A Self-Evolving LLM Agent for Strategic Simulation Games 


Vol. 15,  No. 3, pp. 247-256, Mar.  2026
https://doi.org/10.3745/TKIPS.2026.15.3.247


PDF
  Abstract

Large Language Models (LLMs) have recently evolved into intelligent agents capable of interacting with external environments, with particular attention to self-evolving paradigms that improve performance without external supervision. This study proposes a self-evolving LLM agent for StarCraft II, a real-time strategy (RTS) game environment. The proposed agent employs a hierarchical architecture combining Qwen 3 1.7B as a small-scale model and GPT-3.5 Turbo as a large-scale model, where the small model learns domain-specific strategies through continual learning and the large model provides strategic judgment. The core mechanism is a replay buffer-based continual learning system that dynamically reorganizes training data centered on winning experiences. The agent accumulates successful gameplay experiences and progressively updates its training data to enhance strategic performance without external supervision. Experimental results demonstrate that the proposed method improves the win rate from 33% to 56% at difficulty level 6 over five rounds of learning, achieving 7 times higher performance than the existing baseline (8%). Notably, the agent exhibited the ability to autonomously discover and reinforce counter-strategies against the opponent AI's strategic patterns. This work presents a self-evolving framework that operates solely on win-loss signals without explicit reward functions, demonstrating the potential for autonomous learning of LLM agents in complex decision-making environments.

  Statistics
  Cite this article

[IEEE Style]

S. Kim, D. Ahn, J. Choi, "A Self-Evolving LLM Agent for Strategic Simulation Games," The Transactions of the Korea Information Processing Society, vol. 15, no. 3, pp. 247-256, 2026. DOI: https://doi.org/10.3745/TKIPS.2026.15.3.247.

[ACM Style]

San Kim, Daechul Ahn, and Jonghyun Choi. 2026. A Self-Evolving LLM Agent for Strategic Simulation Games. The Transactions of the Korea Information Processing Society, 15, 3, (2026), 247-256. DOI: https://doi.org/10.3745/TKIPS.2026.15.3.247.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals