Lingxue Song

LongVPO: from anchored cues to self-reasoning for long-form video preference optimization