Abstract page for arXiv paper 2610.10889: SPLIT-RL: Staged Perception-Language Reasoning Training with Claim-Level Advantages