---
license: apache-2.0
language:
- en
tags:
- arxiv:2605.12178
- servicenow
- enterprise
- world-models
- workflows
- business-rules
- agents
- benchmark
pretty_name: CascadeBench
size_categories:
- n<1K
task_categories:
- text-generation
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
---

<div align="center">

<h1>CascadeBench: Do Enterprise Systems Need Learned World Models?</h1>

<p><a href="https://arxiv.org/abs/2605.12178"><img src="https://img.shields.io/badge/Paper-red?logo=arxiv&logoColor=white" alt="Paper" /></a> <a href="https://neurips.cc/Conferences/2026"><img src="https://img.shields.io/badge/NeurIPS-2026%20Main%20Track-purple" alt="NeurIPS 2026" /></a> <a href="https://github.com/ServiceNow/SyGra/tree/scratch/ewm/tasks/examples/wow_state_predictor_da"><img src="https://img.shields.io/badge/Discovery%20Agent-GitHub-black?logo=github" alt="Discovery Agent" /></a></p>

<p>🎉 <b>Accepted to NeurIPS 2026: The Fortieth Annual Conference on Neural Information Processing Systems (Main Track)</b></p>

<p><i>A reasoning-focused benchmark for predicting enterprise business-rule cascades, built on synthetic schemas with rule-level attribution of every field change</i></p>

</div>

<div align="center"><img src="assets/teaser.png" alt="CascadeBench overview" width="90%" /></div>

## About

In enterprise systems, the dynamics come from tenant-specific business logic that varies across deployments and changes over time. Business rules, workflows and schema defaults decide what happens when a record changes. The same action can have different effects on different instances.

**CascadeBench** tests whether a model can predict those effects. Each example gives the current state $s_t$ and an action $a_t$. The model predicts the next state $s_{t+1}$: every field-level change across every table, including all changes made by the chain of business rules the action triggers.

The benchmark accompanies the paper [*Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics*](https://arxiv.org/abs/2605.12178). The paper finds that:

- Offline-trained world models perform well in-distribution but degrade as configurations change. **Enterprise discovery agents** stay more robust because they read the active rules at runtime.
- **Having the rules is necessary but not sufficient.** Accuracy still drops sharply as cascades compose, even when the active rules are in the prompt. Multi-step rule composition limits performance more than retrieval does.

## Key Features

- 🧪 **Real dynamics.** A live ServiceNow rule engine produced every transition. Nothing is simulated.
- 🏷️ **Rule-level attribution.** A custom execution log traces each field change in the audit log to the business rule that caused it.
- 🔒 **Synthetic surface form.** Table and field names are freshly generated with a `u_` prefix, and reserved product namespaces are excluded. Models can't rely on recalling table structures they may have seen in pretraining.
- 📦 **Full context per example.** Each example includes table schemas, business rules, seed records and supporting records. You control how much of this the model sees.
- 🧹 **Clean ground truth.** Audits are restricted to content fields. System IDs, timestamps and bookkeeping fields are removed.
- ✅ **Validated cascades.** Every business rule passes 14 deterministic checks (schema correctness, cycle detection, filter validity, script safety) and is verified through execution.

## Code

- **Discovery agent:** [ServiceNow/SyGra › `wow_state_predictor_da`](https://github.com/ServiceNow/SyGra/tree/scratch/ewm/tasks/examples/wow_state_predictor_da)
- **Dataset construction and evaluation code:** coming soon! 🚧

## Dataset Summary

| | |
|---|---|
| Samples | 37 (one per workflow domain) |
| Domains | 37, e.g. `accounts_payable_processing`, `change_management`, `incident_escalation` |
| Tables per sample | ~10 synthetic tables with foreign-key relationships |
| Business rules per sample | 4–7 (204 total) |
| Cascade topologies | `linear` (21), `flat` (16) |
| Triggering operations | `update` (29), `insert` (8) |
| Audit records (deduped) | 1,395 field-level changes |
| Split | `train` |

### Complexity Tiers

Ground-truth changes are stratified into three tiers:

| Tier | Name | What it covers | Metric |
|---|---|---|---|
| T1 | Schema-deterministic | Defaults, constraints and choices on the action's own table | IoU(T+F) |
| T2 | Rule-composable | Cross-table cascades that require at least one business rule to fire | IoU(T+F) |
| T3 | Execution-inferred | Conflicts where two or more rules write different values to the same field | Strict IoU on (table, field, value) |

## Field Descriptions

| Field | Type | Description |
|---|---|---|
| `domain` | `string` | Workflow domain. Unique per row, so it can serve as a sample ID |
| `topology` | `string` | Cascade topology (`linear` or `flat`) |
| `total_brs_fired` | `int` | Number of business rules that actually fired |
| `expected_br_count` | `int` | Number of business rules expected to fire |
| `audit_count` | `int` | Number of deduplicated audit records |
| `raw_audit_count` | `int` | Number of raw audit records |
| `tool_name` | `string` | The action invoked, $a_t$ |
| `parameters` | `string` (JSON) | `{table_name, operation, fields}` for the action. `fields` keys are domain-specific |
| `seed_data` | `string` (JSON) | Initial state of the target record. Empty for `insert` actions |
| `supporting_data` | `string` (JSON) | Map from related table name to its rows |
| `schema` | `string` (JSON) | Map from table name to its column schema, including foreign keys |
| `business_rules` | `list[struct]` | `name, fires_on_table, filter_condition, trigger_sequence, trigger_type, order, script` |
| `ewm_logs` | `list[struct]` | Execution-log entries attributing each change to a rule (`u_br_name`, `u_table_name`, `u_field_name`, `u_old_value`, `u_new_value`, …) |
| `audits` | `list[struct]` | Ground truth $s_{t+1}$: deduplicated field-level audit (`tablename, fieldname, oldvalue, newvalue, documentkey`) |
| `raw_audits` | `list[struct]` | Raw field-level audit, same shape as `audits` |

> Table and field names in the four JSON-string columns (`parameters`, `seed_data`, `supporting_data`, `schema`) differ per domain. They are stored as serialized JSON to keep the Arrow schema stable. Call `json.loads()` to use them.

## Usage

```python
from datasets import load_dataset
import json

ds = load_dataset("ServiceNow-AI/cascade_bench", split="train")

row = ds[0]
print(row["domain"], row["tool_name"])

schema = json.loads(row["schema"])        # table -> column schema
params = json.loads(row["parameters"])    # the action a_t
seed = json.loads(row["seed_data"])       # current state s_t (target record)
rules = row["business_rules"]             # rules that may fire
gold = row["audits"]                      # ground-truth field changes s_{t+1}
```

### Evaluation Settings

The paper evaluates three settings, which differ in how much context the model gets:

| Setting | Context given to the model |
|---|---|
| **Direct** | Action and state only. No rules and no retrieval |
| **Discovery Agent** | Retrieves rules and schema from the system at inference time |
| **Oracle** | Active business rules supplied in the prompt |

## Example Use Cases

- **Benchmark world models** and transition predictors on enterprise systems whose dynamics are specific to each deployment.
- **Compare internalized and runtime-discovered dynamics** by varying which context fields the model sees.
- **Study multi-step rule composition** by using rule-level attribution to measure how accuracy drops as rule hops increase.
- **Evaluate discovery agents** that query schemas, workflow definitions and business rules before acting.

## Citation

```bibtex
@misc{nair2026enterprisesystemsneedlearned,
      title={Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics},
      author={Jishnu Sethumadhavan Nair and Patrice Bechard and Rishabh Maheshwary and Surajit Dasgupta and Sravan Ramachandran and Aakash Bhagat and Shruthan Radhakrishna and Pulkit Pattnaik and Johan Obando-Ceron and Shiva Krishna Reddy Malay and Sagar Davasam and Seganrasan Subramanian and Vipul Mittal and Sridhar Krishna Nemala and Christopher Pal and Srinivas Sunkara and Sai Rajeswar},
      year={2026},
      eprint={2605.12178},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.12178},
}
```