01
Start with established science.
Use named public measurements, reference compilations and well-accepted reconstructions. Fix the data release, conditions or historical window — not a moving target.
Knowledge · Reconstruction · Nature
A world,
reconstructed.
How much of nature can AI put back together?
The maps, spectra, structures and records of our world — reconstructed from learned knowledge, without searching the internet.
FIG. 01 / THE RECONSTRUCTED WORLDConceptual illustration · not measured data
A BENCHMARK OF SCIENTIFIC KNOWLEDGE
01 / The mission
Nature is the reference.
Knowledge is the question.
The natural sciences have measured and mapped our world in extraordinary detail. Generations of scientists have built a shared record of its patterns, structures and relationships.
Nature’s Last Exam asks how much of that established knowledge a model can reconstruct from what it has learned.
Not just naming a fact. Reconstructing it in detail. We want to understand how recall, learned knowledge and connections across the scientific literature contribute to an AI model’s picture of the natural world.
A test grounded in the famous, public,
established facts of nature.
02 / The approach
From knowledge to reconstruction.
A defined task. A fixed scientific reference. A clear account of what is being evaluated.
02
Ask for the relationships.
Choose scientifically meaningful datasets with a rich literature. Reconstruct fragments, relationships and physical context, beyond a few familiar landmarks, without internet access.
03
Return to the evidence.
Compare the reconstruction with its fixed reference. Keep source provenance and evaluation rules explicit, so that what is being tested stays clear.
03 / The scientific atlas
From the atom
to the observable universe.
Thirteen fields. One natural world.
Explore the scientific scope — from the structures of matter to the history of life.
Selected field01 / 13
Atoms & molecules
Spectra, structures and the relationships that describe matter at the atomic and molecular scale.
Conceptual geometry · illustrative only, not a scientific dataset.
Select a field to explore. Keyboard: Tab to a field; use arrow keys to move, Enter or Space to select.
04 / Scientific judgement
Good science needs
more than a score.
It needs judgement about what is canonical, what is meaningful to reconstruct, and what a fair evaluation should measure.
Expert taste, expert feedback and our own careful task curation are central to this mission.
We curate for scientific substance: established knowledge, meaningful detail and a reference that can be held fixed. Expert-proposed tasks help shape that work, rather than replacing it with a collection of easy-to-check facts.
A clearer picture of
what a model knows.
In development
The benchmark and its evaluation are in development. Public tasks, documentation and results will be shared here as they become ready.