Bengio团队新作：强化学习重做符号回归，后验采样提速10倍

TL;DR - Bengio’s team introduces ERRLESS, a Bayesian symbolic regression method that uses maximum-entropy reinforcement learning to sample joint posteriors over formulas, constants, and noise. It improves uncertainty modeling while running roughly 10× faster than most learning-based symbolic regression methods.

• Reframes symbolic regression from finding one optimal expression to approximating a full posterior distribution.
• Uses constrained bottom-up expression generation and a GFlowNet trajectory-balance objective.
• Achieves 0.924 AUC on the Feynman benchmark, remaining stable across three noise levels.
• Avoids separate constant optimization, but requires retraining for each dataset and degrades on complex expressions.