File size: 8,747 Bytes
4881d06 97c654b 4881d06 97c654b 4881d06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 | ---
license: cc-by-4.0
pretty_name: Science Prediction Universe
language:
- en
size_categories:
- n<1K
tags:
- science
- benchmark
- agent-evaluation
- harbor
- chemistry
- materials-science
- biology
- medicine
configs:
- config_name: default
data_files:
- split: test
path: task_cards.parquet
---
<p align="center"><img src="assets/spu-banner.webp" alt="Science Prediction Universe #Benchmark, UniPat AI" width="100%"></p>
# Science Prediction Universe (SPU)
[Paper](https://arxiv.org/abs/2611.15049) · [Homepage](https://spu.unipat.ai/) · [Blog](https://unipat.ai/blog/SciencePredictionUniverse) · [GitHub](https://github.com/UniPat-AI/SPU)
<p align="center"><img src="assets/spu-overview.png" alt="An SPU task: the experimental definition goes to a wet lab and to an LLM agent; the agent's predicted ranking is scored against the hidden measured ranking by Spearman correlation." width="900"></p>
Science Prediction Universe (SPU) is a benchmark of hidden wet-lab outcomes. Each task fixes the candidates, materials, conditions and measurement protocol of a real experiment, and the agent ranks the candidates by the measured outcome. Rankings are scored by Spearman correlation against the measurements.
This repository contains the **100 public tasks**, built from 85 papers published online between January 2025 and August 2026 in 11 Nature Portfolio journals: 39 chemistry, 29 materials science, 20 biology and 12 medicine tasks in 29 subfields. The 16 private tasks of SPU come from unpublished wet-lab experiments and are not released.
**The reference answers are included.** Every task ships with its measurements, grader and oracle solution, so anyone can run and score it. These tasks are therefore not a held-out test set.
<p align="center"><img src="assets/coverage.png" alt="Coverage of SPU: the 116 tasks by discipline and subfield, their source journals and the number of experiments per task." width="820"></p>
## Results in the paper
Thirteen agent configurations, each run once on every task with an eight-hour budget, reach a best mean below 0.5 on a scale where a perfect ranking scores 1 and random rankings average 0.
<p align="center"><img src="assets/results.png" alt="Thirteen agent configurations ranked by mean Spearman score on all scored tasks, with 95% bootstrap intervals." width="820"></p>
Means over the public and private tasks on which each configuration has a scored run, with 95% bootstrap intervals over tasks. Hatched bars are means on the 28 tasks that all thirteen configurations completed, and the right-hand column ranks the configurations on those tasks. See the paper for task-adjusted means, paired tests and cost.
## Contents
| Path | Contents |
| --- | --- |
| `task_cards.parquet` | One row per task: `task_id`, `title`, `summary`, `category`, `subdomain`, `task_path` (the `default` config, `test` split) |
| `tasks/<task_id>/` | A [Harbor](https://harborframework.com/) task: `instruction.md`, `task.toml`, `environment/` (solver image and public inputs), `tests/` (grader, measurements and answers), `solution/` (oracle) |
| `SOURCES.json` | Source paper of each task: DOI, title, authors, journal and URL |
| `SHA256SUMS` | SHA-256 of every file |
## Usage
```python
from datasets import load_dataset
cards = load_dataset("UnipatAI/SPU", split="test")
```
Run a task with Harbor 0.25.0 (Python 3.12 or later) and Docker. This oracle check replays the reference answer and needs no GPU:
```bash
hf download UnipatAI/SPU --repo-type dataset --local-dir SPU && cd SPU
harbor run -p tasks/fluorene-dehydration-concentrations -a oracle -e docker \
--override-gpus 0 --override-cpus 1 --override-memory-mb 1024 -o ../jobs
```
In the paper, each run had 8 CPU cores, 16 GiB of memory, one NVIDIA H20 GPU, web access and an eight-hour budget, as set in each `task.toml`. The grader runs in a separate container without network access. The solver should see only `instruction.md` and `environment/public/`; keep `tests/`, `solution/` and `SOURCES.json` out of its reach.
## Scoring
Each task has one or more comparison groups. The grader computes the Spearman correlation between the submitted ranking and the measured values in each group and combines the groups with the fixed weights given in `instruction.md`. A perfect ranking scores 1 and random rankings average 0. Tied measurements get average ranks, so on 33 tasks the best possible score is slightly below 1. A missing or malformed submission scores −1, the same as an exactly reversed ranking: check `submission_status` in the grader's `report.json`.
The paper's headline metric, the task-adjusted mean, is fitted to the scores of all thirteen evaluated configurations and cannot be computed from a single run. When reporting results on these tasks, give the mean over scored tasks, the number of scored tasks, the model version, reasoning effort and harness. The paper's audited web tools and source-identifier screening are not part of this release, so results obtained without them are not directly comparable to the paper's.
## Sources
Each task is built from one published paper, cited in `SOURCES.json`. Some biology and medicine tasks include published context about human-derived samples or cell lines from the source studies.
## Known issues
- In `mem-cyclopropene-phosphine-conditions-ranking`, one `catalyst_scaffold` value in `protocol.json` joins the image path and a note (`structures/mem_catalyst.pngwithdrawnstereochemistry`). The image is `environment/public/structures/mem_catalyst.png`, and the experiment description links it correctly.
- In `picalm-lipid-class-abundance`, the donor sex, age, ancestry, APOE genotype and cell-bank IDs given in the paper's runs were removed from `cell_context` in `protocol.json` and `experiment.md`. Donor codes, conditions, measurements and grading are unchanged.
All other task files are as they were run in the paper.
## Authors
**Project Team:** Zengji Tu<sup>1*</sup>, Xiang Liu<sup>1*</sup>, Jiarui Sun<sup>1</sup>, Ruoyu Wu<sup>1</sup>, Zhihui Qi<sup>1</sup>, Chengda Wu<sup>1</sup>, Yutao Zhou<sup>1</sup>, Bo Yang<sup>1</sup>, Ziyan Zhang<sup>1</sup>, Jialong Wu<sup>1</sup>, Liang Chen<sup>1†‡</sup>, Kuan Li<sup>1†</sup>
**Advisory Committee:** Steven Dillmann<sup>2</sup>, Yuanqi Du<sup>3</sup>, Soojung Yang<sup>4</sup>, Shizhe Diao<sup>5</sup>, Jianjun Wu<sup>6</sup>, Junlang Liu<sup>7</sup>, Xuming Hu<sup>8</sup>, Lin Chang<sup>6,9†</sup>, Baobao Chang<sup>10</sup>, Xiaowen Chu<sup>8</sup>, Philip S. Yu<sup>11</sup>, Tong-Yi Zhang<sup>8,12</sup>
**Wet-Lab Collaborators:** Hao Cheng<sup>13</sup>, Qin Han<sup>13</sup>, Yanye Feng<sup>14</sup>, Chenru Duan<sup>15</sup>, Chuncheng Liu<sup>15</sup>, Yan Liu<sup>15</sup>, Yuchen Gao<sup>15</sup>, Zhao Li<sup>15</sup>, Ronghua Zhang<sup>15</sup>, Tory Xu<sup>15</sup>, Renjian Song<sup>16</sup>, Yaokai Fu<sup>16</sup>
<sup>*</sup>Core contributors. <sup>†</sup>Corresponding authors: Liang Chen (<a href="mailto:liangchen@unipat.ai">liangchen@unipat.ai</a>), Kuan Li (<a href="mailto:kuanli@unipat.ai">kuanli@unipat.ai</a>), Lin Chang. <sup>‡</sup>Project leader: Liang Chen.
<details>
<summary>Affiliations</summary>
1. UniPat AI
2. Stanford University
3. Cornell University
4. FutureHouse
5. The Hong Kong University of Science and Technology
6. State Key Laboratory of Photonics and Communications, School of Electronics, Peking University
7. Department of Chemistry and Chemical Biology, Harvard University
8. The Hong Kong University of Science and Technology (Guangzhou)
9. Frontiers Science Center for Nano-optoelectronics, Peking University
10. Peking University
11. University of Illinois Chicago
12. Shanghai University
13. Yurogen Biosystems LLC
14. ABclonal Biotechnology Co., Ltd.
15. Deep Principle
16. Bota Biosciences
</details>
## Citation
```bibtex
@misc{spu2026,
title = {Science Prediction Universe},
author = {Tu, Zengji and Liu, Xiang and Sun, Jiarui and Wu, Ruoyu and Qi, Zhihui and Wu, Chengda and Zhou, Yutao and Yang, Bo and Zhang, Ziyan and Wu, Jialong and Chen, Liang and Li, Kuan and Dillmann, Steven and Du, Yuanqi and Yang, Soojung and Diao, Shizhe and Wu, Jianjun and Liu, Junlang and Hu, Xuming and Chang, Lin and Chang, Baobao and Chu, Xiaowen and Yu, Philip S. and Zhang, Tong-Yi and Cheng, Hao and Han, Qin and Feng, Yanye and Duan, Chenru and Liu, Chuncheng and Liu, Yan and Gao, Yuchen and Li, Zhao and Zhang, Ronghua and Xu, Tory and Song, Renjian and Fu, Yaokai},
year = {2026},
eprint = {2611.15049},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2611.15049},
note = {Project page: https://spu.unipat.ai}
}
```
|