README.md

---
pretty_name: "Humanity's Sixth Sense (HSS)"
language:
  - en
license: mit
task_categories:
  - visual-question-answering
  - video-text-to-text
tags:
  - multimodal
  - visual-reasoning
  - video
  - benchmark
  - rubric-graded
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.jsonl
---

# Humanity's Sixth Sense (HSS)

[Leaderboard](https://scale.com/leaderboard/hss)

**HSS** is a benchmark for *intuitive visual reasoning*: the implicit temporal,
spatial, social, and abstract structure that people infer from an image or a
short clip at a glance. Each task pairs one image or video with a human-written
question, a reference answer, and a rubric of independently checkable criteria.

People reach **93.1%** on HSS. The strongest of the 24 models we evaluated,
GPT-6-astra at maximum reasoning effort, reaches **53.6%**.

## At a glance

| | |
|---|---|
| Tasks | 522 (288 image, 234 video) |
| Rubric criteria | 723 (1–7 per task) |
| Domains / subdomains | 4 / 11 |
| Split | `test` only |

Every task was written by a trained annotator. It was admitted only after
three independent review rounds approved it unanimously. Of 3,466 authored
tasks, 522 survived (15.1%).

| Domain | Subdomains | Tasks |
|---|---|---|
| Physical & Spatial Logic | Hidden & Invisible · Spatial "Alien Viewpoint" · Affordance & Feasibility · Spatial Reachability | 176 |
| Temporal & Causal Dynamics | Retrodiction · Mechanistic Causality · Extrapolation | 134 |
| Social Understanding | Social Role, Norm & Power Dynamics · Theory of Mind | 110 |
| Abstract & Contextual Inference | Change & Consequence · Patterns & Pareidolia | 102 |

## Fields

| Field | Type | Description |
|---|---|---|
| `task_id` | string | Unique task id |
| `domain` | string | One of the 4 domains |
| `subdomain` | string | One of the 11 subdomains |
| `media_type` | string | `image` or `video` |
| `media_path` | string | Path of the media file inside this repo, e.g. `media/images/<task_id>.png` |
| `prompt` | string | The question shown to the model |
| `golden_response` | string | Human-written reference answer |
| `rubric_criteria` | list of `{id, title}` | Criteria a correct answer must satisfy |
| `num_criteria` | int | `len(rubric_criteria)` |
| `evidence_start`, `evidence_end` | string or null | Video tasks: the part of the clip that holds the evidence for the answer, as `m:ss` timestamps. Null for image tasks |
| `evidence_start_sec`, `evidence_end_sec` | float or null | The same window in seconds |

Models are evaluated on the **whole** video; the evidence window is metadata
for analysis (for example, how much of a clip a model needs to see), not part
of the input.

## Loading

```python
from datasets import load_dataset
from huggingface_hub import snapshot_download

repo = "ScaleAI/HSS"
ds = load_dataset(repo, split="test")
root = snapshot_download(repo, repo_type="dataset")  # media files

task = ds[0]
media_file = f"{root}/{task['media_path']}"
```

## Scoring

An LLM judge reads the question, the reference answer, and the model's answer,
then marks each rubric criterion as met or unmet. Because only essential
criteria survived curation, **a task counts as correct only if every one of its
criteria is met**. The headline metric is accuracy over the 522 tasks. In the
paper the judge is Claude Opus 5.

## Updates

- **2026-10-08**: added the video evidence window (`evidence_start`,
  `evidence_end` and their `_sec` versions) for all 234 video tasks. Revised
  the prompt of task `6a7f82ad5cfeb63114fa869b` (a minor issue found during our
  internal audit).

## Notes and limitations

- **Contamination.** The media comes from the public web and may appear in
  pretraining data. The questions were written for this benchmark and are new.
- **Judge dependence.** Scores go through an LLM judge. The paper reports
  agreement across judges and a human-oracle check.
- This is a test set. Please do not train on it.

## License and media takedown

The annotations in this dataset (prompts, reference answers, rubrics, and
taxonomy labels) are released under the [MIT License](https://opensource.org/licenses/MIT).

The images and videos under `media/` were collected from publicly available
sources. They remain the property of their original creators and rights
holders, and the MIT License does not apply to them. They are included only so
that the benchmark can be used for non-commercial research and evaluation.
Users are responsible for complying with any terms that apply to the original
media.

If you own the rights to any media in this dataset and would like it removed,
please open a discussion on this repository with the `task_id` or file path.
We will remove it promptly.

## Authors

Xingang Guo¹\*, Jing Gu²\*, Brian Jang¹\*, Renxiong Wang¹, Utkarsh Tyagi¹, Daniel Quigley¹, Steven Li¹, David Yan², Daniel Yue Zhang¹, Darvin Yi¹, Forrest Huang², HiJae Kim¹, Tianyi Zhang², Jared Lichtarge², Jihua Huang², Le Xue², Manan Tomar², Qiuyi Richard Zhang², Ruofei Yu², Seth Neel², Yaning Hu², Marcella Valentine², Daniel Evans¹, Chenguang Wang¹·³, Dustin Tran², Tong Zhao¹, Yinfei Yang², Yunzhong He¹

¹Scale AI, ²Elorian, ³University of California, Santa Cruz · \*Equal contribution

## Citation

```bibtex
@misc{guo2026hss,
  title  = {Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models},
  author = {Xingang Guo and Jing Gu and Brian Jang and Renxiong Wang and Utkarsh Tyagi and Daniel Quigley and Steven Li and David Yan and Daniel Yue Zhang and Darvin Yi and Forrest Huang and HiJae Kim and Tianyi Zhang and Jared Lichtarge and Jihua Huang and Le Xue and Manan Tomar and Qiuyi Richard Zhang and Ruofei Yu and Seth Neel and Yaning Hu and Marcella Valentine and Daniel Evans and Chenguang Wang and Dustin Tran and Tong Zhao and Yinfei Yang and Yunzhong He},
  year   = {2026}
}
```