Submit a task
Contribute a scientific workflow from a Nature-family paper, including the goal, steps, tools, and expected metrics, for expert review and inclusion as a benchmark task.
AI/CORE Unranked
SciVeri-Bench Task Internal Review 2026
SciVeri-Bench Scientific Task Internal Review 2026
SciVeri-Bench is a department-driven benchmark that evaluates AI agents on their ability to execute real scientific workflows, not just propose hypotheses. It uses tasks derived from Nature-family papers, broken into atomic, verifiable steps across 10 scientific domains, measuring whether agents can drive genuine discovery through experimental loops using real tools and simulations.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Research areas in scope
01
Science Domains
Protein EngineeringGenomicsStructural BiologyComputational ChemistryMaterials ScienceDrug DiscoveryNeuroscienceSystems & Synthetic BiologyClimate & Earth ScienceAstrophysics
Policies worth checking twice
- Tasks are submitted for expert review before inclusion in the benchmark.
- Each task must include atomic, individually verifiable steps with target metrics set by the original scientist.
- The benchmark evaluates agents across four settings: autonomous, workflow-guided, stepwise self-produced, and stepwise human outcome.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026