Recursive self-improvement for science

Scientific data for self-improving AI.

Help language models use scientific evidence to decide which candidates deserve testing. Start with molecular tasks: check answers, compare choices under changed evidence, and test whether a revision transfers.

Phase 1: Molecular DiscoveryLater phases

P: scientific program
I: improvement process

A revised improver makes the next change.

StartAfter a revisionNext revisionP₀P₁P₂I₀I₁I₂Installed & used
Compare with a fixed improverP₀P₁P₂I₀Same improverP may change
Illustrative mechanism. How the loop works

Teaching demo · MolP2SBench Molecular discovery

Recover a missing molecular fragment

Choose a molecular fragment using dipole or ¹³C NMR evidence.

Source: Xie et al., MolP2SBench (2026)

A teaching walkthrough

Which structure fits the property?

Use the computed dipole to choose between two linkers on the same scaffold.

Computed dipole magnitude
1.45724D

Shading marks the hidden group.

About this example

Recover a hidden fragment from computed molecular properties. Known structures provide the reference answers; controlled comparisons help investigate weaknesses.

Explore dipole and ¹³C NMR here. The downloadable starter and full walkthrough use the dipole example.

Citation & underlying data

This checks your choice against public computed records. No model runs here. The proposed revision and tests are plans; this example alone cannot diagnose a solver.

What makes data useful for RSI?

Dataset requirements
  • Checkable outcomes

    A reference answer or reproducible check.

  • Useful feedback

    Outcomes linked to properties, structures and controlled comparisons.

  • Fresh comparisons

    Separate scientific families for assessment.

  • Reproducible use

    Same checks and budgets; recorded versions.

Provide actual records, references or checks, scientific context and a runnable example. Independent assessment cases can be prepared later.

Research notes

All news & research

Research selected on

CISEPreprint · Related method · Materials search

Making scientific feedback more reliable

CISE uses calibrated reward intervals to guide materials search. It addresses a practical dataset question: how can feedback prevent promising-looking but infeasible candidates from steering later trials?

A feedback method for materials search; recursive self-revision is untested.

Original source
AIDE²Preprint · RSI research · Research agents

Can an improved research agent become the next improver?

AIDE² evolves research-agent code. Its separate ignition test puts an evolved agent back into the role of improving agents—the return path that matters for recursive self-improvement.

Later improver reuse was tested; its advantage remains unresolved.

Original source

A shared scientific collection

Small contributions.
Shared discovery.

A dataset from one project can become a starting point for someone else. Contribute original or curated scientific data as actual files, with clear sources, checkable outcomes and useful feedback. A paper is optional.

Explore the community map
ContributorsChoose a city to meet its dataset contributors.Meet the contributors through the dataset link below.

No shared projects yet

Using a dataset from this collection? Share your project and choose whether to appear on the map.

Share your project
Data useProjects shared by researchers, with permission.