EverMine: Dissecting the Self-Evolution of Research Capabilities in Long-Horizon Alpha Research
In briefEverMine: Studying Self-Evolving Research Skills in Long-Term Alpha Discovery
SummaryThis paper introduces EverMine, a framework for testing whether self-evolving agents keep improving their accumulated research skills during long-horizon alpha factor discovery. It splits the research state into history, the current factor portfolio, and reusable capabilities, and compares runs with fixed or evolving capabilities under matched resource limits. Across 18 long trajectories and 48 continuation branches, capability evolution showed no consistent gain, while tuning existing factor structures still helped. The authors also find that a candidate's value depends on the evolving portfolio and the order in which candidates are submitted.