Position: Prioritize Identifying Structure, Not Complex Models, for Scientific Discovery

Collection of recent ML/statistics papers presenting theoretical and methodological advances across several areas: a position paper arguing mechanistic learning is generically underdetermined in high-dimensional regimes and warning that LLMs can produce fluent but spurious mechanistic narratives; a rigorous analysis showing periodic and soft target updates can stabilize linear Q-learning via switched linear system and joint spectral radius arguments; a state-coupled stochastic-volatility latent-state model with particle EM for recovery under partial observation; ScoreStop, a gradient-based, r�