RLVR May Perform Disproportionately Poorly in Science
RLVR 可能在科学领域格外糟糕
Dwarkesh argues that RLVR has a short-feedback weakness in scientific theory validation; the post says validation loops can span decades or centuries, and does not disclose experimental results or benchmark numbers.
Why it matters: Dwarkesh Patel argues reinforcement learning from verifiable rewards may fail at scientific discovery, since theories can take decades or centuries to confirm.