Splitting Documents at Lower Cost: Multi-Split Boundary Decisions for LLM-Based Page Stream Segmentation

What is Multi-Split Boundary Decision (MSBD)?

MSBD is a prompting formulation for page stream segmentation — the task of splitting a concatenated stream of scanned pages into individual documents. Instead of asking the model about one page or one page pair per call (the PC and BD baselines), it packs all K−1 adjacent-page boundary decisions of a K-page window into a single structured-JSON response. Windows are tiled with a stride of K−1 so every adjacent pair is judged exactly once, cutting the number of model requests roughly by a factor of K−1. At K=2 it is identical to the boundary-decision baseline.

How much cheaper is MSBD than calling the model once per boundary?

Savings are large in request count but modest in dollars. On the TABME++ test set, growing the window from K=2 to K=50 cuts model calls from 6,236 to 127 (49× fewer), yet test-set cost drops only about 3.2× (from $2.79 to $0.87 for Gemini 3.1 Flash-Lite with text input), because each remaining call carries many more pages. Per-call latency also rises at large windows, so the real benefit is fewer requests (better throughput and rate-limit headroom), not proportionally lower cost.

What window size should I use for document splitting with an LLM?

There is no universal answer: the useful window depends on both the model and the document corpus. In the paper's experiments, Gemini 3.1 Flash-Lite with OCR text input was the most tolerant, staying close to the one-pair-per-call baseline up to windows of about 5–10 pages on TABME++ and 5–7 on the form-heavy VRDU corpus, while Claude Haiku 4.5 and GPT-5.4-mini degraded at smaller windows. The authors recommend selecting the window size on a small labelled validation set from your own corpus.

How do LLMs fail when asked to detect many document boundaries at once?

Two distinct failure modes emerge at large window sizes. Claude and Gemini increasingly over-segment, predicting cuts inside single documents; GPT-5.4-mini under-segments, merging adjacent documents by predicting too few cuts. Reliability also differs: Claude starts emitting malformed JSON (detectable through schema validation), whereas GPT-5.4-mini's output stays well formed while its accuracy silently collapses — the more dangerous failure, since only accuracy monitoring catches it.

Is the paper's code and data available?

The extracted paper contains no code or data release statement. The exact prompts for all three formulations are reproduced in Appendix C, both datasets are public (TABME++ and VRDU), but the VRDU stream-construction recipe, the evaluation harness, and the proprietary 'Vision Classifier' features used for the trained baselines are not released, so exact replication is not possible from the paper alone.