AI agents that pick stock factors can now be trained and run entirely in-house, without renting reasoning from an outside API provider.
The system, called AlphaDiverse, targets alpha factor mining: the search for signals that predict stock returns. Multi-agent LLM systems already automate parts of this work, but relying on external APIs limits control over cost, availability, and confidentiality. Worse, long research loops tend to keep circling back to the same handful of proven economic mechanisms rather than finding new ones. AlphaDiverse addresses this by generating multiple complementary research plans and varying the research environment across loops, then using those varied traces to fine-tune local "Planner" and "Realizer" agents, which are further trained with a reinforcement learning method (joint GRPO) that rewards both prediction quality and the diversity of each agent's contribution.
The interesting part is the training discipline, not just the diversity trick. Feedback during training comes only from one data period, and the final frozen model is evaluated on a separate, later period, avoiding the common trap of quietly tuning against the test set. That matters for a field where backtests routinely look better than live results.
Still, the claims here are modest by design. Across four Chinese stock universes, AlphaDiverse produced predictions competitive with existing approaches while exploring a wider range of strategies, not predictions that beat them outright. Quant researchers have chased diversified factor discovery through ensembles and randomized search for years; whether training agents locally actually helps in live trading, rather than in a backtest, is the question this paper does not yet answer.