SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding

The researchers launched the SciDocBench process center benchmark, aimed at evaluating the ability to understand scientific documents. The benchmark includes 124 expert questions covering five scientific fields, 19 subtasks, and seven research assistant capability groups, generating a total of 496 evaluation instances. The results showed that the strongest system currently scored only 62.6/100, with significant deficiencies in document perception, evidence tracing, verification, and cross-document reasoning. To address this, the team introduced SciDocIR as a typified evidence graph representation for retaining scientific document objects, layouts, and reference relationships, and created a SciDocDataset containing approximately 15,000 supervised fine-tuning samples and 8,000 reinforcement learning samples. SciDocBench, SciDocIR, and SciDocDataset together form an evaluation and training framework for diagnosing and improving scientific document assistants.