When a Verified World Model Still Loses: Play-Adequacy vs Prediction-Accuracy in LLM-Synthesized Code World Models

In the rapidly evolving landscape of artificial intelligence, Large Language Models (LLMs) are increasingly being leveraged to synthesise complex systems, including the rules of a game as executable code – known as Code World Models (CWMs). These CWMs are then typically integrated with classical planners to navigate and strategise within simulated environments. The prevailing method for validating these models has been to assess their transition accuracy on sampled trajectories, with high accuracy often being accepted as a benchmark for reliability.

However, new research from arXiv:2607.14169v1 presents a compelling argument that this focus on prediction accuracy might be fundamentally misguided when it comes to the practical application of CWMs in planning scenarios. The study posits that a CWM, even one achieving a perfect 100% transition accuracy on sampled data, can still be inadequate for effective planning. This introduces a critical distinction between 'prediction-accuracy' and 'play-adequacy'.

The researchers demonstrate four key findings that challenge current validation paradigms:

1. Passing the Sampling Gate, Failing the Plan: An LLM-synthesised CWM can achieve 100% transition accuracy and maintain $\geq 98\%$ state accuracy on the sampled data, yet still fall short when used for actual planning.
2. The Misleading Nature of High Accuracy: This high prediction accuracy can create a false sense of security, leading developers to believe the CWM is robust enough for planning, when in reality, it contains subtle flaws that undermine its utility.
3. Introducing Play-Adequacy: The study advocates for a new metric: 'play-adequacy'. This concept focuses on whether the CWM accurately represents the environment in a way that allows a planner to achieve its objectives, rather than merely predicting individual transitions correctly.
4. Implications for Future Development: The findings suggest a need for a paradigm shift in how CWMs are evaluated. Moving forward, validation methods should prioritise metrics that directly assess a model's fitness for purpose in planning contexts, rather than relying solely on traditional accuracy measures.

This research underscores a crucial point for the AI community: the metrics we choose to validate our models profoundly impact their real-world applicability. For LLM-synthesised CWMs, merely predicting the next state correctly might not be enough. The true test lies in whether the model facilitates successful planning and decision-making within the environment it purports to represent. As AI systems become more sophisticated and autonomous, ensuring their 'play-adequacy' will be paramount for their reliable deployment.