Source-linked AI summary
Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness
Jiajun Chen, Haoyu Wu, Mingda Jia, Xihui Liu
TL;DR
Runnable games do not necessarily deliver an enjoyable player experience, and translating gameplay into experience-oriented revision remains challenging. Recursive Game Creator links design, implementation, executable play, and preference-informed review in a recursive harness. It reports state-of-the-art overall performance across GameCraft-Bench and GameASG-Bench, with GameCraft-Bench improving across three refinement rounds.
Problem
Coding agents can produce runnable games, but successful execution does not establish play quality, and translating gameplay into experience-oriented revision remains challenging.
Source: §1 · p. 2Method
Recursive Game Creator iteratively refines games through Designer, Builder, Player, and Reviewer roles, using reusable programmatic play policies and behavioral, visual, and preference feedback.
Source: §1 · p. 2, §3.3 · p. 6Results
The harness achieves the highest overall score on GameCraft-Bench and GameASG-Bench among baselines; three rounds raise GameCraft-Bench’s overall score from 72.70 to 77.89 (+5.19).
Source: §1 · pp. 2–3Takeaways & Limitations
The policy-to-trajectory-to-feedback loop enables continued game evolution with fixed model parameters.
Source: §5 · p. 11Proxy strategies may not represent the full human player population, and proxy playtime alone does not show human interest or enjoyment.
Source: §3.2 · p. 5, §3.3 · p. 6
Abstract
from arXiv · showhide
Recent game design agents have made substantial progress in generating playable games. However, program correctness does not ensure an enjoyable experience for players. We present Recursive Game Creator, an experience-oriented harness to advance agentic game development from rough game prototypes into entertaining games. Recursive Game Creator organizes recursive development around four components: Designer, Builder, Player, and Reviewer. The Designer translates user instructions and Reviewer's feedback into detailed plans. The Builder turns these plans into candidate games. The coding-native Player creates and executes reusable policies through programmatic interfaces to efficiently collect diverse gameplay trajectories, mitigating evaluation bias caused by slow GUI-based collection. The Reviewer uses carefully designed trajectory-based metrics to induce player preferences, integrating with visual evidence and explicit textual preferences to evaluate games against game-specific criteria. Finally, the Reviewer accepts the better version and provides improvement reviews for the next round, closing the recursive loop. Our method achieves state-of-the-art overall performance of 77.89 on GameCraft-Bench. On GameASG-Bench, it achieves a strict task success rate of 53.2%, a 34.1% improvement over the same-model baseline, and the highest mean runtime-check pass rate at 93.4% among compared methods. A user study shows longer playtime and higher ratings. Code is coming soon.
1 INTRODUCTION
Runnable games do not necessarily provide a good player experience, and translating gameplay into experience-oriented revision remains challenging. Recursive Game Creator addresses this with a four-role iterative harness that uses gameplay and review feedback to refine games.
1 INTRODUCTION
- Runnable execution alone does not establish engaging play; the paper frames repeated playtesting and adaptation to intended players as necessary for product-level games.Experience depends on mechanics, difficulty, presentation, and responsive feedback.Source: §1 · p. 2
- The Coding-Native Player executes reusable policies through programmatic interfaces to collect frequent, diverse gameplay trajectories efficiently.Direct access to state and event information supports subsequent analysis, while separating fast policy execution from policy construction and visual review.Source: §1 · p. 2
- The Reviewer combines general measures with game-specific criteria, visual observations, and explicit or inferred preferences to guide revisions.Criteria can differ by genre, such as responsive steering in racing games or meaningful story and dialogue in visual novels.Source: §1 · p. 2
- The harness links Designer, Builder, Player, and Reviewer roles in a multi-round process for product-level game evolution and user-directed customization.The Designer plans gameplay, story, feedback, and asset requirements; the Builder implements the game and manages its assets and project.Source: §1 · p. 2
- Three refinement rounds raise GameCraft-Bench’s overall score from 72.70 to 77.89 (+5.19); the harness achieves the highest overall score on GameCraft-Bench and GameASG-Bench among baselines.On GameASG-Bench, it raises success from 9.47 to 25/47 tasks and records a 93.4% mean runtime-check pass rate.Source: §1 · pp. 2–3
2 RELATED WORK
Related work applies coding agents to video, embodied control, and game development, and separately studies iterative feedback and preference alignment. Recursive Game Creator connects executable play and preference analysis to repeated revision of the game artifact.
2 RELATED WORK
- Coding-agent applications span video generation, including Code2Video (Chen et al., 2025b) and VideoAgent, and embodied control, including Code as Policies (Liang et al., 2023).These applications motivate executable policies for gameplay testing.Source: §2 · p. 3
- Game-development agents include role-based collaboration such as GameGPT and engine or template workflows such as AutoUE and OpenGame (Yin et al., 2026, Jiang et al., 2026).The passage distinguishes producing or repairing runnable artifacts from establishing engaging player experience.Source: §2 · p. 3
- Recursive game-development work includes RSIGame, which uses iterative evaluation feedback to refine generated games.The paper distinguishes its focus by connecting policy generation, trajectory collection, experience and preference analysis, and multi-round revision.Source: §2 · p. 3
- Feedback research ranges from output revision methods such as Self-Refine and Reflexion (Madaan et al., 2023, Shinn et al., 2023) to preference alignment methods such as RLHF (Christiano et al., 2017, Ouyang et al., 2022).Prior game-related work also links player modeling or target affective patterns to content adaptation (Yannakakis & Togelius, 2011, Barthet et al., 2024).Source: §2 · p. 3
- The harness applies structured preference feedback to revise game artifacts while keeping model parameters fixed; explicit user requests and preferences inferred from play inform later rounds.Its workflow combines behavioral and visual evidence with shared and game-specific criteria for revision feedback.Source: §2 · p. 4, Figure 2 · p. 4
3 RECURSIVE GAME CREATOR, EXPERIENCE-ORIENTED GAME REFINEMENT
Recursive Game Creator organizes game refinement around a Designer, Builder, Player, and Reviewer. Its loop connects planning and construction to policy-based play, evidence-based review, version retention, and subsequent revision.
3 RECURSIVE GAME CREATOR, EXPERIENCE-ORIENTED GAME REFINEMENT
- The four-role loop links game construction, policy-based play, experience assessment, and iterative refinement across versions.The Designer translates requirements and evaluation feedback into a structured plan, while the Reviewer uses evidence to inform version retention and the next plan.Source: §3 · p. 4
3.1 DESIGNER & BUILDER
- The Designer turns the user brief, current game, and recent feedback into an experiential hypothesis, concrete changes, acceptance goals, and asset requests; the Builder develops and integrates a playable candidate.The Builder checks assets in gameplay and uses targeted self-checks alongside broader Player evidence.Source: §3 · p. 4, §3.1 · p. 5
3.2 CODING-NATIVE PLAYER
- Executable policies use exposed game state to choose actions and collect trajectories recording behavior and outcomes across varied strategies.Proxy strategies can reveal failures but do not establish that they represent the full human player population; human trajectories and explicit feedback add preference evidence.Source: §3.2 · p. 5
3.3 EXPERIENCE-ORIENTED REVIEWER
- The Reviewer assesses play using shared and game-specific criteria, combines trajectory patterns with visual findings and explicit user input, and compares candidate and retained versions to guide revisions and retention.Sampled rendered views complement behavioral records; when a failure's cause is uncertain, feedback can specify what further evidence would help resolve it.Source: §3.3 · p. 6
4 EXPERIMENTS
Experiments assess benchmark performance, iterative qualitative revisions, coding-native policy exploration, and expert-rated experience-oriented customization.
4.1.1 GAMECRAFT-BENCH
- 77.89 overall on GameCraft-Bench exceeds the strongest reported baseline by 6.37 points, and all five categories improve across three rounds.The score rises from 72.70 in round one to 77.89 in round three, a 5.19-point gain; it is 6.63 points above the same-model baseline.Source: Table 1 · p. 7, §4.1.1 · p. 7
4.1.2 GAMEASG-BENCH
- 53.2% of GameASG-Bench tasks pass; mean L1 and L2 pass rates are 98.3% and 93.4%, with the highest reported mean L2 and P2 scores.Task success requires valid delivery, completed evaluation, and passing every L1 check and applicable L2 P0/P1 check.Source: §4.1.2 · p. 7, Table 2 · p. 9
4.2 QUALITATIVE ANALYSIS OF MULTI-ROUND RSI
- Across action-game, visual-novel, and Meme Arena versions, successive revisions expand playable content and clarify game state, actions, and feedback.Action-game revisions add combat and progression feedback; visual-novel revisions use scene-grounded choices and illustrated events; Meme Arena revisions develop skill effects and movement guidance.Source: Figure 3 · p. 8, §4.2 · p. 9
4.3 PLAYER EXPLORATION ANALYSIS
- 48 coding-native policy rollouts cover 97.1% of the FPS test-scene grid, including 11 interiors; illustrative GUI routes are not a speed comparison.A reusable policy collects rollout evidence without a model call for every action.Source: Figure 4 · p. 10, §4.3 · p. 10
4.4 USER STUDY OF EXPERIENCE-ORIENTED CUSTOMIZATION
- From the first to third round, average playtime rises from 1.50 to 12.41 minutes, while controls, playability, depth, and art ratings also increase.The respective ratings change from 2 to 8, 1 to 7, 3 to 6, and 2 to 8 on a 1–10 scale.Source: Table 3 · p. 10, §4.4 · p. 10
5 CONCLUSION
Recursive Game Creator links design, tool-assisted implementation, executable play, and preference-informed review in a loop that supports continued game evolution. The conclusion reports strong benchmark performance and development examples and expert assessments of improvements.
5 CONCLUSION
- Recursive Game Creator achieves state-of-the-art overall performance on GameCraft-Bench and the highest mean runtime-check pass rate among compared methods on GameASG-Bench.Development examples and expert assessments characterize improvements in controls, content, and presentation.Source: §5 · p. 11
- Its coding-native Player collects reusable gameplay trajectories through programmatic interfaces, while the Reviewer combines behavioral evidence, visual observations, and user input to guide version selection and revisions.The policy-to-trajectory-to-feedback loop supports continued game evolution with fixed model parameters.Source: §5 · p. 11
- The harness connects design, tool-assisted implementation, executable play, and preference-informed review to provide a foundation for advancing functional prototypes toward product-level games.It also supports evaluating continued adaptation to individual players.Source: §5 · p. 11