Research & results

What we’ve learned.
What comes next.

We’re building autonomous researchers that choose better experiments, develop stronger measurements, and improve their methods. Explore the experiments shaping that work.

Latest / Research methods

October 12, 2026 · Development experiments

Learning which research methods earn their place

Draft-only search improved final quality on two gCastle runs at higher cost. Focused drafting did not help. A coordinated research team completed a separate six-attempt smoke test.

Separate experiments on one exposed task, followed by our plans for objective engineering and recursive improvement.

01 / Safety classification

June 2026 · Two single-seed runs

Making optimization work when labels are scarce

Headline recall: 73.1% for the starting classifier, 70.2% for direct GEPA, and 77.7% for Gnosys.

About 200 verified labels. Test-derived thresholds used a nominal 5% false-positive budget; actual rates differed. Includes regressions and a second early run.

02 / Candidate selection

September 2026 · Retrospective study

Choosing the result that holds up

On a causal-inference task, an additional evaluation changed which candidate Gnosys selected. That candidate had 5.45% lower test error than the validation-only choice.

One search run · Retrospective comparison.

03 / Autonomous ML research

September 2026 · One task, one seed

An early comparison of eight research agents

Gnosys reached 78.45% few-shot test accuracy, compared with AIDE’s 79.21%. Explore the full eight-agent comparison, search behavior, and runtime analysis.

Same model · 25-step search budget. Runtime methodology is detailed in the study.

The research agenda · October 12, 2026

From better experiments
to better researchers.

01

Objective engineering
Study in progress

Measure how a research team’s results change when agents build and revise their own feedback. Compare ten candidate attempts per arm and track the cost of the additional measurements.

02

Recursive improvement
Study design

Test whether an improved researcher becomes better at producing further improvements. Start with causal-inference research, fresh tasks, and controlled handoffs between researcher generations.

03

Learning through discovery
Research thesis

Use historical knowledge and grounded experimental environments to develop research skills. Test whether those skills transfer to unfamiliar problems without answer leakage.

We’d like to use one cookie to count unique visits. Privacy