single mention · evaluation
LLM-as-a-Verifier
LLM-as-a-Verifier uses an LLM as a verifier to select one candidate trajectory from a pool, without extra training or a separately trained verifier.1 It grades candidates on multiple criteria with a graded scale, and combines pairwise comparisons via a Bradley–Terry preference probability in a tournament.2
Trade-off It requires no additional parameter updates or separately trained verifier.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
The same checkpoint then serves as a verifier, selecting one candidate using LLM-as-a-Verifier (LAV
— NVIDIA Nemotron 3.5 Super and Lightning Technical Reportconverted into a preference probability using the Bradley–Terry model
— NVIDIA Nemotron 3.5 Super and Lightning Technical ReportThis requires neither additional parameter updates nor a separately trained verifier
— NVIDIA Nemotron 3.5 Super and Lightning Technical Report
Also called LLM-as-a-Verifier candidate selection, LLM-as-a-Verifier (LAV).
Filed under evaluation › judge