Specification-Driven Generation and Evaluation of Discrete-Event World Models via the DEVS Formalism
arXiv:2603.03784v1 Announce Type: new Abstract: World models are essential for planning and evaluation in agentic systems, yet existing approaches lie at two extremes: hand-engineered simulators that offer consistency and reproducibility but are costly to adapt, and implicit neural models that are flexible but difficult to constrain, verify, and debug over long horizons. We seek a principled middle ground that combines the reliability of explicit simulators with the flexibility of learned models, allowing world models to be adapted during online execution. By targeting a broad class of environments whose dynamics are governed by the ordering, timing, and causality of discrete events, such as queueing and service operations, embodied task planning, and message-mediated multi-agent coordination, we advocate explicit, executable discrete-event world models synthesized directly from natural-language specifications. Our approach adopts the DEVS formalism and introduces a staged LLM-based g
arXiv:2603.03784v1 Announce Type: new Abstract: World models are essential for planning and evaluation in agentic systems, yet existing approaches lie at two extremes: hand-engineered simulators that offer consistency and reproducibility but are costly to adapt, and implicit neural models that are flexible but difficult to constrain, verify, and debug over long horizons. We seek a principled middle ground that combines the reliability of explicit simulators with the flexibility of learned models, allowing world models to be adapted during online execution. By targeting a broad class of environments whose dynamics are governed by the ordering, timing, and causality of discrete events, such as queueing and service operations, embodied task planning, and message-mediated multi-agent coordination, we advocate explicit, executable discrete-event world models synthesized directly from natural-language specifications. Our approach adopts the DEVS formalism and introduces a staged LLM-based generation pipeline that separates structural inference of component interactions from component-level event and timing logic. To evaluate generated models without a unique ground truth, simulators emit structured event traces that are validated against specification-derived temporal and semantic constraints, enabling reproducible verification and localized diagnostics. Together, these contributions produce world models that are consistent over long-horizon rollouts, verifiable from observable behavior, and efficient to synthesize on demand during online execution.
Executive Summary
This article presents a novel approach to generating and evaluating discrete-event world models using the DEVS formalism and a staged LLM-based generation pipeline. The proposed method aims to balance the reliability of explicit simulators with the flexibility of learned models, enabling world models to be adapted during online execution. The approach is evaluated through structured event traces validated against specification-derived temporal and semantic constraints, allowing for reproducible verification and localized diagnostics. The resulting world models are consistent over long-horizon rollouts, verifiable from observable behavior, and efficient to synthesize on demand. This work has significant implications for planning and evaluation in agentic systems, particularly in environments governed by discrete events.
Key Points
- ▸ The article introduces a staged LLM-based generation pipeline that separates structural inference of component interactions from component-level event and timing logic.
- ▸ The proposed approach uses the DEVS formalism to synthesize explicit, executable discrete-event world models directly from natural-language specifications.
- ▸ The method evaluates generated models through structured event traces validated against specification-derived temporal and semantic constraints.
Merits
Improvement over existing approaches
The proposed method balances the reliability of explicit simulators with the flexibility of learned models, addressing the limitations of existing approaches that lie at two extremes.
Increased efficiency
The staged LLM-based generation pipeline enables world models to be synthesized on demand during online execution, making the approach more efficient than traditional methods.
Enhanced verification and diagnostics
The use of structured event traces and specification-derived temporal and semantic constraints allows for reproducible verification and localized diagnostics, improving the overall accuracy and reliability of the approach.
Demerits
Limited domain applicability
The proposed approach is specifically designed for environments governed by discrete events, which may limit its applicability to other domains.
Potential complexity of DEVS formalism
The use of the DEVS formalism may introduce complexity and require significant expertise, potentially limiting the accessibility of the approach.
Expert Commentary
The article presents a novel and significant contribution to the field of artificial intelligence, particularly in the area of world modeling and planning. The proposed approach has the potential to address some of the key limitations of existing methods, such as the balance between reliability and flexibility. However, the approach may require significant expertise in the DEVS formalism and the use of structured event traces. Furthermore, the limited domain applicability of the approach may limit its impact. Nevertheless, the work has significant implications for the development of agentic systems and highlights the importance of formal methods in software engineering.
Recommendations
- ✓ Further research is needed to explore the applicability of the proposed approach to other domains and to investigate the potential benefits of integrating the DEVS formalism with other formal methods.
- ✓ The authors should provide more detailed guidance on the implementation and use of the staged LLM-based generation pipeline, as well as the structured event traces and specification-derived temporal and semantic constraints.