University of Macau, Macau, China · Westlake University, Hangzhou, China · Harbin Institute of Technology, Shenzhen, China +1
This paper introduces EvolvingWorld, a framework and benchmark for character and world co-evolution in interactive literary worlds. Existing systems either treat interactive literary simulation as static persona imitation or isolated scene generation, failing to capture how characters and worlds evolve together over time. To address this, EvolvingWorld models literary simulation as a long-horizon process where characters interact, scenes progress, and character and world states are persistently updated. Unlike prior systems relying on fixed schemas, EvolvingWorld adopts an open-schema framework to support simulation across diverse literary worlds. The framework consists of two coupled modules: a Character Agent for multi-character role-play and persistent profile evolution, and an LLM-based World Model for global and location/entity-level state maintenance and scene progression. Based on this architecture, we formulate 7 trainable tasks for scene initialization, interaction generation, and state update. We construct a dataset from 57 books, producing 138,596 supervised training samples and 222 snapshots for testing. Furthermore, we introduce a trajectory-level LLM-as-Judge evaluation protocol spanning 10 dimensions and 20 metrics. Experiments show that EvolvingWorld can improve long-horizon simulation by effectively maintaining persistent, coherent character and world development.
Qing Zong, Yue Guo, Mengxin Yang +2
Hong Kong University of Science and Technology · LIGHTSPEED · Huazhong University of Science and Technology +1