* Equal contribution

    Correspondence: [CORRESPONDING AUTHOR], [EMAIL]

    Animated pipeline diagram: a source dataset of gold passage, question, and answer triples flows into context construction, answer generation by four LLMs, response annotation by a judge model, and finally preference-pair construction. Animated pipeline diagram: a source dataset of gold passage, question, and answer triples flows into context construction, answer generation by four LLMs, response annotation by a judge model, and finally preference-pair construction.

    Pipeline diagram: context construction, answer generation by four LLMs, response annotation by a judge model, and preference-pair construction. Pipeline diagram: context construction, answer generation by four LLMs, response annotation by a judge model, and preference-pair construction.

    Radar chart comparing four Ministral-8B training conditions on LegalRewardBench. The length-balanced-augmentation condition (gold) forms the largest polygon on every dimension. Radar chart comparing four Ministral-8B training conditions on LegalRewardBench. The length-balanced-augmentation condition (gold) forms the largest polygon on every dimension.