Reinforcement Failing guides the discovery of emergent physical dynamics in adaptive tumor therapy

Artificial intelligence is revolutionizing scientific discovery in medicine, with reinforcement learning (RL) emerging as a promising tool for optimizing therapeutic strategies. Yet applying RL to complex scenarios such as therapy dynamics in solid tumors is constrained by the challenge of constructing training environments that are both computationally efficient and mechanistically interpretable. Here, we introduce Reinforcement Failing, an AI-guided, human-in-the-loop discovery framework that shifts the focus from agent policy optimization to the refinement of the training environment itself. By combining multi-fidelity RL with group-relative performance evaluation across agent cohorts, Reinforcement Failing systematically reveals emergent mechanisms that first-principles models overlook. We apply this framework to adaptive therapy in solid tumors, which seeks to delay resistance-mediated treatment failure. In this setting, Reinforcement Failing uncovered a coupling between the mechanically driven collective motion of cells and spatially-heterogeneous proliferation that strongly influences therapy outcomes. We integrated these emergent physical mechanisms with existing domain knowledge into an augmented training environment, resulting in improved cross-environment therapeutic performance and exposure of potential pitfalls during translation. More broadly, these findings position Reinforcement Failing as a powerful artificial scientific discovery framework, capable of deciphering high-complexity processes at the interface of physics, machine learning, and medicine.