Researchers propose Transition Complexity Profile to standardize difficulty measurement in game world modeling and RL

A new position paper introduces the Transition Complexity Profile (TCP), a small, reproducible set of metrics designed to characterize the difficulty of the transition prediction problem in game world modeling (GWM) and reinforcement learning (RL) environments. The authors argue that current research often conflates environment difficulty with model performance due to the lack of standardized difficulty metrics at the declared interface (e.g., pixels, tokens, or latents with finite history).

The TCP framework proposes three core components: (i) intrinsic one-step branching, which measures the number of possible next states from a given state; (ii) interaction-induced uncertainty and opponent influence, which captures variability introduced by multi-agent interactions when observable; and (iii) temporal and spatial dependency span, assessed via standardized probe curves. Each component is reported alongside an explicit reference distribution, protocol stochasticity, and a versioned measurement budget that includes sampling/resampling and fixed probe compute, enabling comparable numbers across benchmarks.

The paper outlines how common game families and modern "neural game engine" domains populate the TCP landscape, providing a qualitative map of where different environments fall in terms of transition complexity. The authors advocate for TCP to become standard benchmark metadata and a required statistic in GWM and RL papers, positioning it as a necessary step toward more rigorous and reproducible research in these fields.

The proposal was accepted by the ICML 2026 Position Paper Track and is authored by Lele Cao. While the paper does not introduce empirical results, it frames TCP as a methodological contribution aimed at improving the interpretability and comparability of research findings in game world modeling and reinforcement learning.