Discovering effective alphas from noisy, non-stationary financial data remains a challenging open problem. Recent LLM-based alpha discovery systems improve search automation, but they still largely operate within human-specified research protocols and the evolving object is typically the alpha formulas or mining trajec- tory, rather than the researcher itself. To address these gaps, we introduce STAR, a self-tuning agent framework for studying alpha-research policy evolution. STAR couples an LLM-based alpha researcher with a metacognitive module that can revise the research scaffold, while fixed and isolated validation/test data and a host- side evaluator preserve the integrity of the utility channel, and the search policy prioritizes lineages with stronger descendants. Under a leakage-controlled forward protocol on the most recent 2026Q1 deployment window on CSI300, evolved STAR researchers generate substantially higher quality alphas than the base model and a comprehensive set of strong baselines. Qualitative analysis further indicates that these gains arise from genuine researcher evolution, including a 6.9× expan- sion of operator vocabulary, new tools, and improved research protocols. These results highlight the effectiveness of self-evolving agents in improving capabilities in noisy, weak-supervision domains, and position them as a promising pathway toward building more capable financial researchers.
William F. Shen (Mon,) studied this question.