Research map / Multi-agent systems and games / Human-AI and goal-conditioned

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

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: World-model-based agents · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CascadeBench and enterprise discovery agents that query live ServiceNow configuration at inference, showing offline-trained world models degrade under configuration shift while runtime discovery stays robust.

Problem

Learned world models trained on historical enterprise transitions become brittle when tenant-specific business rules vary across deployments and evolve over time; the paper asks whether agents should instead discover dynamics from readable configuration at runtime.

Approach

Treat enterprise dynamics as a contextual transition function of instance configuration c and compare fine-tuned world models against a ReAct-style discovery agent that retrieves business rules, schemas, records and SLA definitions from the live instance via a single query tool at inference time, with a frozen LLM and no weight updates.

Key results

• On CascadeBench w/o BR, prompted models score about 7-12 IoU(T+F)-range while w/ BR they reach 38-51; discovery agent with GPT-5 reaches 32.1 IoU vs 9.82 prompted.
• Fine-tuned Gemma-4-31B reaches 91.6 in-distribution test IoU but falls to roughly 40-41 IoU on CascadeBench under shift.
• Discovery agent with Opus 4.6 raises WoW k=1 IoU from 0.40 to 0.45 and improves over the prompted baseline at every horizon k=1..5.

Surprising findings

• Fine-tuned Gemma-4-31B-LoRA underperforms the discovery agent-free baseline on WoW for some conditions, with discovery agents on 27-31B open-weight models bounded by tool-use reliability.
• SFT with business rules in context is not uniformly beneficial; several models gain little or regress.

Limitations

• Discovery agent assumes business rules are readable on the live instance; access controls reduce it to the prompted baseline.
• Evaluation limited to a single platform (ServiceNow) and a small set of open-weight models for the same-model comparison.
• Tier 3 execution-order dynamics are not recoverable from configuration and remain largely unaddressed.

Experimental setup

Datasets: Enterprise Gym transition corpus World of Workflows (WoW)

Benchmarks: CascadeBench WoW

Metrics: IoU IoU(T+F) state-prediction IoU across horizons k=1..5

Baselines: Claude Sonnet 4.6 Claude Opus 4.6 GPT-5 Gemini 3 Pro Qwen-3.5-27B Qwen-3.6-27B Gemma-4-31B World of Workflows

Builds on: ReAct World of Workflows LoRA SyGra

Closest papers by meaning

• EnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings (ICML 2026 · similarity 0.86)
• Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.85)
• Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback (NeurIPS 2026 · similarity 0.85)
• Synthesis Through Simulation: Generating Coherent Enterprise Data via Scalable Agent-System Interaction (NeurIPS 2026 · similarity 0.84)
• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.84)
• Beyond Prediction: Steering VLM Agents with Retrospective World Modeling (NeurIPS 2026 · similarity 0.84)