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

Nikhil Reddy Pottanigari; Sepideh Kharaghani; Saverio Vadacchino; Alejandro Posada; Ying Zhang

cs.AISep 18, 2026

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

Authors: Nikhil Reddy Pottanigari, Sepideh Kharaghani, Saverio Vadacchino, Alejandro Posada, Ying Zhang

Organizations: ServiceNow Canada

Abstract

Scanned mail, uploaded PDFs, and consolidated attachments often arrive as page streams that must be split into individual documents before downstream classification, extraction, or routing. Zero-shot large language models can detect document boundaries without task-specific training, but standard Page Classification (PC) and Boundary Decision (BD) formulations resolve only one boundary per model call. We introduce Multi-Split Boundary Decision (MSBD), which predicts multiple boundaries within a page window in a single call, reducing the number of inference requests. We evaluate MSBD across multiple language models, document collections, input modalities, and window sizes. The results reveal a model- and corpus-dependent operating range in which MSBD preserves strong segmentation accuracy while substantially improving inference efficiency, followed by a sharp decline at larger windows. MSBD provided the strongest overall accuracy--efficiency trade-off, while large windows expose distinct over- and under-segmentation behavior across models. These findings show that multi-boundary prediction can make zero-shot page stream segmentation more efficient when the window size is selected for the target corpus.

Figures & tables

Figure 1: Zero-shot PSS formulations. PC classifies each page independently, BD evaluates one adjacent-page pair per call, and MSBD predicts all $K{-}1$ boundaries within a $K$ -page window in a single call.

Figure 2: Boundary F1 for PC and BD. Adjacent-page context substantially improves TABME++, but provides limited benefit on the form-heavy VRDU dataset.

Dataset	Setting	Cut	Continue	Micro
TABME++	PC	0.637	0.824	0.757
	BD-Concat	0.619	0.830	0.769
	BD-LSTM	0.623	0.926	0.817
VRDU	PC	0.988	0.782	0.878
	BD-Concat	0.992	0.769	0.872
	BD-LSTM	0.992	0.750	0.862

Table 1: Accuracy of trained boundary classifiers using frozen Vision Classifier † features. PC uses one page, BD-Concat concatenates adjacent-page features, and BD-LSTM processes the same pair sequentially. Cut and Continue are per-class accuracies; Micro is overall accuracy. Best Micro per dataset is in bold . † In-house proprietary feature extractor; results are indicative rather than independently reproducible.

Dataset	Feature source	Cut	Continue	Micro
TABME++	Vision Classifier †	0.688	0.851	0.792
	FCOS	0.185	0.981	0.693
	GUSE	0.434	0.587	0.532
VRDU	Vision Classifier †	0.188	0.863	0.550
	FCOS	0.685	0.319	0.489
	GUSE	0.573	0.405	0.483

Table 2: Prototype-based feature-similarity results using Euclidean distance. Each adjacent-page pair is assigned to its nearest cut or continue prototype. Cut and Continue are per-class accuracies; Micro is overall accuracy. Best Micro per dataset is in bold . † In-house proprietary Vision Classifier features.

TABME++						VRDU
Model	Input	Method	B-F1	Lat	Cost	Model	Input	Method	B-F1	Lat	Cost
Claude Haiku 4.5	text	PC	0.674	2.12	10.33	Claude Haiku 4.5	text	PC	0.961	2.25	2.30
	text	BD	0.866	2.42	13.75		text	BD	0.969	2.50	3.26
	text	MSBD ( $K{=}3$ )	0.844	1.19	6.94		text	MSBD ( $K{=}3$ )	0.900	1.06	1.79
	text	MSBD ( $K{=}5$ )	0.817	1.30	5.18		text	MSBD ( $K{=}5$ )	0.795	1.15	1.39
	vision	PC	0.711	2.75	14.22		vision	PC	0.955	2.70	2.91

Table 3: PC, BD, and MSBD results across datasets, models, and input modalities (B-F1 = Boundary F1; Lat = per-call latency in seconds; Cost = full test set, USD). MSBD is shown at

K{=}3

and

K{=}5

for both modalities; the complete sweep (

K{=}2

–

50

) with costs and call counts appears in Tables 4 – 5 . Per dataset and model, the strongest B-F1 and the lowest cost among the rows shown are in bold ; larger windows reduce cost further. † Page-level F1 reported by Heidenreich et al. (2024) ; not directly comparable.

Figure 3: MSBD behavior across window sizes on TABME++ (top) and VRDU (bottom); solid and dashed curves show text and vision inputs. Left : Boundary F1 versus window size $K$ —the useful window (knee) depends on the model and dataset, with GPT-5.4-mini degrading earliest. Center : Boundary F1 versus test-set cost as $K$ increases—Gemini provides the strongest cost–accuracy trade-off. Right : predicted cut rate versus $K$ —Claude and Gemini increasingly over-segment, whereas GPT-5.4-mini under-segments at large $K$ ; horizontal dashed lines mark the ground-truth cut rates.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

	TABME++						VRDU
$K$	Cla-txt	Cla-vis	Gem-txt	Gem-vis	GPT-txt	GPT-vis	Cla-txt	Cla-vis	Gem-txt	Gem-vis	GPT-txt	GPT-vis
2	0.866	0.837	0.883	0.882	0.873	0.876	0.969	0.969	0.971	0.931	0.783	0.962
3	0.844	0.844	0.847	0.887	0.828	0.846	0.900	0.928	0.940	0.938	0.751	0.869
5	0.817	0.799	0.870	0.881	0.793	0.811	0.795	0.755	0.943	0.888	0.609	0.800
7	0.782	0.751	0.863	0.863	0.739	0.765	0.722	0.633	0.925	0.846	0.589	0.727
10	0.726	0.713	0.861	0.800	0.693	0.716	0.616	0.619	0.856	0.872	0.545	0.618

Appendix

Table 4: Boundary F1 across the complete MSBD sweep. The

K{=}2

setting is equivalent to BD. Cla, Gem, and GPT denote Claude Haiku 4.5, Gemini 3.1 Flash-Lite, and GPT-5.4-mini; txt and vis denote text and vision inputs. Results at large

K

on VRDU are based on relatively few windows and should therefore be interpreted cautiously. Costs and call counts for the same configurations appear in Table 5 .

	TABME++							VRDU
$K$	Calls	Cla-txt	Cla-vis	Gem-txt	Gem-vis	GPT-txt	GPT-vis	Calls	Cla-txt	Cla-vis	Gem-txt	Gem-vis	GPT-txt	GPT-vis
2	6236	13.75	21.25	2.79	4.74	7.72	12.39	1102	3.26	4.50	0.66	0.85	1.84	2.22
3	3118	6.94	12.63	1.61	3.13	4.28	7.84	551	1.79	2.73	0.41	0.56	1.11	1.41
5	1559	5.18	9.90	1.22	2.49	3.18	6.14	275	1.39	2.15	0.32	0.44	0.86	1.11
7	1039	4.58	8.97	1.09	2.27	2.81	5.57	183	1.24	1.96	0.29	0.41	0.77	1.00
10	692	4.18	8.35	1.00	2.13	2.57	5.19	122	1.15	1.84	0.27	0.38	0.72	0.94

Appendix

Table 5: Test-set cost in USD and model-call counts across the complete MSBD sweep. The

K{=}2

setting is equivalent to BD. Call counts depend only on the dataset and window size; individual runs may contain one or two fewer completed requests when a model call fails. Boundary F1 for the same configurations appears in Table 4 .

Figure 4: Boundary F1 across datasets, models, input modalities, and window sizes. Darker cells indicate higher F1. Gemini retains accuracy over the widest range of $K$ , while GPT-5.4-mini generally degrades earliest.

Figure 5: Mean latency per model call across the MSBD sweep. Text-input latency remains broadly stable at small and medium windows before increasing at larger $K$ , while vision latency generally rises more quickly as additional pages are included.

Figure 6: Correctly handled boundaries (VRDU). On clear transitions (distinct letterheads, advertisers, or form types), the models agree with the ground truth in both the cut and continue directions. These easy cases sit behind VRDU’s high scores at small $K$ .

Figure 7: Over-segmentation (VRDU): predicted cuts fall inside what is really a single document (ground truth continue ), fragmenting one filing across several predicted documents. Top row: all three models over-cut a continuation page (a contract’s second page; a line-item table continuing across pages). Middle row: only Claude over-cuts while Gemini and GPT-5.4-mini correctly continue. Bottom row: Claude and GPT-5.4-mini over-cut an agreement form’s own terms page (left), and in the rare reverse only Gemini over-cuts a multi-page invoice (right).

Figure 8: Under-segmentation (VRDU): a genuine boundary (ground truth cut ) is predicted as continue , merging two documents into one. Top row: all three models miss the boundary on visually ambiguous transitions (a schedule form giving way to an invoice; an invoice to a new station’s order). Bottom row: only Claude (left) or only Gemini (right) merges the documents while the other two models cut correctly.

Figure 9: Under-segmentation by GPT-5.4-mini (VRDU): in each of these windows GPT-5.4-mini alone predicts continue at a genuine boundary (ground truth cut ), merging adjacent filings, whereas Gemini and Claude both cut correctly.

Document Chunking

Document chunking splits long or structured documents into coherent units for retrieval, question answering, and other language-model tasks.

See all 53 papers

DocSplit: A Comprehensive Benchmark Dataset and Evaluation Approach for Document Packet Recognition and Splitting
Feb 17, 2026Md Mofijul Islam, Md Sirajus Salekin, Nivedha Balakrishnan +6Document Image Understanding Document Parsing
Amazon Web Services
HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding
Feb 13, 2026Wenhui Liao, Hongliang Li, Pengyu Xie +15Efficient VLM Inference Speculative Decoding
South China University of Technology, Guangzhou, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China · Nanjing University, Nanjing, China +1
Khondo: A Multimodal Benchmark for Document Packet Splitting of Bangla Forms
Jul 23, 2026Abu Tyeb Azad, Fahim Ahmed, Ishita Sur Apan +7Document Chunking Multilingual VLM Evaluation
Wichita State University, USA · CCDS, Independent University, Bangladesh · Amazon GenAI, USA

Feb 17, 2026J/K move · Enter open · S save

DocSplit: A Comprehensive Benchmark Dataset and Evaluation Approach for Document Packet Recognition and Splitting

Md Mofijul Islam, Md Sirajus Salekin, Nivedha Balakrishnan, Vincil C. Bishop+5

Amazon Web Services

Document Image Understanding Document Parsing Multimodal Model Evaluation Document Chunking

Open paper AskPDF arXiv

Document understanding in real-world applications often requires processing heterogeneous, multi-page document packets containing multiple documents stitched together. Despite recent advances in visual document understanding, the fundamental task of document packet splitting, which involves separating a document packet into individual units, remains largely unaddressed. We present the first comprehensive benchmark dataset, DocSplit, along with novel evaluation metrics for assessing the document packet splitting capabilities of large language models. DocSplit comprises five datasets of varying complexity, covering diverse document types, layouts, and multimodal settings. We formalize the DocSplit task, which requires models to identify document boundaries, classify document types, and maintain correct page ordering within a document packet. The benchmark addresses real-world challenges, including out-of-order pages, interleaved documents, and documents lacking clear demarcations. We conduct extensive experiments evaluating multimodal LLMs on our datasets, revealing significant performance gaps in current models' ability to handle complex document splitting tasks. The DocSplit benchmark datasets and proposed novel evaluation metrics provide a systematic framework for advancing document understanding capabilities essential for legal, financial, healthcare, and other document-intensive domains. We release the datasets to facilitate future research in document packet processing.

Abstract

Document Chunking

Explore similar work

DocSplit: A Comprehensive Benchmark Dataset and Evaluation Approach for Document Packet Recognition and Splitting

HSD: Training-Free Acceleration for Document Parsing Vision-Language Models with Hierarchical Speculative Decoding

Khondo: A Multimodal Benchmark for Document Packet Splitting of Bangla Forms