LLM_judge question

If I want to use qwen3 embedding to generate and ground truth to calculate the similarity after embedding as a reward, does this need to be modified to the llm_judge format?

Start by locating the existing llm_judge format and the reward paths that handle embeddings and ground truth similarity. Compare their inputs and outputs to determine whether this use case fits the current format or requires a new integration. Done should be a decided scope with the relevant implementation and validation points identified.

Written by the indexing model from the issue text.