cs.CLNov 1, 2026
SaveTowards Self-Evolving Narrative Systems: Real-Time Adaptation and Long-Term Optimization with LLMs
Organizations: Duke Kunshan University & Duke University · Duke University
Abstract
Traditional interactive narrative systems offer strong authorial control and structural coherence but often restrict player agency. In contrast, LLM-driven narratives enable open-ended generation and greater flexibility, yet frequently suffer from inconsistency and diminished narrative control. To navigate this trade-off, we propose a hybrid narrative framework that dynamically balances pre-authored content with real-time LLM generation via semantic similarity alignment. Additionally, sustaining narrative quality over extended gameplay remains a core challenge in LLM-based storytelling. To address this, we introduce a self-evolving, feedback-driven optimization framework that leverages in-context learning to iteratively refine prompts, eliminating the need for manual prompt engineering. Through a case study and participant-based evaluation, we show that our framework enhances narrative coherence while preserving player agency, effectively aligning generated content with both authorial intent and evolving player behavior.
Figures & tables
Figure 1: System Architecture Overview. Our framework addresses coherence, adaptability, and continual improvement via three interdependent phases: I . Pre-generated Mode transforms author-defined Game Scripts into modular Game Plots offline using an LLM Generator, ensuring narrative structure and quality; II . Gameplay Mode dynamically selects between pre-authored and generated content using an adaptive narrative strategy that responds to player choices; and III . Self-Evolving Optimization uses player feedback to iteratively refine prompts via an LLM Optimizer, enhancing future generation. These refined prompts guide the LLM Generator to produce contextually grounded narratives that reflect both authorial design and player behavior.
Figure 2: Feedback-driven prompt refinement. (1) Feedback Synthesis : Players provide ratings (1–5) and qualitative comments on NPC Consistency , Narrative Coherence , and Player Agency , stored alongside interaction history. (2) Diagnose Analysis : The LLM Optimizer retrieves this data to identify issues such as inconsistent NPC behavior, disconnected narrative flow, or illusory agency. (3) Prompt Refinement : Refined system, scripted, or generated prompts are produced to resolve diagnosed issues and improve future narrative generation.
Figure 3: Track A Survey Results. Compared to the flat system, our structured system improves narrative coherence (+0.5) and NPC consistency (+0.5), with only a modest reduction in player agency (–0.2).
| Metric | Flat System | Structured System |
|---|---|---|
| Immersion | 4.0 | 4.3 ( 0.3) |
| Handling unexpected inputs | 3.6 | 4.1 ( 0.5) |
| Willingness to replay | 3.4 | 3.8 ( 0.4) |
Table 1: Additional Track A metrics (5-point Likert).
Figure 4: Self-evolving optimization vs. static baseline (5-point Likert). Iterative refinement improves coherence and NPC consistency by +0.3 each and agency by +0.2. Gains accumulate across cycles. Takeaway : Lightweight player-feedback loops tighten character voice and narrative logic without retraining.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Narrative Architecture Overview. In the pre-generated mode , authors design a Game Script that specifies the environment, NPC traits and goals, key questions, and ending conditions. This script is used to construct a Game Plot —a modular, hierarchical narrative structure comprising two types of branches: Scripted Branches , realized by LLM Generator using Scripted Prompt , follow the author’s scripts. To support player-driven variability, the LLM Generator dynamically expands Pre-generated Branches in response to decision points using Generated Prompt . A global System Prompt governs overarching narrative rules and constraints.
Figure 6: Adaptive Narrative Strategy: Example scene with branching hierarchy. At each layer, if the player does not reach a key question (key decision), they remain in the current node ( 1. Freeform Exploration ). Reaching a key question moves the narrative to the next layer: either to a best-matching pre-generated node ( 2. Authored Progression ) or an LLM-generated node ( 3. Emergent Adaptation ). This balances coherence with controlled variation in narrative.
| ID | Scene Name | Scene Summary | Key Questions (Authored) |
|---|---|---|---|
| 1 | Cell of Whispers | Fane reveals a hidden vent and offers escape in exchange for assistance. | k1. Do player promise to recover Fane’s memories for passage? k2. Do player lower head and ask what ’fealty’ truly demands? k3. Do player press ear to brick and listen for hollow beyond? |
| 2 | The Mercenary Oath | Ifan proposes an ambush operation in exchange for a key. | k1. Do player agree to lure Magister Ralvo behind the catapult? k2. Do player step into the shade and ask Ifan for the plan’s details? k3. Do player accept the seventy-thirty split? |
| 3 | Royal Gate of Sludge | The Red Prince blocks passage and demands fealty. | k1. Do player accept the Red Prince as a travelling companion? k2. Do player trust his promise of lizard mercenaries? k3. Will player pledge to seek the Dreamer who foretells his throne? |
| 4 | Pact in Suspicious Lab | Magister Yarrow offers forged documents in exchange for silence. | k1. Will player keep Yarrow’s experiments secret for freedom? k2. Do player lean close to read page twelve? k3. Do player pocket a vial for leverage? |
| 5 | Last Boat on the Sand | Ifan prepares to secure the last seaworthy vessel. | k1. Do player trust his oil-barrel plan? k2. Do player call for Ifan’s covering fire? k3. Will player promise to share Bishop Alexander’s bounty? |
| 6 | Moonlit Monstrosity | Fane provides guidance against a Voidwoken threat. | k1. Do player trust Fane’s lesson about the weak spot? k2. Do player hold back Source bursts and drive your blade where instructed? |
Table 2: Overview of the "Escape from Fort Joy" Narrative Arc Scenes and Key Questions
| Theme | Definition (Desired Direction) | Representative Quote |
|---|---|---|
| NPC Consistency Dimension | ||
| Memory Failures | NPC contradicts prior facts or forgets commitments. (Lower is better) | “There were definitely moments of inconsistency. For example, after I killed an enemy, I entered ’he actually didn’t die, just injured,’ Ifan believed it.” (P11) |
| Consistent Persona | NPC maintains tone, goals, and speech patterns. (Higher is better) | “Fane stayed sly and dark-humored, even when I tried to call his bluff about the brick. It felt like I was actually negotiating with Fane.” (P5) |
| Narrative Coherence Dimension | ||
| Opaque Endings | Scene fails to close despite player indicating completion. (Lower is better) | “I have to give a very strictly-formatted input to pass the game.” (P12) |
| Earned Closure | Scene ends after deliberate choice + in-world acknowledgment. (Higher is better) | “When I promised Fane, he said ’Good. Now we move.’—it felt real.” (P10) |
Table 3: Complete thematic map. Themes are grouped by evaluation dimension, with color highlighting for each dimension: NPC Consistency , Narrative Coherence , and Player Agency . Representative quotes include both positive and negative experiences.
Figure 7: Game Interface. Left: Narrative Map; Right: Dialogue area. Players interact with NPCs through natural-language text input in the chat interface. The orange arrow indicates entered pre-generated node or generated node.
Figure 8: Backend Monitor Dashboard. The interface tracks ( Left ) authored vs. generated nodes, ( Top Right ) real-time semantic alignment and branching when player input diverges, and ( Bottom Right ) prompt optimization cycles with key metrics (e.g., coherence improvement).
Figure 9: Narrative Branching at Layer N. Emergent Adaptation (Phase II) forms a complete node parallel to Scripted/Pre-generated Branch (Phase I). The system chooses the path using semantic alignment and, after executing Emergent Adaptation, returns to the main backbone with updated state.
| Flat System (Baseline) | Structured System (Ours) | |
|---|---|---|
| Scenario | Player input during tense escape scene: “ I look closely at the puddle on the floor. ” | Same input in same context ( ) |
| System Response | “ The puddle reflects a beautiful rainbow and you see a magic sword inside. ” ✗ Unconstrained generation; deviation leads to hallucination or context drift. | “ The puddle is mixed with blood and sewer sludge. It reflects the flickering torchlight. Eyes up. The filth isn’t going anywhere. We move. ” ✓ Contextually Bounded: 1. Defers narrative progression. 2. Treats deviation as a design signal for exploration. |
Table 4: Freeform Exploration.
Figure 10: Illustration of narrative optimization. ( Left ) Static baseline; ( Center ) player feedback guiding the optimizer across rounds; ( Right ) refined narrative structure after optimization. The optimization improves narrative coherence, enhances NPC consistency, and increases player agency.
Figure 11: Optimization Addresses “Opaque Endings”. The top box shows the diagnosis-to-refinement workflow. The bottom box compares player-input handling before vs. after optimization.
More in this topic
Interactive Storytelling
Interactive storytelling studies how people and AI create responsive narratives across choices, dialogue, games, and embodied media.
See all 28 papers