Notes from the laboratory.
Apex is a frontier AI lab. This is the research log — what we learn while training a foundation model for the physical world. Architectures, evaluations, case studies, biosecurity methodology, and the long unglamorous middle where most progress actually happens. Reproduce the work before you bet a roadmap on it.
One backbone. Every modality.
The research agenda underneath every paper on this page is the same single model. CAD, mesh, RTL, molecules, spectra, plate data, code and sensor streams all flow into shared weights — and out through discipline heads.
Four threads, one model.
Architecture, evaluation, safety, platform. The research agenda underneath Aether — the questions we publish on, the questions we are working on, and the ones we have not yet figured out.
Architecture
How Aether is trained, what we put in the corpus, how the latent state spans physics, chemistry, biology and silicon. Cross-domain transfer and scaling-law work.
Evaluation
Benchmarks that mean something — measured against experiment, not against other models. Reproduction kits, error bars, failure cases.
Safety
Biosecurity refusals, capability gating, the architecture of the dual-use boundary. Red-team methodology and external review.
Platform
The agent runtime — long-horizon planning, deterministic memory, eval-gated deploys. What it takes to ship autonomous science in production.
Every paper, by topic.
Six topic threads, one running archive. Each paper carries a tag, a date, an abstract, the reproduction kit and the authors. The full set is below the feature card.
Reproducible by default.
Three principles we hold to on every research post. They are part of the charter — if a published number stops reproducing, we treat it as an incident.
If it doesn't reproduce, we don't publish
Every benchmark on this site ships with a reproduction kit — the case, the mesh, the random seed, the model version. If it stops reproducing, we treat it as a P1 bug.
Failure modes alongside the wins
We publish what doesn't work as carefully as we publish what does. Sweeping the negatives makes everyone's models worse.
Named accountability
Every paper has named authors with named contributions. When numbers move, we say who moved them.
Where else to look.
The research grid is not the whole story. Three more places to read what we are doing and why.
Reproduce before you trust.
If you spot a number that doesn't replicate, we want to know. Reproduction reports go to research@apexworldlabs.com — we acknowledge within a business day.