RSIBench

Recursive self-improvement

Can a model get better at making itself better?

RSIBench gives a model a goal, time, compute, and a working research environment. It can change permitted parts of a model system, then test whether the change helped.

How the measurement works →Recorded work

The three levels

  1. Level 3Research-process coordinationOpen question
  2. Level 2Model-improvement researchResearch controller
  3. Level 1Target-model improvementMeasured consequence

Verification and orchestration applies declared checks, preserves the record, and does not choose the research method.

Current statusLevel 1 and Level 2 execution contracts are under construction. Level 3 remains an open research question.

01

What counts as recursive improvement

A target-model gain can show that one intervention helped. It does not show that the system that chose the intervention became better at research. That question begins only when an updated research controller returns to make later decisions under a matched condition.

One update

P₀ → evidence → P₁

A candidate research controller is produced from a declared update.

Later work

P₁ → new work → P₂

The changed controller must contribute to a further verified improvement.

02

What a cycle record contains

A cycle record keeps the declared starting state, the model's intervention, the candidate artifact, the independent evaluation, protected checks, and complete resource use together. A result is not reduced to a green run label or a single aggregate score.

Declared condition→Intervention and candidate→Independent result

The record also distinguishes a valid regression, no movement, an unresolved comparison, and broken execution. A broken join is repaired; it is not treated as model behavior.

The interface is a separate working surface for following a cycle event by event.

03

What remains visible in the record

  • Declared conditionsObjective, starting state, evaluator, resource envelope, and validity boundary.
  • ConsequencesLocal change, frontier relation, protected capabilities, uncertainty, and complete cost.
  • LineageAccepted, rejected, incomplete, and invalid branches—without turning missing evidence into zero.
  • Traceable evidenceArtifacts and receipts are linked as records, while hidden evaluation material and private reasoning remain private.