AIAny
Icon for item

Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models

Proposes treating game development as a recursive data engine and introduces RLHEV (Reinforcement Learning with Human-Engine Verification) to combine dense engine checks (collision, physics, navigability) with human acceptance feedback, producing trajectory data and rewards for post-training world models.

Introduction

Why this matters

Most spatial/world-model research relies on scraped video and fuzzy proxies (CLIP, FVD) that provide noisy, biased rewards and poorly support iterative RL post-training. The paper's core insight is that game development naturally exposes executable, verifiable world specifications: game engines provide dense, low-cost structural checks (collision, physics, navigability, bounded playability) while developers provide sparse but highly aligned acceptance judgments. Combining these two signals creates a recursive feedback engine that can drive RL-based post-training and iterative improvement of world models.

Key Findings
  • A concrete post-training paradigm: RLHEV (Reinforcement Learning with Human-Engine Verification) — uses engine checks as dense rewards and developer accept/reject decisions as sparse, aligned supervision. So what: enables continuous RL improvement rather than relying solely on static proxy metrics.

  • An agent and workflow: Agentic World Model (AWoMo) — proposes scene edits, executes them in engines, observes failures and human decisions, and converts accepted/ repaired traces into training data. So what: turns development traces into actionable multimodal trajectories for learning.

  • Empirical signals of benefit: on a 200-example UnitySceneBench edit evaluation, RLHEV achieves top scores; AWoMo-augmented training shows positive transfer to other engines (Unreal, Godot) and improves embodied-task performance on benchmarks such as R2R and Gym/D4RL MuJoCo. So what: these results suggest the approach yields transferable gains across engines and downstream embodied tasks.

  • Practical advantage over current proxies: engine execution exposes precise, local failures that are inexpensive to check (e.g., collisions, unreachable areas), making reward signals less gameable than global perceptual scores. So what: this opens a scalable path to RL-driven spatial model scaling analogous to how executable code enabled RL improvements in coding agents.

Who it's for and tradeoffs

Great fit if you research or build world models, multimodal agents, or scene synthesis and can instrument a game engine or development pipeline — the approach needs an executable environment and access to developer verification signals. Look elsewhere if your target application cannot be expressed as an executable scene (purely generative video without an underlying simulator) or you cannot collect developer feedback at scale. The method trades human labeling cost for much higher-quality, iterative supervision: engine checks provide dense cheap signals, but the loop still depends on developer acceptance to align global goals.

Where it fits

This work positions game development as a missing piece between large-scale scraping and an RL-driven, self-improving loop for spatial intelligence. It complements benchmarks and agentic evaluation efforts by supplying a practical reward engine and a recipe for converting development traces into training data.

Information

  • Websitearxiv.org
  • OrganizationsInfRec, Cardinal AI Lab, University of California, Berkeley, Independent Researcher, Hong Kong University of Science and Technology
  • AuthorsPengfei Zhou, Hexin Wang, Zhengfeiyang Zhang, Yixing Ma, Zhenglin Wan, Kaipeng Zhang, Wangbo Zhao, Yang You
  • Published date2026/08/26

More Items

Provides a real-scale 3D Hong Kong sandbox to evaluate whether multimodal LLM agents can turn local street-view perception into sustained spatial action, supporting closed-loop first-person interaction, an interactive map, and controlled tests of grounding, long-range navigation, and robustness.

Trains compact conversational agents to adapt at runtime to changing 'Harness' configurations (Skills, Hooks, prompts, tools) using Harness-Aware Training (HAT): Harness-State Augmentation, on-policy distillation, and RL to preserve generality while meeting low-latency deployment constraints.

Analyzes how to generate useful interaction data for LLM agents and proposes the ACE lens — Accuracy, Complexity, divErsity — while factorizing agentic data as (E, q, τ, v). Surveys verification, difficulty calibration, and coverage strategies and outlines implications for training and benchmarks.