Agentic Retriever–Perceiver–Reasoner Framework for Multimodal Scientific Document Reasoning
The paper introduces an agentic Retriever–Perceiver–Reasoner framework for reasoning over scientific documents where evidence is spread across text, figures and tables. The Retriever selects source-aligned evidence under a context budget, the Perceiver builds a question-conditioned latent ontology over multimodal evidence, and the Reasoner combines that latent guidance with the original sources to produce grounded answers. The three components are trained via diagnostic-guided co-evolution, using a shared answering reward plus Retrieval Potential Feedback and Latent Efficacy Feedback to target resampling. Across four benchmarks and seven evaluation settings, the framework reports average scores of 69.9 and 75.0 with 7B and 8B Reasoner backbones, exceeding the strongest baseline by 4.0 and 9.1 points.