AWhat changed
Solving a problem can produce a useful procedure or reveal a mistake worth avoiding. EvoLib stores these as modular skills and reflective insights in an evolving library. A new problem retrieves candidate abstractions, the model attempts a solution, and the system updates which abstractions are sampled. Its weighting mechanism considers both immediate usefulness and whether an item helped produce a useful successor. The researchers test this no-weight-update approach on mathematics, coding and interactive tasks. The contribution is the evolving-library algorithm, with its reported comparisons and ablations.
Scientific scope
Solving a problem can produce a useful procedure or reveal a mistake worth avoiding. EvoLib stores these as modular skills and reflective insights in an evolving library. A new problem retrieves candidate abstractions, the model attempts a solution, and the system updates which abstractions are sampled. Its weighting mechanism considers both immediate usefulness and whether an item helped produce a useful successor. The researchers test this no-weight-update approach on mathematics, coding and interactive tasks. The contribution is the evolving-library algorithm, with its reported comparisons and ablations. The library relies on self-evaluation, synthetic tests or majority judgments during learning; these can be wrong. Static experiments revisit a task pool, while continual experiments use a stream, so their results are not interchangeable. Some results tie a strong baseline. Weighted token counts are a specified cost proxy, not measured current spending. Generalization to open-ended scientific work remains unknown. The library’s internal credit heuristic is not a determination of scientific authorship or a proof that its abstractions are valid.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The library relies on self-evaluation, synthetic tests or majority judgments during learning; these can be wrong. Static experiments revisit a task pool, while continual experiments use a stream, so their results are not interchangeable. Some results tie a strong baseline. Weighted token counts are a specified cost proxy, not measured current spending. Generalization to open-ended scientific work remains unknown. The library’s internal credit heuristic is not a determination of scientific authorship or a proof that its abstractions are valid.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.9. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction