Support Selection Survives under Official-Context RAG: A Same-Backbone Boundary Study
Abstract
We study bounded Agentic retrieval-augmented generation (RAG) for knowledge-intensive question answering. The task is to answer multi-hop or domain questions from auditable external evidence under a fixed inference budget. Existing Agentic RAG evaluations often vary the model, retriever, corpus interface, controller, training signal, and search budget together. As a result, observed gains do not show whether planning, repair, active retrieval, or support selection is the useful mechanism. We conduct a same-backbone boundary study with fixed Qwen-based retrieval, generation, evaluation, and call budgets. We decompose bounded agentic control into same-stack mechanism controls and seven ablations. The evaluation uses official-context HotpotQA, 2Wiki, and MuSiQue. It also includes a secondary native-retrieval Chinese aerospace QA stress test. Across three seeds, the full Agentic workflow is not the best answer-F1 design. Corrective achieves the highest F1 on the public benchmarks, and Active retrieval achieves the highest F1 on the domain set. In this access regime, Agentic mainly adds lexical grounding and trace structure at higher latency. The most consistent positive component is sentence-level support selection. Removing it reduces F1 on every dataset. Bootstrap CIs support this drop on HotpotQA, MuSiQue, and the domain stress test, but not on 2Wiki.