Scanned emails, uploaded PDFs, and merged attachments often arrive as page streams that must be split into individual documents before downstream classification, extraction, or routing. Zero‑shot large language models (LLMs) can detect document boundaries without task‑specific training, yet traditional Page Classification (PC) and Boundary Decision (BD) formulations resolve only a single boundary per model call, leading to a large number of inference requests.
This work introduces Multi‑Split Boundary Decision (MSBD), which predicts multiple boundaries within a page window in a single call, thereby substantially reducing the number of inference requests. We evaluate MSBD across various language models, document collections, input modalities, and window sizes.
The results reveal a model‑ and corpus‑dependent operating window range: within this range MSBD retains segmentation accuracy comparable to single‑boundary methods while markedly improving inference efficiency; beyond it, accuracy drops sharply. Larger windows expose distinct over‑segmentation and under‑segmentation behaviors across models.
Overall, MSBD offers the strongest accuracy‑efficiency trade‑off, demonstrating that multi‑boundary prediction can make zero‑shot page‑stream segmentation more efficient when the window size is tuned to the target corpus.
Review