RLVR this thing is indeed a bit shortsighted in scientific validation; something that takes decades to falsify, it can't wait that long.

ME News message, May 17 (UTC+8), RLVR (Reinforcement Learning and Verification) may show disproportionate flaws in the validation of scientific theories. The validation cycle for scientific theories can last several decades, even several centuries, and the theories currently considered more optimal often end up making worse predictions. This contradiction reveals a fundamental conflict between the reinforcement learning paradigm based on short-term feedback and the long-term, complex nature of scientific exploration, highlighting the structural limitations of existing AI methods in tackling tasks with ultra-long feedback cycles such as scientific discovery. (Source: AiHot)