Abstract: Large Language Models(LLMs) encounter accuracy issues such as semantic distortion and hallucination, as well as comprehensiveness issues including information redundancy and the lack of crossdocument semantic associations in multidocument summarization tasks. Existing solutions mitigate these limitations by designing specialized workflows. However, it remains uncertain whether enhancing LLMs capabilities results in diminishing marginal returns and whether the introduction of workflows can counteract these diminishing effects. This study develops a workflow incorporating automatic document classification and hierarchical summarization to investigate the marginal effects of LLMs on both accuracy and comprehensiveness, and to determine whether workflows can overcome these marginal constraints. Experimental results reveal that, in the absence of workflows, text accuracy(measured by ROUGE score) stabilizes at 11.6, while comprehensiveness approaches 0.025. With the proposed workflow, the improvement in accuracy diminishes from 45% to less than 5%, indicating a pronounced effect of diminishing marginal returns. In contrast, the enhancement in comprehensiveness remains substantial, with the improvement margin decreasing from 50 times to approximately 25 times, yet continuing to grow as model capabilities advance. This suggests that workflows can effectively bolster comprehensiveness and overcome marginal effects. The study confirms the presence of marginal effects in both comprehensiveness and accuracy within multidocument summarization using LLMs and underscores the pivotal role of workflows in overcoming these limitations.

Key words: Large Language Models(LLMs), Multidocument summarization, Marginal effect, Workflow, Digital transformation in education