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

TL;DR

The paper tackles persona drift and narrative incoherence in LLM-driven characters by introducing Living Novel, a mobile-friendly system that transforms novels into timeline-aware, multi-character conversations. It combines a data-free Deep Persona Alignment (DPA) with Coherence and Robustness Enhancing (CRE) to enforce diegetic time constraints and improve consistency, using a Diegetic Knowledge Graph and story-time gated retrieval. Through a three-phase evaluation (automatic, lab, in-the-wild) and Jules Verne as a test case, the approach achieves superior persona fidelity compared with GPT-4o and near-perfect coherence and robustness, while offering practical design guidelines for narrative systems. The work demonstrates strong potential for scalable, immersive, mobile-first interactive narratives and outlines future extensions to embodied, multimodal storytelling.

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.

Figures (9)

• Figure 1: An overview of the Living Novel system. Our end-to-end pipeline automatically transforms any novel (left) into a mobile-accessible conversational experience (right). The system uses Deep Persona Alignment (DPA), and Coherence and Robustness Enhancing (CRE) (center) to generate in-character, spoiler-free agents that users can interact with at any point in the story.
• Figure 2: The three-stage pipeline for creating Character Aligned LLMs. First, Pre-processing extracts character profiles and a diegetic timeline graph from the novel. Stage 1 (Deep Persona Alignment) uses these profiles to generate in-character training data, fine-tuning a base model for persona fidelity. Stage 2 (Coherence and Robustness Enhancing) uses the timeline graph and profiles to generate "Immersion Training Data," further training the model to be timeline-aware and robust to out-of-domain questions.
• Figure 3: Interface of the system. (a) In the main interface, users can upload a novel, and the system automatically extracts basic information (e.g., character profiles, setting, and worldview). (b) Single chat interface. Users select a primary character and engage in multi-turn dialogue with that character. (c) Group chat interface. Users can choose multiple primary characters and converse with them simultaneously.
• Figure 4: Distribution of the average CharacterBox scores for each character. The black bar indicates median and red bar shows the mean value.
• Figure 5: Automatic evaluation of Timeline-coherence (TT) and Robustness (RT). Scores represent the percentage of correct responses (out of 100). Results demonstrate our story-time–aware memory (present in Full and only CRE) is highly effective.