💡 PaLMR aligns multimodal LLM reasoning processes with visual perception via a perception-aligned data layer and hierarchical process reward fusion, reducing reasoning hallucinations and achieving SOTA