Apple Improves Open-Domain QA Alignment with Search-Distilled Rubric-Reward

Apple introduced the Rubric-Reward framework to generate high-quality answers in open-domain question answering (QA). Previously, it was difficult to design effective reward signals because capturing multiple aspects of answer quality in a single scalar objective function was challenging.

The new framework generates per-query Rubrics based on retrieved evidence and decomposes them into multiple quality dimensions to provide fine-grained supervision during the post-training stage. Conditionally applying Rubrics to retrieved evidence strengthens factual grounding, while decomposition into quality-specific dimensions improves coherence, organization, and adherence to query requirements.

Evaluation results showed that this approach achieved 6.5% improvement over the instruction-tuned baseline and 4% improvement over the flat Rubric variant. Consistent improvements were confirmed across three evaluation axes: composition, grounding, and instruction adherence. These results demonstrate that targeted, multi-dimensional Rubrics provide more effective reward supervision for complex open-domain QA.