EBM tutorial thread: best-of-N and RL fine-tuning share one KL-regularized objective

daibond_alpha · x · 2026-11-20

Part 4/5 of the EBM tutorial thread: the same view covers inference-time computation — best-of-N, verifiers, and RL fine-tuning all target one KL-regularized EBM, differing only in whether the cost is paid at training or inference time. Diagrams show amortization and best-of-N sampling fine-tuning, with gradients flowing through the EBM form of the policy.

Original post →

More from Research

Research channel →