Source-linked AI summary
GameGPT: Multi-agent Collaborative Framework for Game Development
Dake Chen, Haoyang Zhang, Hanbin Wang, Yunhao Huo, Yuzhao Li, Junjie Wang
TL;DR
LLM-based agents used for game development face concerns about hallucination and redundancy. GameGPT combines multi-agent collaboration, lexicons, and code decoupling, and its empirical results demonstrate effective decision-making and decision-rectifying throughout the game development process.
Problem
The paper identifies hallucination and redundancy as concerns for LLM-based agents in game development.
Source: §1 · pp. 2–3Method
GameGPT uses dual collaboration, in-house lexicons, and code decoupling to address hallucination and redundancy in game development.
Source: §1 · p. 3Results
GameGPT demonstrates effective decision-making and decision-rectifying throughout the game development process.
Source: §1 · p. 3Takeaways & Limitations
GameGPT is presented as a multi-agent framework tailored to facilitate automated game development.
Source: §1 · p. 3
Abstract
from arXiv · showhide
The large language model (LLM) based agents have demonstrated their capacity to automate and expedite software development processes. In this paper, we focus on game development and propose a multi-agent collaborative framework, dubbed GameGPT, to automate game development. While many studies have pinpointed hallucination as a primary roadblock for deploying LLMs in production, we identify another concern: redundancy. Our framework presents a series of methods to mitigate both concerns. These methods include dual collaboration and layered approaches with several in-house lexicons, to mitigate the hallucination and redundancy in the planning, task identification, and implementation phases. Furthermore, a decoupling approach is also introduced to achieve code generation with better precision.
1 Introduction
GameGPT targets automated game development by addressing LLM-agent hallucination and redundancy through a multi-agent framework and complementary mitigations.
1 Introduction
- Game-development agents risk hallucinations in planning and code generation, while unnecessary tasks or code snippets create an additional redundancy problem.The paper says the domain’s trend pressure calls for precision and concision, and using one LLM across the development cycle is impractical and costly.Source: §1 · pp. 2–3
- GameGPT combines LLMs with small expert models and execution–review roles, alongside in-house lexicon instruction tuning and code decoupling.The authors propose these approaches to mitigate hallucination and redundancy.Source: §1 · p. 3
- GameGPT introduces a multi-agent framework tailored to automate game development.The framework is designed for the game-development process.Source: §1 · p. 3
- The authors report that GameGPT enables effective decision-making and decision-rectifying throughout game development.Source: §1 · p. 3
2 GameGPT
GameGPT coordinates specialized agents and expert models across planning, task classification, code generation, execution, and result summarization. It uses layered reviews, templates, and code decoupling to reduce hallucination and redundancy.
2.1 Overview
- The framework uses five stages—planning, task classification, code generation, task execution, and result summarization—with expert models supporting task identification and critics addressing hallucination and redundancy.Its dual collaboration combines LLMs with task-specific expert models and collaboration among agents in different roles.Source: §2.1 · pp. 3–4
2.2 Multi-agent Framework
- GameGPT assigns distinct agents to roles across the development cycle because long, complex projects challenge a single LLM’s memory, context limits, scalability, and reliability.Agents maintain private memories while consulting a shared public discussion; reviewer agents support collaboration to address hallucination and redundancy.Source: §2.2 · p. 4
2.3 Game Development Planning
- Planning starts with an LLM-generated task list prone to hallucinated or redundant tasks, addressed through genre templates, plan review, instruction tuning, and user rectification.These four strategies are described as orthogonal and layerable; users can review and refine interim plans through the frontend.Source: §2.3 · pp. 4–5
2.4 Game Development Task Classification
- Task classification uses a game-development-trained BERT classifier and a separate LLM to populate task-specific argument templates, with reviewer checks and user queries for missing information.The reviewer also maps task dependencies into a directed acyclic graph and uses breadth-first search to determine execution order.Source: §2.4 · pp. 5–6
2.5 Code generation
- GameGPT decouples Lua scripts into manageable snippets, then uses snippet examples for in-context inference to reduce hallucination and redundancy.The framework generates multiple code candidates for testing and user feedback, fine-tunes on a labeled in-house code-snippet lexicon, and uses code review and user suggestions to refine code.Source: §2.5 · p. 6
2.6 Game Development Task Execution and Result Summary
- The testing engineer executes tasks in dependency order, tracks and consolidates logs into a user-facing summary, and reports tracebacks that may require code or execution adjustments.Each task’s code is sent to the game engine for execution.Source: §2.6 · pp. 6–7