AQuA closes a recursive self-improvement loop in quant research: IC ~0.190 on crypto factors, held-out Sharpe up to +2.50 on US equities

arXiv·medium signalXBlueskyLinkedInCopy link

AQuA (arXiv 2608.12841, 2026-08-13) tests recursive self-improvement at the level of the research process rather than the model: two separate LM-driven systems — one for symbolic factor discovery, one for trainable model development — that share no agents, memories, candidate spaces or research state, each retaining validated evidence to guide later proposals. Each runs in a sealed sandbox fixing data splits, feature and label definitions and the evaluator, so the system can act only through constrained factor expressions or configuration diffs. The manager-mediated factor pipeline reaches a combined information coefficient of ~0.190 on a crypto universe; the config-driven model loop reaches +0.0843 per-stock IC on US equities and a held-out Sharpe up to +2.50 at two-leg cost.