VLM-guided key information extraction for Chinese visually-rich documents

Current document KIE systems are typically optimized for end-task accuracy, but deployment also requires evidence for predicted links, complete processing of long documents, and stable joint SER+RE training. In this work, we study KIE reliability through three requirements: inspectable decision evidence, coverage completeness under long-document inputs, and optimization stability in joint SER+RE learning. We instantiate this perspective with a unified framework consisting of three modules. Global Entity Contextualization (GEC) introduces global entity interaction before pair scoring and exposes entity-level attention patterns as supporting evidence for inspection. Coverage-Preserving Inference (CPI) extends sliding-window chunked decoding to evaluation with deterministic SER merging and cross-chunk relation reasoning, reducing truncation-induced silent failures. Span-Conditioned Relation Initialization (SCRI) initializes relation learning on realistic predicted spans, reducing early-stage gradient conflict in joint optimization. On HUST-CELL (ICDAR 2023 SVRD), the framework improves CompScore from 54.88% to 59.93%, surpassing the reported competition champion (56.45%) by 3.48 points. Beyond benchmark gains, the method provides inspectable relation evidence and robust long-document behavior.