Complex problems often lack sufficient context—e.g., scientific claims—requiring extra knowledge for reasoning. To address this, we introduce the SMART framework, which captures the implicit state of each entity involved in a problem context to acquire the needed knowledge. SMART leverages multiple large language models (LLMs) together with a structured role‑based prompting strategy, curating the state information into diverse, complementary, and solution‑salient premises.\n\nThese premises are verified against real‑world sources before being added to the knowledge base, then used to guide the solving process. Across three scientific‑claim datasets, SMART’s knowledge‑guided reasoning improves validation accuracy by up to 20%. Ablation studies show that grounding premises in real‑world knowledge contributes an additional ~5% gain.\n\nReview: SMART’s explicit modeling of participant states and role‑driven prompting enable efficient knowledge aggregation and verification‑oriented reasoning, highlighting the promise of multi‑agent collaboration for complex scientific inference.