Reaction
The ‘bad mixing’ argument about finite-step MCMC sampling
A commenter argues that finite MCMC steps reweight a base model rather than enforce a hard consistency constraint.
TLDR
A commenter argues that the approach relies on ‘bad mixing’: finite MCMC steps reweight a base model using a function of the expert-prompted base model’s distribution, rather than impose a hard consistency constraint. With infinite mixing, they say, the process would essentially return to sampling from the base model.
Combined views
456
2 Sources, first seen ago
