Living the Novel: A System for Generating Self-Training Timeline-Aware Conversational Agents from Novels (2512.07474v1)

Abstract: We present the Living Novel, an end-to-end system that transforms any literary work into an immersive, multi-character conversational experience. This system is designed to solve two fundamental challenges for LLM-driven characters. Firstly, generic LLMs suffer from persona drift, often failing to stay in character. Secondly, agents often exhibit abilities that extend beyond the constraints of the story's world and logic, leading to both narrative incoherence (spoiler leakage) and robustness failures (frame-breaking). To address these challenges, we introduce a novel two-stage training pipeline. Our Deep Persona Alignment (DPA) stage uses data-free reinforcement finetuning to instill deep character fidelity. Our Coherence and Robustness Enhancing (CRE) stage then employs a story-time-aware knowledge graph and a second retrieval-grounded training pass to architecturally enforce these narrative constraints. We validate our system through a multi-phase evaluation using Jules Verne's Twenty Thousand Leagues Under the Sea. A lab study with a detailed ablation of system components is followed by a 5-day in-the-wild diary study. Our DPA pipeline helps our specialized model outperform GPT-4o on persona-specific metrics, and our CRE stage achieves near-perfect performance in coherence and robustness measures. Our study surfaces practical design guidelines for AI-driven narrative systems: we find that character-first self-training is foundational for believability, while explicit story-time constraints are crucial for sustaining coherent, interruption-resilient mobile-web experiences.

• The paper introduces a two-stage pipeline combining Deep Persona Alignment and Coherence and Robustness Enhancing for self-training conversational agents.
• It achieves high persona fidelity and near-perfect timeline coherence while mitigating spoiler leakage and frame-breaking responses.
• User studies and in-the-wild deployments confirm the system's low-latency mobile performance and engaging, narrative-driven interactions.

Living the Novel: End-to-End Timeline-Aware Conversational Agents from Literary Texts

System Motivation and Technical Contributions

The Living Novel system formalizes the transformation of static literary works into interactive, multi-character conversational environments, uniquely integrating persona fidelity and diegetic narrative awareness. LLM-driven characters have failed to maintain consistent roles due to pervasive persona drift and narrative incoherence—issues including spoiler leakage and frame-breaking responses. Living Novel introduces a two-stage pipeline: Deep Persona Alignment (DPA), utilizing data-free reinforcement fine-tuning to solidify character identity, and Coherence and Robustness Enhancing (CRE), enforcing narrative constraints via a temporal knowledge graph and gated retrieval. The system targets mobile accessibility through decoupled client-server deployment, offloading computational complexity.

Figure 1: An overview of the Living Novel system architecture automating the transformation of any novel into a mobile-accessible, multi-agent conversational application.

Multi-Stage Architecture and Training

The system pipeline comprises four critical stages: pre-processing, DPA, CRE, and deployment. Initially, a foundation LLM extracts structured character profiles and constructs a diegetic knowledge graph anchoring entities, relations, and events in narrative time. DPA then generates synthetic, contrastive preference datasets via timeline-cycled, emotion-varied, and intent-diverse prompt generation. Preference pairs are synthesized: ideal, in-character responses versus persona-drifting or frame-breaking alternatives, enabling reinforcement fine-tuning (Group Relative Policy Optimization, GRPO) of Qwen-2.5-7B adapters without human annotation.

In CRE, narrative-specific memory is engineered by organizing information in a story-time-gated retrieval database, architecturally preventing spoiler leakage and robustly refusing out-of-domain queries. A retrieval-grounded alignment pass leverages the diegetic graph, generating distinct training data targeting factual QA, temporal adversarial, and frame-breaking test cases, enhancing both context-awareness and robustness.

Figure 2: Pipeline overview for character-aligned LLM generation, detailing persona alignment and timeline-aware immersion training.

Interface and User Experience

Living Novel operationalizes this backend through a web interface supporting single-character and group dialogues, enabling users to anchor conversations at arbitrary story points, with the system ensuring only context-appropriate, in-character responses attuned to the selected timeline coordinate. The decoupled architecture guarantees low latency (<2s/turn) across mobile devices.

Figure 3: UI components for uploading novels, single/group character chat, and timeline-based session navigation.

A dual-level, story-time-gated retrieval mechanism parses queries into low- and high-level keyword searches over entities and relations, with all retrieval strictly limited by the narrative present.

Empirical Evaluation

Automatic Metrics

The paper deploys Jules Verne’s "Twenty Thousand Leagues Under the Sea" as an evaluation substrate, benchmarking four system variants (Baseline, DPA-only, CRE-only, Full) and GPT-4o across persona fidelity (CharacterBox metrics), coherence, and robustness. The Full system achieves top scores in all evaluation axes, notably outperforming GPT-4o in Behavioral Accuracy and Personality Traits.

Figure 4: CharacterBox score distribution per character, indicating median and mean for persona fidelity across systems.

CRE and Full system architectures achieve near-perfect timeline coherence (TT=90.8/100), decisively preventing spoiler leakage, while the Full system demonstrates strong defense against frame-breaking out-of-domain queries (RT=73.5/100).

Figure 5: Automatic evaluation for timeline-coherence and robustness, demonstrating efficacy of story-time-aware memory constraints.

Controlled Lab Study

A within-subject design finds highly significant user preference for the Full system in direct head-to-head comparisons; median ranking of 1.0 for Full, with 77.4% of turns ranked #1 (Wilcoxon $p<0.001$ for all pairwise comparisons). Participants cite authenticity in persona and crucial narrative gating as key to immersion.

Figure 6: User interface used for LLM ranking during the lab paper.

Figure 7: Per-turn distribution of user rankings, highlighting dominant preference for the full Living Novel system.

In-the-Wild Deployment

A longitudinal diary paper on mobile devices reveals "narrative snacking"—high-frequency, short-duration sessions leveraging the timeline as practical anchor for re-engagement. The system sustains high perceived coherence (>4/5 Likert) and user satisfaction, with participants attributing ease of interruption recovery to explicit timeline context management.

Figure 8: Screen capture of in-the-wild user sessions, exemplifying the mobile usability of the system.

Figure 9: Survey schematic with session lengths, use contexts, and overall user rating distributions for real-world deployments.

Implications for Narrative AI and Future Directions

The Living Novel architecture demonstrates that deep, personalized model alignment and diegetic time constraints are both necessary and computationally feasible for believable multi-character embodiments. Design recommendations include: 1) leveraging scalable, data-free RL pipelines over manual annotation, 2) hard architectural narrative gating—not semantic similarity alone—to prevent spoilers and maintain trust, 3) explicit timeline tools for mobile, interruption-resilient engagement.

Potential expansions include multimodal embodiment, auditory interaction via high-fidelity TTS, integration of personalized episodic memory, and AR interfaces for physical-digital character presence. These extend Living Novel beyond entertainment, into education and document-grounded conversational systems.

Conclusion

Living Novel substantiates a mobile-first, timeline-aware framework for generating coherent, robust, persona-faithful conversational agents from arbitrary literary texts. Results validate both the DPA and CRE pipeline stages as essential for surmounting the dual challenges of persona drift and narrative incoherence, outperforming prompt-only and general-purpose SOTA LLMs in both automatic and user studies. The system establishes a durable architectural foundation for future interactive narrative AI and mobile engagement modalities.

(2512.07474)