AIHill Sampling’s loop: parallel edits to the best verified solution
A user describes generating and evaluating many edits in parallel—512, for example—and making an edit the language model’s new context if it is at least as good.
TLDR
A user describes Hill Sampling as a repeating loop: condition a language model on the single best verified solution found so far, then generate and evaluate many edits in parallel. If an edit is at least as good, make it the new context and repeat.
Combined views
469
1 Source, first seen 11h ago
Hill Sampling’s loop: parallel edits to the best verified solution
A user describes generating and evaluating many edits in parallel—512, for example—and making an edit the language model’s new context if it is at least as good.
TLDR
A user describes Hill Sampling as a repeating loop: condition a language model on the single best verified solution found so far, then generate and evaluate many edits in parallel. If an edit is at least as good, make it the new context and repeat.