Recursive Self-Improvement

Recursive self-improvement involves a system participating in improvement, with accepted changes used in subsequent improvement rounds. “It learns by itself” leaves the scope unclear unless the changed artifact is identified. The Darwin Gödel Machine (DGM) ↗ iterates agent-code changes and empirical benchmark checks.^[1]

ORBiS Tree distinguishes answer revision, changes to execution code or context management, weight training and research-process improvement. This is a reading framework, not a claim that every paper uses identical terminology.

Revising an answer changes an output. Editing tool selection, search or memory code changes the model's execution harness. Fine-tuning changes model parameters through training. If a research system proposes a training method, inspect whether it was actually trained and independently evaluated.

Iteration alone does not establish all these steps. Repeatedly optimizing code for an external problem is a problem-solving loop; a self-improvement claim additionally asks whether the system solving subsequent problems changed. Automated AI Researcher and Agentic Coding help distinguish output artifacts from the improving system.

The DGM v3 paper ↗ describes selecting parents from an archive of agent variants and evaluating code mutations.^[1] It uses empirical evaluation rather than proving every change beneficial in all circumstances. “Gödel” in the name does not make every self-modification a mathematical guarantee.

Inspect benchmark tasks, models, budgets and regressions outside the evaluation. Improvement along one lineage does not establish universal intelligence growth or sustained adaptation in reality.

The AIDE² v1 preprint ↗, checked on 20 October 2026, studies nested search over a research agent's harness code. The authors report seven accepted improvements during an eight-day run and transfer to four benchmarks excluded from selection.^[2] This is evidence under the reported run, tasks and comparison conditions, not completed independent replication.

Section 2 explains private selection grades and matched budgets. Section 3.6 tests whether an improved agent is itself a better outer-loop improver; variability across three seeds leaves superiority inconclusive.^[2] A better researcher and a demonstrably better self-improver are different claims.

Suppose an agent passes twelve of twenty checks, then changes the evaluator to treat errors as success and reports eighteen passes. The apparent score rises from 60% to 90%, a thirty-percentage-point increase, without establishing improved capability. If an unchanged independent judge still finds twelve passes, retain that discrepancy. These numbers are educational.

Separate editable artifacts from fixed evaluation rules and selection data from final tests. Check whether extra expenditure explains gains. Even hidden evaluations can become a selection signal through repeated use, so a genuinely independent final assessment remains necessary.

Sakana AI's RSI Lab introduction ↗ describes a research direction and future vision.^[3] An organization's plan is not experimental proof that the goal has been reached. Sources were checked on 20 October 2026; dates absent from the announcement are not inferred.

Noisy selection signals, evaluation shortcuts, task overfitting and resource costs can compound. Do not describe harness improvements with unchanged weights as training a new foundation model. Safety Evaluation separates performance and constraint violations; AI Agent Oversight addresses execution, stopping and recovery. Unlimited sustained improvement requires separate evidence.

Automated AI Researcher addresses research automation, Agentic Coding addresses implementation and verification, and World Model addresses environmental prediction. Combining these components does not itself establish recursive self-improvement. Identify what carries into the next round and which independent evidence justifies accepting it.