ContributeExpert-reviewed · 10 domains

Submit a task & the workflow you actually used

Scientists propose a discovery task drawn from a real paper, broken into atomic, individually verifiable steps with the target scores they require. After a multi-round expert review it becomes a University of Scientific Workflow benchmark task.

New here? Load the HACS flagship example to see the expected shape.

Task basics

The open-ended discovery goal and how it is framed.

Task title*

Domain*

Sub-domain· science macro · optional

Task coverage· verifiability class

Outcome is known and quantitatively verifiable; the workflow is measurable step by step.

Difficulty

Medium

Estimated duration· optional

Task description· 2–3 sentences · hub-card blurb

Extended description· optional · full procedure context

End goal*· the open-ended objective

Tags· comma-separated

Constraints — software· comma-separated · optional

Constraints — hardware· comma-separated · optional

Datasets· one per line · name — description · optional

Source paper

A real, peer-reviewed paper grounds the task.

Paper title

Year

Venue· e.g. Nature Methods

URL / DOI· optional

Paper PDF· optional · max 24 MB

Authors

Who proposed the task — we credit and contact you.

Author name

Email· for review contact

Department

Affiliation

Job title / position

Allowed tools & simulations

Name the simulations essential to this task. Agents also get a larger pool at eval time — so include only the tools the experiment genuinely requires.

Connecting to SimulationHub…0 selected

Workflow builder

The scientist's procedure — atomic steps, each with verifiable target metrics.

1Step 1

Step title*

Description

Substeps· one per line

Target metrics1

Submission guidelines

What makes a good task

Strong proposals follow our task-construction guideline — the same rubric reviewers apply in the multi-round loop.

Grounded in a real paper
Drawn from published, peer-reviewed work with a concrete outcome.
Atomic, verifiable steps
Each step is independently checkable — not one monolithic instruction.
Explicit target scores
Every step carries the metrics and thresholds you require.
Real tools & simulations
Pick from the SimulationHub catalog — or add any tool that isn't listed yet.

What happens next

01
Propose
Ph.D. student
A Ph.D. student proposes a practical, meaningful task following the task-construction guideline.
02
Evaluate
In-lab advisor
The student's advisor — a domain expert in the same lab — judges whether the direction is promising enough to anchor a Nature-family paper.
03
Revise for agents
Lead student
The lead student lightly revises the proposal — dataset paths, the main workspace path — so an agent can execute it in a computer environment.
04
Multi-round review
OpenReview
The loop runs on OpenReview in participant-restricted mode, mirroring agents4science.stanford.edu.

See the full review process

Contributors may be invited as co-authors

Researchers who contribute high-quality tasks — especially Ph.D. candidates and faculty at recognized institutions — may be invited to join the benchmark as co-authors. Add your affiliation and position above so we can reach you.

Title, domain, end goal & one named step required