Formulaic alpha mining seeks executable expressions that capture predictive market signals within a massive, noisy search space. Existing LLM‑based methods rely on prompting, iterative search, or seed‑factor evolution, leaving direct factor construction via policy learning largely unexplored. To fill this gap we introduce AlphaRLM, a reinforcement‑learning framework that trains large language models to build executable symbolic factors from financial literature without any input factor expressions.\ \ The core of AlphaRLM is the Literature‑Guided Operator Assembly (LEGO) module, which conditions factor construction on ideas extracted from literature and on a set of executable operators, thereby providing semantic priors that steer the search toward meaningful regions.\ \ For policy alignment, AlphaRLM defines two reward families:\
- Executable‑validity reward: guarantees that generated factors can be run without errors;\
- Grouped‑return‑separation reward: captures non‑monotonic structures across factor quantiles, formalized as $$\text{Reward}=\sum_{g}\|\mu_{g}^{+}-\mu_{g}^{-}\|,$@@@MATH_BLOCK1@@@\mu{g}^{+}@@@MATH_BLOCK2@@@\mu{g}^{-}$ are the mean returns of positive and negative groups within quantile $g$.\ \ We conduct extensive experiments on the CSI 300, CSI 1000, and S&P 500 indices. AlphaRLM consistently outperforms strong factor‑mining baselines and prior LLM approaches in annualized excess return. Notably, the 8‑billion‑parameter model retains robust cross‑market performance, highlighting the promise of literature‑conditioned, market‑aligned factor generation.\ \ Review: AlphaRLM bridges textual finance knowledge and reinforcement learning, offering an end‑to‑end pipeline from literature to tradable factors and opening new avenues for automated symbolic factor discovery.