[Paper Review] Specification-Driven Generation and Evaluation of Discrete-Event World Models via the DEVS Formalism

Zheyu Chen, Zhuohuan Li+1 more|arXiv (Cornell University)|Mar 4, 2026

Simulation Techniques and Applications0 citations

TL;DR

The paper presents a pipeline that synthesizes executable discrete-event world models from natural-language specifications using the DEVS formalism, with trace-based evaluation to ensure specification conformance and long-horizon consistency. It enables on-demand adaptation and verifiable, modular world models combining explicit structure with learned behavior.

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.

Motivation & Objective

• Motivate a middle ground between hand-engineered simulators and implicit neural models for world models in agentic systems.
• Propose accessible, executable discrete-event world models synthesized from natural-language specifications using the DEVS formalism.
• Develop a two-stage LLM-driven generation pipeline that separates structural inference from component-level behavior.
• Introduce a trace-based evaluation framework to verify models against specification-derived temporal and semantic constraints.
• Enable on-demand synthesis and adaptation of world models during online execution.

Proposed method

• Adopt the DEVS formalism to model world dynamics as interacting atomic and coupled components with explicit state, timing, and event semantics.
• Use a staged LLM-based generation pipeline that first performs structural synthesis (PlanTree) and then Behavioral Synthesis to implement atomic models and assemble coupled models.
• Employ interface contracts and JSON schemas to decouple components and enable parallel, scalable generation.
• Implement adaptive coupling via interface summarization to align parent models with ground-truth subcomponent interfaces.
• Generate simulators that emit structured event traces (JSONL) enabling specification-driven evaluation of observable behavior.
• Evaluate models via trace-based conformance against temporal/semantic constraints derived from natural-language specifications.

Experimental results

Research questions

• RQ1 How can we generate executable discrete-event world models from natural-language environment descriptions?
• RQ2 Can DEVS-based modular synthesis provide reliable long-horizon behavior and verifiability for generated simulators?
• RQ3 How can we evaluate such simulators when no unique ground-truth implementation exists?
• RQ4 Does a trace-based, specification-driven evaluation framework enable targeted diagnostics and robust adaptation during online execution?

Key findings

• The approach yields executable DEVS-based world models synthesized from natural language.
• Structural synthesis via a PlanTree enables scalable, contract-driven component generation.
• Behavioral synthesis creates per-component timing and state-transition logic that preserves modularity.
• Trace-based evaluation checks temporal and semantic constraints against specifications, enabling localization of violations.
• The framework supports on-demand synthesis and adaptive coupling to maintain consistency with the implemented components.