EvoSpark : Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution

Shiyu He, Minchi Kuang, Mengxian Wang, Bin Hu, Tingxiang Gu · 2026

Read EvoSpark : Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution by Shiyu He, Minchi Kuang, Mengxian Wang, Bin Hu, Tingxiang Gu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-12776.txt.

Introduction

Shiyu He and four colleagues sustain long stories in multi-agent story worlds. Generative emergence is stochastic, causing two failures. Social memory piles up clashing states without resolution. Place logic drifts from plot. EvoSpark counters with a layered memory that digests experiences into living thought. A staging mechanism aligns role, location, and plot. An engine turns random sparks into lasting characters. A tiny premise thus grows into an open-ended story world. Tests show EvoSpark beats baselines across diverse paradigms.

Catalog record

CREATORShiyu He, Minchi Kuang, Mengxian Wang, Bin Hu, Tingxiang Gu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54683
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2604.12776

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

EvoSpark : Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution Shiyu He, Minchi Kuang, Mengxian Wang, Bin Hu, Tingxiang Gu Abstract 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, a 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. 1 Introduction The integration of Large Language Models (LLMs) into Multi-Agent Systems (MAS) has fundamentally reshaped the landscape of generative storytelling, enabling agents to simulate complex social interactions with unprecedented fluency Park et al. 2023 ; Piao et al. 2025 . However, while current systems excel at generating short-term vignettes, achieving long-horizon story evolution—where a simulation evolves autonomously from a minimal seed into an unbounded, self-sustaining, and logically coherent narrative ecosystem—remains an elusive goal Xia et al. 2025 . As narratives expand in complexity, two critical systemic deficits emerge in current architectures, preventing the sustainability of long-term logic. First, systems suffer from social memory stacking, where the conventional append-only memory architecture leads to the accumulation of conflicting relational states (e.g., distinct memories of being both a friend and a foe), causing behavioral incoherence Platnick et al. 2025 ; Zhong et al. 2023 . Second, text-based agents face narrative-spatial dissonance. Lacking a mechanism to synchronize narrative progression with spatial states, agents often generate ungrounded interactions that violate the essential role-location-plot logic—such as characters appearing in disjointed locations during plot-critical transitions—thereby severing the logical link between the story and its setting Ran et al. 2025 ; Chen et al. 2025 . Beyond these functional deficits, the field is structurally constrained by a paradigm schism. Traditional interactive narratives rely on rigid script adherence, ensuring logic but sacrificing autonomy Sun et al. 2025 . Conversely, recent LLM-based simulations often prioritize open-ended emergence Park et al. 2023 ; Yang et al. 2025 , leading to uncontrollable chaotic emergence. Although recent works imply control mechanisms Han et al. 2024 ; Wang et al. 2024 , existing fragmented architectures fail to support the full spectrum of narrative control necessary for diverse simulation needs—ranging from strict hierarchical planning to open-ended evolution Sun et al. 2025 . To address these challenges, we propose EvoSpark , a unified framework that integrates narrative control, cognitive evolution, and spatial grounding to foster interactive agent societies driven by endogenous character emergence. Our core contributions are: • Unified Narrative Operation Engine (NOE): We operationalize sparking —stochastic LLM hallucinations—not as errors, but as drivers for creativity via the Emergent Character Grounding Protocol (ECGP) . Through Ontological Promotion , the system validates and transforms fleeting narrative-induced hallucinations into persistent, legitimate main characters, effectively turning stochastic noise into structural assets for infinite world expansion. • Role Socio-Evolutionary Base (RSB): To resolve social memory stacking , we introduce the RSB as a mutable, living cognition base. Unlike static or append-only RAG approaches, the RSB employs event-driven reflection to continuously metabolize experiences—updating personality, social graphs, and goals via in-place modifications. This ensures the agent’s internal state evolves in real-time, maintaining consistency with the shifting socio-dynamic landscape. • Generative Mise-en-Scène (GMS): We mitigate narrative-spatial dissonance through GMS mechanism, which acts as a virtual stage manager to enforce strict Role-Location-Plot (RLP) alignment. It dynamically synchronizes character presence and transitions with the evolving narrative flow, ensuring spatial contexts remain logically congruent with story progression. 2 Related Work 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. Endogenous Multi-Agent Systems Research on MAS has transitioned from static orchestration to dynamic, endogenous evolution. Early frameworks like MetaGPT and Camel Hong et al. 2024 ; Li et al. 2023 relied on fixed Standard Operating Procedures, limiting adaptability. To address this, recent architectures enable self-improvement: CoMAS and AFlow Xue et al. 2025 ; Zhang et al. 2025d optimize policies via interaction rewards, while the Darwin Gödel Machine Zhang et al. 2025c allows agents to recursively modify their own code. In the realm of simulation, AgentSociety, OASIS, and Generative Agents Piao et al. 2025 ; Yang et al. 2025 ; Park et al. 2023 scale interactions to observe emergent norms, while BookWorld Ran et al. 2025 constructs societies directly from fictional texts. However, current endogenous systems exhibit a teleological mismatch for open-ended storytelling. First, evolutionary frameworks optimize for metric convergence (e.g., pass rates) Xue et al. 2025 ; Zhang et al. 2025d , treating the stochasticity vital for narrative expansion as noise. Second, social simulators like S3 and WarAgent Gao et al. 2023 ; Hua et al. 2023 suffer from social memory stacking—accumulating interaction logs without a metabolic mechanism to transform transient experiences into persistent structural updates, inevitably degrading long-horizon coherence. Socio-Evolutionary Dynamics and Memory To support long-term narratives, research primarily targets static consistency to mitigate identity drift. Frameworks like ID-RAG Platnick et al. 2025 employ identity knowledge graphs to ground agent personas, while Open-Theatre Xu et al. 2025 and MemoryBank Zhong et al. 2023 introduce hierarchical stores to ensure retrieval accuracy. Similarly, S3 and AgentSociety Gao et al. 2023 ; Piao et al. 2025 utilize memory modules to maintain coherent behavioral patterns across social simulations. However, these architectures typically treat memory as an accumulative log. This rigidity leads to social memory stacking, where obsolete relational states persist and conflict with new developments. While Generative Agents Park et al. 2023 uses reflection to synthesize observations, it focuses on preserving character states. Recent works attempt to address this: G-Memory Zhang et al. 2025a evolves hierarchies based on interaction trajectories, MemEvolve Zhang et al. 2025b meta-optimizes memory architectures, and DOME Wang et al. 2025 tracks temporal state changes. Despite these advances, a gap remains in enabling a socio-evolutionary metabolism that fundamentally transforms agent personality and relations over time, rather than merely accumulating context. Generative Mise-en-Scène Current text-based agents often suffer from Narrative-Spatial Dissonance, where generated narratives detach from coherent environmental contexts. While Generative Agents Park et al. 2023 grounds behavior in sandboxes and AgentSociety Piao et al. 2025 models urban mobility via AOIs, these environments often function as passive containers. BookWorld Ran et al. 2025 advances this by introducing discrete geospatial tracking and travel constraints. Similarly, NarrativeGenie and HAMLET Kumaran et al. 2024 ; Chen et al. 2025 dynamically position props and adjudicate physical interactions to match

Read the complete text