<p>Document-level Relation Extraction (DocRE) aims to identify relational connections between entity pairs across multiple sentences within a document. Compared with sentence-level extraction, it presents substantially greater complexity due to the need for comprehensive document-level reasoning and long-range dependency modeling. Current DocRE methods face several fundamental challenges: relation identification often relies on limited subsets of sentences rather than exploiting all relevant evidence; evidence selection and reasoning are typically optimized as independent processes, preventing end-to-end synergy; and entities are frequently sparsely distributed, requiring sophisticated multi-hop reasoning that existing approaches struggle to capture effectively. To address these limitations, we propose ESTR-DocRE, an Evidence-Supervised Transformer Reasoning framework that unifies evidence awareness with cross-entity logical reasoning through two core components. The Evidence-Supervised Representation Module (ESRM) extracts critical evidence sentences through adaptive relevance scoring and integrates them with entity representations via multi-head attention. The Transformer-Driven Reasoning Module (TDRM) constructs joint representations of entity pairs and enables information exchange across related pairs through a streamlined transformer architecture, thereby supporting effective multi-hop reasoning and cross-entity understanding. Extensive experiments on four DocRE benchmarks demonstrate that ESTR-DocRE consistently outperforms strong baselines. Detailed analyses further validate the advantages and effectiveness of our approach.</p>