Open Role-Playing with Delta-Engines

TL;DR

The paper introduces open role-playing games (ORPGs) powered by Delta-Engine, a neural-augmented in-game module consisting of a base engine and a neural proxy that incrementally expands a character’s code in response to natural language instructions. It validates the approach through a Free Pokémon playground, showing how NL prompts trigger new methods and moves via incremental code updates, with retrieval to maintain scalability. A human–AI co-design data-generation pipeline targets novelty and interestingness using prototypes and Tags of Interest, improving out-of-domain generalization over purely synthetic data. Experiments reveal that co-designed data plus retrieval yields higher execution and accuracy, especially on hard tasks, indicating better generalization and scalability for evolvable game worlds. The work suggests meaningful future directions for fully realized virtual worlds, improved data alignment, and safety considerations in open-ended game AI systems.

Abstract

Game roles can be reflections of personas from a parallel world. In this paper, we propose a new style of game-play to bridge self-expression and role-playing: \emph{open role-playing games (ORPGs)}, where players are allowed to craft and embody their unique characters in the game world. Our vision is that, in the real world, we are individually similar when we are born, but we grow into unique ones as a result of the strongly different choices we make afterward. Therefore, in an ORPG, we empower players with freedom to decide their own growing curves through natural language inputs, ultimately becoming unique characters. To technically do this, we propose a special engine called Delta-Engine. This engine is not a traditional game engine used for game development, but serves as an in-game module to provide new game-play experiences. A delta-engine consists of two components, a base engine and a neural proxy. The base engine programs the prototype of the character as well as the foundational settings of the game; the neural proxy is an LLM, which realizes the character growth by generating new code snippets on the base engine incrementally. In this paper, we self-develop a specific ORPG based on delta-engines. It is adapted from the popular animated series ``Pokémon''. We present our efforts in generating out-of-domain and interesting role data in the development process as well as accessing the performance of a delta-engine. While the empirical results in this work are specific, we aim for them to provide general insights for future games.

Figures (5)

• Figure 1: Free Pokémon system. Please see our supplementary materials for web demonstration.
• Figure 2: Input-output template for incremental prediction in Free Pokémon. The engine is implemented using Python. For brevity, we omit some elements: the engine state $y_{t-1}$, instruction $x_t$.
• Figure 3: Human and AI design (co-design). At the top left, we illustrate the process we leverage prototypes to enhance the LLM's design. We align the descriptions of the prototype and its associated design result using colors.
• Figure 4: Histograms of 100 sampling. We highlight the number of evolution steps in the training data as a baseline. On the right, we show a concrete case of the retrieval process.
• Figure 5: Comparison of the roles crafted by different methods and a concrete case on the right. We visualize them from the semantics and interestingness space.