Posted on Nov 19 β’ AI-assisted
Synergizing Evidence Retrieval and Utilization for Self-Improving Multimodal Scientific Document Reasoning
Analyzing multimodal scientific documents (combining text, tables, and charts) frequently fails due to retrieval errors or a model's inability to effectively utilize retrieved evidence. This study introduces the RPR (RetrieverβPerceiverβReasoner) agent pipeline, which gathers evidence under a strict context budget, compresses it into a query-specific latent ontology, and autonomously evolves its modules by diagnosing whether failures stem from the retrieval or evidence-reasoning stage.