jishnunair commited on
Commit
0ef6210
·
verified ·
1 Parent(s): f7bcf8d

Update dataset card: paper details, NeurIPS 2026, teaser, citation

Browse files
Files changed (2) hide show
  1. README.md +118 -34
  2. assets/teaser.png +3 -0
README.md CHANGED
@@ -3,14 +3,19 @@ license: other
3
  language:
4
  - en
5
  tags:
 
6
  - servicenow
 
 
7
  - workflows
8
  - business-rules
9
  - agents
10
  - benchmark
11
- pretty_name: Cascade Bench
12
  size_categories:
13
  - n<1K
 
 
14
  configs:
15
  - config_name: default
16
  data_files:
@@ -18,42 +23,87 @@ configs:
18
  path: data/train.jsonl
19
  ---
20
 
21
- # Cascade Bench
22
 
23
- A benchmark of ServiceNow workflow trajectories ("World of Workflows++"). Each row is a
24
- single tool action applied to a domain, together with the seed/supporting data, the table
25
- schemas, the business rules that can fire, and the resulting cascade of execution logs and
26
- audit records.
27
 
28
- ## Summary
 
 
 
29
 
30
- - **37 samples**, one per workflow domain (e.g. `accounts_payable_processing`,
31
- `incident_escalation`, `change_management`).
32
- - One split: `train`.
33
 
34
- ## Columns
35
 
36
- | Column | Type | Description |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37
  |---|---|---|
38
- | `domain` | string | Workflow domain / sample id |
39
- | `topology` | string | Cascade topology (e.g. `linear`) |
40
- | `total_brs_fired` | int | Business rules that actually fired |
41
- | `expected_br_count` | int | Expected number of business rules |
42
- | `audit_count` | int | Number of deduped audit records |
43
- | `raw_audit_count` | int | Number of raw audit records |
44
- | `tool_name` | string | The action invoked |
45
- | `parameters` | string (JSON) | `{table_name, operation, fields}` — `fields` keys are domain-specific |
46
- | `seed_data` | string (JSON) | Initial record state for the target table |
47
- | `supporting_data` | string (JSON) | Map of related table name → rows |
48
- | `schema` | string (JSON) | Map of table name → column schema |
49
- | `business_rules` | list[struct] | `name, fires_on_table, filter_condition, trigger_sequence, trigger_type, order, script` |
50
- | `ewm_logs` | list[struct] | Execution-log entries (`u_br_name`, `u_table_name`, `u_field_name`, `u_old_value`, `u_new_value`, …) |
51
- | `audits` | list[struct] | Deduped field-level audit (`tablename, fieldname, oldvalue, newvalue, documentkey`) |
52
- | `raw_audits` | list[struct] | Raw field-level audit, same shape as `audits` |
53
-
54
- > The four JSON-string columns (`parameters`, `seed_data`, `supporting_data`, `schema`)
55
- > hold dicts whose keys (table/field names) differ per domain, so they are stored as
56
- > serialized JSON to keep a stable Arrow schema. Call `json.loads()` to use them.
57
 
58
  ## Usage
59
 
@@ -65,7 +115,41 @@ ds = load_dataset("ServiceNow-AI/cascade_bench", split="train")
65
 
66
  row = ds[0]
67
  print(row["domain"], row["tool_name"])
68
- schema = json.loads(row["schema"]) # variable-key dict
69
- params = json.loads(row["parameters"])
70
- brs = row["business_rules"] # native list of structs
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71
  ```
 
3
  language:
4
  - en
5
  tags:
6
+ - arxiv:2605.12178
7
  - servicenow
8
+ - enterprise
9
+ - world-models
10
  - workflows
11
  - business-rules
12
  - agents
13
  - benchmark
14
+ pretty_name: CascadeBench
15
  size_categories:
16
  - n<1K
17
+ task_categories:
18
+ - text-generation
19
  configs:
20
  - config_name: default
21
  data_files:
 
23
  path: data/train.jsonl
24
  ---
25
 
26
+ <div align="center">
27
 
28
+ <h1>CascadeBench: Do Enterprise Systems Need Learned World Models?</h1>
 
 
 
29
 
30
+ <p>
31
+ <a href="https://arxiv.org/abs/2605.12178"><img src="https://img.shields.io/badge/Paper-red?logo=arxiv&logoColor=white" /></a>
32
+ <img src="https://img.shields.io/badge/NeurIPS-2026%20Main%20Track-purple" />
33
+ </p>
34
 
35
+ <p>🎉 <b>Accepted to NeurIPS 2026: The Fortieth Annual Conference on Neural Information Processing Systems (Main Track)</b></p>
 
 
36
 
37
+ <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>
38
 
39
+ </div>
40
+
41
+ <div align="center"><img src="assets/teaser.png" alt="CascadeBench overview" width="90%" /></div>
42
+
43
+ ## About
44
+
45
+ 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.
46
+
47
+ **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.
48
+
49
+ 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:
50
+
51
+ - 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.
52
+ - **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.
53
+
54
+ ## Key Features
55
+
56
+ - 🧪 **Real dynamics.** A live ServiceNow rule engine produced every transition. Nothing is simulated.
57
+ - 🏷️ **Rule-level attribution.** A custom execution log traces each field change in the audit log to the business rule that caused it.
58
+ - 🔒 **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.
59
+ - 📦 **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.
60
+ - 🧹 **Clean ground truth.** Audits are restricted to content fields. System IDs, timestamps and bookkeeping fields are removed.
61
+ - ✅ **Validated cascades.** Every business rule passes 14 deterministic checks (schema correctness, cycle detection, filter validity, script safety) and is verified through execution.
62
+
63
+ ## Dataset Summary
64
+
65
+ | | |
66
+ |---|---|
67
+ | Samples | 37 (one per workflow domain) |
68
+ | Domains | 37, e.g. `accounts_payable_processing`, `change_management`, `incident_escalation` |
69
+ | Tables per sample | ~10 synthetic tables with foreign-key relationships |
70
+ | Business rules per sample | 4–7 (204 total) |
71
+ | Cascade topologies | `linear` (21), `flat` (16) |
72
+ | Triggering operations | `update` (29), `insert` (8) |
73
+ | Audit records (deduped) | 1,395 field-level changes |
74
+ | Split | `train` |
75
+
76
+ ### Complexity Tiers
77
+
78
+ Ground-truth changes are stratified into three tiers:
79
+
80
+ | Tier | Name | What it covers | Metric |
81
+ |---|---|---|---|
82
+ | T1 | Schema-deterministic | Defaults, constraints and choices on the action's own table | IoU(T+F) |
83
+ | T2 | Rule-composable | Cross-table cascades that require at least one business rule to fire | IoU(T+F) |
84
+ | T3 | Execution-inferred | Conflicts where two or more rules write different values to the same field | Strict IoU on (table, field, value) |
85
+
86
+ ## Field Descriptions
87
+
88
+ | Field | Type | Description |
89
  |---|---|---|
90
+ | `domain` | `string` | Workflow domain. Unique per row, so it can serve as a sample ID |
91
+ | `topology` | `string` | Cascade topology (`linear` or `flat`) |
92
+ | `total_brs_fired` | `int` | Number of business rules that actually fired |
93
+ | `expected_br_count` | `int` | Number of business rules expected to fire |
94
+ | `audit_count` | `int` | Number of deduplicated audit records |
95
+ | `raw_audit_count` | `int` | Number of raw audit records |
96
+ | `tool_name` | `string` | The action invoked, $a_t$ |
97
+ | `parameters` | `string` (JSON) | `{table_name, operation, fields}` for the action. `fields` keys are domain-specific |
98
+ | `seed_data` | `string` (JSON) | Initial state of the target record. Empty for `insert` actions |
99
+ | `supporting_data` | `string` (JSON) | Map from related table name to its rows |
100
+ | `schema` | `string` (JSON) | Map from table name to its column schema, including foreign keys |
101
+ | `business_rules` | `list[struct]` | `name, fires_on_table, filter_condition, trigger_sequence, trigger_type, order, script` |
102
+ | `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`, …) |
103
+ | `audits` | `list[struct]` | Ground truth $s_{t+1}$: deduplicated field-level audit (`tablename, fieldname, oldvalue, newvalue, documentkey`) |
104
+ | `raw_audits` | `list[struct]` | Raw field-level audit, same shape as `audits` |
105
+
106
+ > 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.
 
 
107
 
108
  ## Usage
109
 
 
115
 
116
  row = ds[0]
117
  print(row["domain"], row["tool_name"])
118
+
119
+ schema = json.loads(row["schema"]) # table -> column schema
120
+ params = json.loads(row["parameters"]) # the action a_t
121
+ seed = json.loads(row["seed_data"]) # current state s_t (target record)
122
+ rules = row["business_rules"] # rules that may fire
123
+ gold = row["audits"] # ground-truth field changes s_{t+1}
124
+ ```
125
+
126
+ ### Evaluation Settings
127
+
128
+ The paper evaluates three settings, which differ in how much context the model gets:
129
+
130
+ | Setting | Context given to the model |
131
+ |---|---|
132
+ | **Direct** | Action and state only. No rules and no retrieval |
133
+ | **Discovery Agent** | Retrieves rules and schema from the system at inference time |
134
+ | **Oracle** | Active business rules supplied in the prompt |
135
+
136
+ ## Example Use Cases
137
+
138
+ - **Benchmark world models** and transition predictors on enterprise systems whose dynamics are specific to each deployment.
139
+ - **Compare internalized and runtime-discovered dynamics** by varying which context fields the model sees.
140
+ - **Study multi-step rule composition** by using rule-level attribution to measure how accuracy drops as rule hops increase.
141
+ - **Evaluate discovery agents** that query schemas, workflow definitions and business rules before acting.
142
+
143
+ ## Citation
144
+
145
+ ```bibtex
146
+ @misc{nair2026enterprisesystemsneedlearned,
147
+ title={Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics},
148
+ 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},
149
+ year={2026},
150
+ eprint={2605.12178},
151
+ archivePrefix={arXiv},
152
+ primaryClass={cs.AI},
153
+ url={https://arxiv.org/abs/2605.12178},
154
+ }
155
  ```
assets/teaser.png ADDED

Git LFS Details

  • SHA256: d6c6e163a117043146005587644efa382bee9ed4743ce65acf9f4743b0733ca1
  • Pointer size: 131 Bytes
  • Size of remote file: 634 kB