SeerWorld-Bench

SeerWorld-Bench is a 120-case benchmark for evaluating out-of-sight world evolution in generative video models. The model reasons about the resulting world state and formulates its own revisit continuation. The provided blind_revisit_prompt is an optional reference prompt that can be used by default when a model or evaluation pipeline does not generate its own continuation; it requests a return without prescribing a particular reaction. A successful rollout must realize the event, return to the original subject and scene, and render a causally plausible change while maintaining visual consistency.

The benchmark accompanies the paper “SeerWorld: Towards Out-of-Sight World Evolution in Generative Video Models.”

Repository structure

SeerWorld-Bench/
├── README.md
├── LICENSE
├── THIRD_PARTY_NOTICES.md
├── annotations.jsonl
└── data/
    └── test/
        ├── metadata.jsonl
        └── images/
            ├── seerworld_bench_000.png
            └── ...

Benchmark input schema

Each line in data/test/metadata.jsonl contains:

Field	Description
id	Stable public benchmark identifier.
file_name	Relative path to the initial image.
initial_prompt	Establishes the initial subject, scene, and move-away trajectory.
event_prompt	Introduces the event at the appropriate prompt switch.
blind_revisit_prompt	Optional reference prompt for revisiting the original region without prescribing a ground-truth reaction; models may instead reason about and formulate their own continuation.

Example:

{
  "id": "seerworld_bench_000",
  "file_name": "images/seerworld_bench_000.png",
  "initial_prompt": "...",
  "event_prompt": "...",
  "blind_revisit_prompt": "..."
}

Annotation schema

annotations.jsonl contains one row per public benchmark ID:

Field	Description
id	Public benchmark identifier.
event_strength	strong or weak.

A strong event requires an urgent protective, evasive, or defensive response. A weak event permits either a mild or a stronger causally plausible response. The benchmark does not expose or require a predefined reaction type.

License

SeerWorld-Bench is released under the custom SeerWorld Research-Only License v1.0 (research-only-v1.0). Benchmark images, prompts, and annotations may be used only for non-commercial research. See LICENSE for the complete terms and THIRD_PARTY_NOTICES.md for generated-media notices.

Links

• Project page
• Benchmark repository
• Training dataset

Citation

Citation information will be added when the accompanying paper is made public.