@@ -214,7 +214,8 @@ Although incorporating CoT rationales into generative reward models improves pre
\subsubsection{Rubric-based Reward Models}
For open-ended tasks, output quality is often multi-dimensional and highly dependent on the specific instruction. As a result, directly making a holistic judgment requires the reward model to implicitly determine both \emph{what aspects should be evaluated} and \emph{how well the output satisfies them}. This is challenging. In practice, such holistic evaluation can lead to biased, inconsistent, and less transparent judgments. For example, when evaluating an output to a writing task, the model may need to simultaneously consider factuality, relevance, clarity, and style. A single holistic reward score (or preference label) may overemphasize one aspect, such as fluency, while overlooking others. To provide comprehensive preference judgments, we can employ rubric-based reward modeling. This method is ....
The generative reward models introduced above improve preference modeling by using the reasoning capability of LLMs. However, they still typically produce an overall preference judgment for each response or response pair. For complex open-ended tasks, such holistic evaluation may be insufficient because response quality often depends on multiple instruction-specific aspects. To address this issue, we can employ rubric-based reward modeling. The idea is to introduce a set of task-specific criteria to guide the reward model on \emph{what aspects to evaluate} when predicting preferences.
