Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics

Abstract: World models enable agents to anticipate the effects of their actions by internalizing environment dynamics. In enterprise systems, however, these dynamics are often defined by tenant-specific business logic that varies across deployments and evolves over time, making models trained on historical transitions brittle under deployment shift. We ask a question the world-models literature has not addressed: when the rules can be read at inference time, does an agent still need to learn them? We argue, and demonstrate empirically, that in settings where transition dynamics are configurable and readable, runtime discovery complements offline training by grounding predictions in the active system instance. We propose enterprise discovery agents, which recover relevant transition dynamics at runtime by reading the system's configuration rather than relying solely on internalized representations. We introduce CascadeBench, a reasoning-focused benchmark for enterprise cascade prediction that adopts the evaluation methodology of World of Workflows on diverse synthetic environments, and use it together with deployment-shift evaluation to show that offline-trained world models can perform well in-distribution but degrade as dynamics change, whereas discovery-based agents are more robust under shift by grounding their predictions in the current instance. Our findings suggest that, in configurable enterprise environments, agents should not rely solely on fixed internalized dynamics, but should incorporate mechanisms for discovering relevant transition logic at runtime.

• The paper shows that static learned world models fail in dynamic enterprise environments, emphasizing the need to retrieve live configuration context for accurate state prediction.
• It introduces CascadeBench, a benchmark that simulates cascading transitions across varied enterprise configurations with precise ground-truth datasets.
• Empirical results reveal that discovery agents outperform fine-tuned models, particularly in high-complexity transition tiers under deployment shifts.

Contextual Dynamics in Enterprise System Modeling: An Analysis of "Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics"

Problem Motivation and Enterprise Dynamics

Enterprise systems are fundamentally shaped by tenant-specific configuration artifacts—business rules, workflows, policies—that evolve continuously and vary across deployments. These configurations introduce environment dynamics that are neither stationary nor globally consistent, contrasting sharply with traditional world model assumptions where the transition function is fixed and unknown. As delineated in the paper, this gives rise to a contextual transition model:

$s_{t+1} \sim P(s_{t+1} \mid s_t, a_t, c)$

where $c$ represents the instance configuration that is both explicit and readable in enterprise platforms.

Such variability induces distribution shifts not encountered in classical model-based RL or fixed-world environments. Consequently, LLM agents, which are increasingly leveraged for enterprise automation and planning, face brittle generalization if restricted to internalized, static world models. The paper thus interrogates the fundamental utility of learned world models in scenarios where system transition logic remains inspectable at inference time through configuration queries, categorizing transition complexity into three tiers:

• Tier 1: Schema-determined transitions fully specified by data dictionary;
• Tier 2: Rule-composable cascades requiring composition of explicit, configurable business rules;
• Tier 3: Execution-inferred transitions contingent on platform-internal behaviors (e.g. rule execution order, race conditions).

CascadeBench: Benchmarking Under Configuration Shift

To dissect the impact of configurability and instance-specific logic, the authors introduce CascadeBench, a synthetic yet highly-structured benchmark that extends "World of Workflows" methodology to diverse, systematically generated enterprise environments.

Figure 1: Overview of CascadeBench and conceptual comparison between learned world models—trained on historical data—and enterprise discovery agents that recover transition logic from live configurations.

CascadeBench enables fine-grained control of environment diversity (industries, organizational sizes, rule families) and precise delivery of relevant context (schemas, business rules), isolating reasoning capabilities from memorization or retrieval artifacts that confound real-world audit logs. The task—predicting the next state $s_{t+1}$ given $(s_t, a_t)$—necessitates accurate modeling of cross-table cascades and multi-step rule firings.

Figure 2: The data generation pipeline: audit logs of isolated, sandboxed action executions are subjected to normalization and filtering, resulting in a ground-truth dataset of state transitions and labeled cascade paths.

Experimental Framework: Prompting, SFT, and Discovery Agents

Three principal approaches to transition modeling are compared:

• Prompted Baseline: Zero-shot inference by general-purpose (frontier and base) LLMs on minimally contextualized prompts;
• Learned World Models: Models fine-tuned over transition data, internalizing transition dynamics in their parameters;
• Enterprise Discovery Agents: Agents augmented with runtime retrieval, enabling on-the-fly inspection of current configuration artifacts (business rules, schemas, records) to predict action effects.

The discovery agent follows a retrieve-then-reason paradigm, adaptively selecting relevant context $\tilde{c} \subseteq c$ at inference. By comparing zero-shot, fine-tuned, and discovery-augmented agents—across both in-distribution and deployment-shifted settings—the study exposes the inductive limitations of purely learned world models for systems with evolving logic.

Empirical Results and Analysis

In-Distribution vs. Configuration Shift

Fine-tuned world models achieve strong in-distribution performance (IoU up to 91.6 for Gemma-4-31B), demonstrating effective memorization of training-regime dynamics when business rules are static. However, this performance collapses under deployment shift in CascadeBench, reverting to around 41 IoU—a minimal advantage over base model prompting with full rule context. This delineates a sharp generalization boundary for SFT approaches: internalization of enterprise transitions is not robust across diverse or unseen configurations.

On the Limits of Prompting and SFT

When deprived of business rule context, both prompted and fine-tuned models perform near chance on CascadeBench (IoU $<$ 13), contrasting with their moderate performance in other benchmarks and emphasizing that business rules encode the critical causal dynamics. SFT offers marginal gains when context is absent, but is ineffective in substituting for the active, ground-truth configuration necessary for cross-instance generalization.

Superiority of Runtime Discovery Methods

The discovery agent architecture consistently improves multi-step state prediction accuracy over static prompting—across all rollout horizons $k=1$ to $5$—and substantially narrows the gap to the in-context “oracle” ceiling.

Figure 3: The Discovery Agent maintains superior IoU over prompted baselines at all rollout horizons, mitigating error compounding when retrieving transition logic on-demand.

Notably, runtime discovery is most advantageous for transitions of higher complexity (Tiers 2 and 3), where static schema reasoning is insufficient and configuration-dependent rule composition dominates.

Figure 4: Tier-stratified assessment on CascadeBench: Prompted models suffice for simple, schema-limited effects, but exhibit steep degradation as transition cascades become more complex. Runtime discovery recovers much of the attainable information for Tiers 1 and 2.

Crucially, discovery agents can match or exceed fine-tuned models (even with no retraining) in deployment-shifted and cross-tenant scenarios. The main residual failure modes involve deep cascade resolution and record insertions, highlighting the persistent reasoning challenges for multi-step transitions even given perfect retrieval.

Implications and Future Directions

This research presents a decisive claim: in dynamically configurable enterprise settings where the transition function is explicit and inspectable, offline-trained world models alone are insufficient for reliable cross-instance or cross-tenant transition prediction. Instead, robust agents should integrate learned priors with runtime retrieval and compositional reasoning over the instance’s active configuration.

Practically, this motivates departure from monolithic, internalized simulators for enterprise automation agents, and a shift toward architectures capable of adaptive, configuration-sensitive reasoning. Theoretically, it challenges prevailing world-model paradigms for agent planning in real-world, non-stationary environments with readable dynamics. The distinction between what should be learned vs. what should be discovered at inference is particularly salient for highly customizable multi-tenant platforms (e.g., ServiceNow, Salesforce).

Future research should examine scalable methods for learning retrieval policies (when, what, and how to retrieve context), the interplay between internalization and runtime grounding, and the generalizability of these findings to other high-variability software systems. Additionally, addressing the limitations in cross-table deep cascade prediction and learning compositional multi-hop rule application remains an open area for improvement.

Conclusion

The findings of this work rigorously demonstrate that when the governing logic of an environment is both readable and non-stationary, world model approaches anchored solely in fixed transition internalization are inadequate. Enterprise discovery agents—entities that couple off-policy learning with adaptive context retrieval—exhibit superior robustness under variation in deployment configuration. These results have immediate relevance for the design and deployment of enterprise AI systems, supporting hybrid architectures that incorporate both model-based priors and runtime discovery capabilities for accurate, context-sensitive action modeling.