---
license: apache-2.0
pretty_name: GlobeBench
language:
- en
task_categories:
- text-generation
tags:
- world-models
- benchmark
- agents
- llm-evaluation
size_categories:
- 1K<n<10K
configs:
- config_name: items
  data_files:
  - split: test
    path: data/items.jsonl
- config_name: results
  data_files:
  - split: test
    path: data/results.jsonl
---

# GlobeBench

GlobeBench evaluates whether a language model can act as a text-based world model: given an agent's trajectory of actions and environment observations, it must predict the environment's response to the next action. The benchmark has 1,632 execution-grounded items across five domains (Code, Enterprise, Games, Math and Science, Dialogue) and 19 simulation skills.

- Code, full model outputs, and judge records: https://github.com/patronus-ai/globebench

## Configurations

### `items` (1,632 rows)

| Field | Description |
|---|---|
| `example_id` | Item identifier |
| `messages` | The rendered trajectory shown to the model (system prompt, prior turns, final action) |
| `target` | The execution-derived gold next observation |
| `metadata` | Domain, primary skill (`primary_sub_eval`), all tested skills (`sub_evals`), reach-back (`h_reachback`), impact radius (`radius_level`), scoring contract, source, and source license |

### `results` (32,640 rows: 20 models x 1,632 items)

| Field | Description |
|---|---|
| `model` | Model name as reported in the paper |
| `model_key` | Model identifier used in the runs |
| `item_id` | Item identifier (matches `example_id`) |
| `prediction` | The model's predicted observation (null when there was no usable response) |
| `served_state` | `served`, or the reason no usable response was returned |
| `refusal` | True for provider refusals (`finish_reason: content_filter`) |
| `passed` | Final task-success decision under the paper's hybrid grader |
| `resolution_rule` | Whether the deterministic grader or the LLM judge decided the outcome |
| `format_outcome` | Format compliance, reported separately from correctness |

Rows without a usable response count as failures. Pooled task success is 6,423 / 32,640 = 19.68%.

## Leaderboard

| Rank | Model | Task success |
|---:|---|---:|
| 1 | GPT-6 Astra | 49.57% |
| 2 | Grok 4.6 | 28.13% |
| 3 | Claude Opus 5.5 | 26.35% |
| 4 | Qwen 3.8 Max | 25.31% |
| 5 | GPT-5.5 | 25.18% |
| 6 | GPT-6 Sol | 25.00% |
| 7 | GPT-5.6 Sol | 23.65% |
| 8 | Qwen 3.8 Flash Next | 23.22% |
| 9 | Gemini 3.7 Flash | 21.75% |
| 10 | Claude Fable 5.1 | 21.63% |
| 11 | Claude Opus 4.8 | 20.04% |
| 12 | GPT-6 Luna | 17.77% |
| 13 | Claude Opus 5 | 15.93% |
| 14 | GLM 5.3 | 14.77% |
| 15 | Qwen-AgentWorld-35B-A3B | 14.46% |
| 16 | Kimi K3 | 13.97% |
| 17 | Inkling | 11.89% |
| 18 | Claude Fable 5 | 7.29% |
| 19 | Digital World Model (DWM) | 5.21% |
| 20 | Laguna S2.1 | 2.45% |

## Scoring

Each prediction is first compared with the gold by a deterministic grader that accepts differences of notation but not of state. When the grader can neither certify a match nor find a concrete contradiction, an LLM judge (GPT-5.6 Terra, with GPT-5.6 Sol as fallback) decides. The grader, judge prompt, and an offline rescoring script are in the GitHub repository.

## License

The code and the items authored by Patronus AI are released under Apache-2.0. Items adapted from upstream projects keep their source license, recorded per item in `metadata.source_license`; one item derived from python-chess is GPL-3.0-or-later. Model outputs may also be subject to the respective providers' terms. See the GitHub repository's `THIRD_PARTY_NOTICES.md` for details.

## Citation

```bibtex
@misc{saban2026globebench,
  title  = {{GlobeBench}: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations},
  author = {Saban, Nicholas and Bannur, Charvi and Vasileva, Mariya and Deshpande, Darshan},
  year   = {2026},
  note   = {arXiv preprint}
}
```