Multimodal document understanding
205 evidence items · also tracked in Computer Vision & Sensing
AI that understands documents containing mixed content — tables, diagrams, images, and text — extracting meaning from each. Includes chart reading, diagram interpretation, and table extraction; distinct from standard OCR which processes text rather than mixed visual elements.
Overview
Multimodal document understanding reads documents as people do, taking tables, charts, diagrams and text together rather than flattening them into a stream of characters, and it matters to anyone whose knowledge is locked in visually rich files. It is a leading-edge practice, steady: production tooling is generally available and credible deployments report gains, but the path from purchase to dependable output is not yet clear. Off-the-shelf tools drop structure and invent values, benchmarks cannot be compared with each other, vendors withhold the specifications a buyer needs, and structure-aware processing carries a steep premium over plain text recognition. Until independent analysts recognise the practice as mature and clean outcomes outweigh the friction, it rewards teams able to build their own verification.
Current Landscape
Vision-language extraction is undercutting incumbent per-page pricing. Aliteq's comparison puts AWS Textract at $1.50 per 1,000 pages. Google deprecated its legacy Document AI processors on June 30, 2026. Billing rules also diverge. Synthorai measured Claude adding 1,558 tokens per PDF page and Qwen 3.8 adding 2,098, against about 15 for GPT-6.
Cloud platforms are folding language-model reasoning into their document services. Microsoft's Azure Content Understanding layers LLM reasoning on Document Intelligence, with DataSnipper, FinHero and Wolters Kluwer named as users. Salesforce now processes complex tables with the Docling parser and an LLM. Specialist platforms such as Indico, Convr and LandingAI continue to compete on hard document types.
Capital and analyst attention are following. Cervicorn Consulting sizes the document AI market at USD 15.28bn in 2025 and forecasts USD 169.49bn by 2035, though it does not disclose its method. The same report lists Reducto's $75m Series B and Coupa's acquisition of Rossum. It also warns that vision-language models can produce "plausible but entirely wrong data".
Named deployments report gains on visually dense documents. TC Energy's Pipeline IQ agent cut review time by 85% on 120-page technical drawing packages. Nexturn reports that an unnamed enterprise moved from 60–70% accuracy with OCR to 90–95% with generative AI, reaching 6,400 documents a day and saving $1.2M a year in licensing. M3 runs production extraction on non-standard Japanese form tables, with technical mitigations.
Open-source parsing is maturing quickly. Docling ships regular releases covering chart extraction, table structure and VLM integration, and Red Hat has demonstrated it at scale with Ray Data on OpenShift AI. Datalab reports marker-pdf 2.1.0 scoring 76.0% overall on olmocr-bench and 83.5% on born-digital PDFs. A Baidu-hosted analysis claims 95.69 on OmniDocBench for MinerU2.5. It also credits the tool with lifting an unnamed financial institution from 50K to 230K documents a day, but neither figure is independently verified.
Compact specialist parsers now claim to match far larger models, though every figure is self-reported. OpenBMB's 1.3B MiniCPM-OCR-1.0 reports 93.68 on Real5-OmniDocBench and converts charts to tables. LlamaIndex reports a mean ParseBench score of 0.87 for its 4B OCRia-1, against 0.75 for the next-best system. NVIDIA's open-weight Nemotron Parse 2.0 adds chart-to-table parsing but publishes no benchmark scores. Cohere's Parse 5 scores a 79.2 ParseBench average, against 90.20 for LlamaParse Agentic Plus.
The input route decides whether visual content is read at all. Synthorai found that extracted text answered 0 of 150 chart questions and 0 of 105 questions on a scanned document. The same chart questions scored 130/130 when sent as page images. Routing pays off in cost as well: selecting 4 pages of a 120-page PDF cut input from 253,822 to 8,758 tokens.
Converting visual content into text at index time is emerging as an alternative to serving visual embeddings. Bloomberg researchers report that MIDR reaches 0.6219 average nDCG@10 on five English ViDoRe V3 domains, a 23.0% relative gain over BM25, with about 9× smaller index memory than ColQwen2.5. D-RAC's multimodal conversion produced 95.7% fewer output tokens than agentic chunking. On engineering rulebooks, MCERF's ColPali-based retrieval improves average accuracy by a relative 32.6% over baseline RAG on DesignQA.
Benchmarks show uneven perception across document types. On SciDocBench, Claude-Opus-5 scores 62.6/100, with equations and figures the weakest areas. TeleOCR-VL, a specialised system, ranked first on table extraction at ICDAR 2026 with 41.81. A study of six vision-language models on an industrial machine manual found strong performance on images and schematics but notable degradation on quantitative elements. An Italian pilot study of five end-to-end systems on newspaper front pages found good bibliographic extraction but weak layout and visual analysis.
Reading text in an image is not the same as using it. A preprint finds that Qwen3-VL-8B recognises 84% of gold evidence rendered at 15 px. On those questions it answers 55.7% from the image, against 66.4% from its own transcript. The authors measure this read-use gap at 7 to 14 points across three model families. Scaling raised recognition without narrowing it.
Long documents and cross-document reasoning remain the hardest cases. ExtractBench reports F1 falling from 87.9% on short documents to 27.9% on long ones. The KYNEX harness self-reports 92.65 unified value F1 on ExtractBench, ranking 12 of 50, but scores 79.73 on long documents, and its rules were tuned on the test set. On MMEnterprise-Bench, GPT-5.6 Sol reaches 52.6% with 10 retrieved records. It scores 27.5% on multi-image questions against 83.3% on single textual records.
Conflicting evidence across sources is a further weakness. MMEnterprise-Bench reports cross-source conflict accuracy of 11.8% to 36.3% across ten models. Retrieving more records improved evidence recall without consistently raising answer accuracy.
Gains depend heavily on document condition. In banking, Matil reports that barely legible documents reach 41–49% accuracy with template OCR but 96–99% with vision-language classifiers. Clean typed scans show a narrower gap, 91.3% against 96.8%.
Healthcare adds its own verification burden. Tavrn reports a word error rate of 0.50 for handwriting on English medical forms and finds that 50.1% of EHR text is duplicated from prior notes. Clinical use still requires human review regardless of confidence scores.
Verification, not capability, now blocks broader adoption. NEC Labs names three barriers: grounding, the cost of token-priced context, and the expense of converting whole archives. Hallucination detection for multimodal output has no established method. Confidence scores are insufficient for safe automation. The pharma field still lacks a single independent audit of table, footnote and unit extraction, so human-in-the-loop review remains the default safeguard.
Tier History
Evidence (205)
— Limitation: Qwen3-VL-8B recognises 84% of gold evidence but answers only 55.7% from the image, against 66.4% from its own transcript. The gap is 7–14 points across model families, and scaling does not close it.
— Bloomberg's EMNLP paper: MLLM conversion of tables and charts at index time reaches 0.6219 nDCG@10 on ViDoRe V3 (+23.0% over BM25). The index is about 9× smaller than ColQwen2.5's.
— Academic ColPali-based retrieval and reasoning over engineering rulebooks that mix text, tables and illustrations. It reports a 32.6% relative accuracy gain over baseline RAG on DesignQA.
— Marker 2.1.0, an active listed tool, reports 76.0% on olmocr-bench (83.5% on born-digital PDFs) and falls back to a VLM for low-confidence tables. The benchmark is vendor-run.
— OpenBMB's open-weight 1.3B model parses tables, formulas and charts (chart-to-table). It self-reports 93.68 on Real5-OmniDocBench and throughput 3.62× that of MinerU2.5 Pro.
200 more · latest 2026-10-22 →
— Market report: USD 15.28bn in 2025 to USD 169.49bn by 2035. Cites VLM handling of mixed-content tables, Reducto's $75m Series B and Coupa acquiring Rossum, and warns of plausible but wrong VLM output.
— Negative signal: the best MLLM reaches 52.6% (k=10). It scores 27.5% on multi-image questions against 83.3% on single text records, and 11.8–36.3% on cross-source conflicts in enterprise slides, charts and tables.
— Open-source agentic harness self-reports 92.65 F1 on ExtractBench (rank 12 of 50), falling to 79.73 on long documents. Table tools and page-vision reading drive the gains, but the rules were tuned on the test set.
— Independent pilot of five end-to-end systems: good on bibliographic metadata, weaker on layout, localised extraction and visual analysis. Trap questions test hallucination. Small sample of ten pages.
— Describes MinerU2.5's layout-then-content VLM parsing. Claims 95.69 on OmniDocBench and an unnamed financial institution scaling from 50K to 230K documents a day; all figures are unverified.
— Negative signal: across 15 models, extracted text answered 0 of 150 chart and 0 of 105 scanned-page questions, while page-image routes scored 130/130 on charts. Mixed visual content needs vision input.
— Hands-on run of MarkItDown, Docling and Marker on three synthetic PDFs: none kept heading levels, MarkItDown scrambled the table, and Marker dropped sections 2 to 40 of a 40-page document.
— Peer-reviewed prototype that arbitrates between visual reasoning and OCR/layout tools: gains of 5.8% on DocVQA and 4.9% on SROIE with over 50% fewer tool calls. Benchmark-only, no deployment.
— List-price survey read 3 October 2026: plain OCR is $1.50 per 1,000 pages on all three clouds, while structure-aware processing costs 2.7 to 66 times more. A cost barrier; the author ran no services.
— Microsoft's own retrospective: LLMs beat purpose-built models on unstructured documents but lose on rigid forms; page-image input gave mixed results, and flattened tables produced an invented cell value.
— LlamaIndex's 4B document-parsing VLM claims a mean ParseBench score of 0.87 against 0.75 for the next-best system, leading on tables, chart data and formatting. The results are self-reported and the model is unreleased.
— GA of a multimodal embedding family for visually rich document retrieval; Cohere self-reports 85.8 on ViDoRe V3 for Embed 5 Pro, an 8.8 gain on Embed 4. No named customers or independent check.
— Vendor comparison of local OCR-VLMs: model-card scores come from non-comparable benchmark versions (OmniDocBench v1.5 vs v1.6), with licence traps and no coverage of stamped or annotated documents.
— AWS production deployment: TC Energy's PIPER agent cuts technical document review time 85% with 95–100% accuracy on 120-page engineering packages, supporting 100+ active users in infrastructure operations.
— Research paper on D-RAC showing multimodal-LLM-driven PDF normalization and Markdown conversion produces 95.7% fewer output tokens and 75% faster chunking than agentic approaches, directly optimizing RAG pipeline cost.
— Release notes show ongoing engineering investment in chart extraction, table handling refactoring, and vision-language picture description, evidencing continued open-source tool maturity.
— Research identifies three structural constraints blocking enterprise adoption beyond model capability: grounding (no verified evidence-to-claim mapping), cost (LLM API pricing by context volume), scale (offline archive digitization infeasible).
— Domain analysis quantifies healthcare-specific constraints: handwriting word-error-rate 0.50, 50.1% of EHR text duplicated from prior notes (AHIMA violation risk), 83.8% of AI-scribe errors are omissions; confidence scores insufficient for clinical safety.
— Unnamed large enterprise transition from OCR-based IDP to GenAI extraction improves accuracy from 60–70% to 90–95%, increases throughput to 6,400 documents/day, eliminates $1.2M annual licensing cost.
— M3's production system uses Azure Document Intelligence on non-standard healthcare forms, documenting specific workarounds: multi-byte character correction tables, LLM cross-checks for merged cells, row-span misjudgement resolution.
— RFP guide for IDP procurement: enterprise documents lack uniformity (dozen invoice formats, handwritten notes, scattered clauses); Google Document AI improves from 10 sample docs; confidence thresholds are business decisions, not technical settings.
— Red Hat webinar demonstrating Docling-based multimodal document processing at enterprise scale (tens of thousands of complex PDFs with tables, multi-column layouts, charts) orchestrated via KubeRay with live end-to-end demo.
— Production challenges in multimodal AI: visual invention, audio substitution, temporal drift, cross-modal conflict; confidence scoring insufficient; benchmarks rarely replicate degraded inputs, conflicting modalities, or edge cases.
— Analyst review of experimental multimodal model: hard limit at 384 tokens/image (~0.64 megapixels), no standard doc benchmarks published, independent testing ~97-98% field accuracy; positioned for lightweight chart reading, not production OCR.
— Workflow-centered benchmark on scientific documents (text, equations, figures, tables, code) across 7 capability groups; Claude-Opus-5 achieves 62.6/100 with pronounced gaps in perception, grounding, and cross-document reasoning.
— AWS reference architecture interposing Textract preprocessing before Bedrock RAG; explicitly documents failure modes (incomplete extraction, hallucination, format inconsistency) and provides deployable CloudFormation stack with production recommendations.
— Critical evidence review: no single independently audited benchmark tests table extraction, footnotes, units, normalization together on real pharma docs; patchwork of vendor benchmarks (DocLayNet, TableBench, RD-TableBench) with incomparable metrics.
— Production pattern: document variability (scanned PDFs, handwritten, new layouts) drives real-world exception handling; human-in-the-loop feedback loops essential for models to improve on multimodal edge cases post-deployment.
— Consultancy analysis of production failure modes: fixed-resolution degradation, shortcut learning, multilingual brittleness, captioning latency; gap between benchmark and reliability is architectural, not a tuning problem.
— Independent coverage of Cohere Parse 5 GA (Aug 27): proprietary 2.3B VLM on ParseBench averages 79.2 (table extraction 92.6); leaderboard context shows LlamaParse Agentic Plus leads at 90.20, indicating competitive benchmark-driven market.
— Banking IDP guide with real field accuracy: barely legible documents 41-49% (template OCR) vs 96-99% (vision-language), clean scans 91.3% vs pristine 96.8%; five-stage multimodal-aware pipeline emphasizing layout-agnostic extraction.
— Peer-reviewed ICDAR 2026 competition paper: specialized OCR-VL systems (41.81 table extraction, first place) outperform general VLMs on scientific charts, chart summary generation remains harder than structure recovery.
— Comprehensive research benchmark (3.7k diagrams, 18.3k questions, 6 scientific domains, 12 MLLMs) showing models excel at reasoning (86% DQA accuracy) but struggle at diagram-to-code generation; agentic evaluation reveals tool use improves parsing but degrades reasoning, establishing production architectural trade-offs in scientific document understanding.
— NVIDIA Enterprise RAG Ingestion Scaling Guide with quantitative performance benchmarks: GPU acceleration reduces multimodal PDF ingestion from days to hours; pre-processing extraction saves 93% end-to-end time; multimodal ingestion 10x slower than text-only; chunk-size trade-offs documented for accuracy vs context and throughput.
— Peer-reviewed agentic document understanding research: Q-Guide agent using deliberate perception (targeted OCR, zoom, grounding) outperforms direct prompting (65% vs 40% on DocVQA2026) and multi-agent alternatives; improvement holds across Claude 4.6 Opus, Sonnet, Opus 4.5; accuracy scales with perception budget, most gains within 2-3 deliberate rounds.
— Azure Content Understanding GA (August 2026) technical tutorial demonstrating multimodal service combining OCR, schema-defined extraction with LLM models, and schema-proposal with Qwen2.5-VL-7B; production-ready configuration-driven approach enables tuning without model retraining.
— Databricks Document Intelligence Precision Mode GA with 7-point benchmark gains; named production customers ICE (millions of financial documents monthly), Panasonic, EY-Parthenon; agentic architecture decomposes large extraction jobs, spawns parallel subagents, handles long documents and complex schemas requiring reasoning.
— Industry maturity assessment: document AI has 'crossed from pilot project to production tool'; extraction from native documents achieves mid-to-high 90s accuracy with sensible verification; clear boundary holds (extraction works, interpretation does not) across all model generations; democratizes capability to four-to-twenty-person firms.
— SlideAgent hierarchical multimodal framework (Georgia Tech + JPMorgan) breaks multi-page documents into three levels (full, page, element), achieving 7.9% improvement over proprietary base model on financial presentations and 9.8% on open-source models; especially strong on cross-slide comparison and visual relationship understanding.
— PerceptionBench news coverage: 16 current multimodal models all score below 60% on isolated visual perception; top model (GPT-5.6-Sol) 59.7%, but only 26.9% on hallucination (correctly answering 'object absent'); documents practical risk for document/invoice/screenshot accuracy and error compounding before reasoning.
— Critical assessment of Mistral OCR 4.1 release identifying vendor transparency gaps: no published accuracy, throughput, latency, or language-coverage specifications; confidence scoring calibration undocumented, creating production risk for auto-upgraded deployments.
— Peer-reviewed comprehensive evaluation framework (7,093 samples, 32 languages, 17 multimodal LLMs across 5 core capabilities) showing models with comparable overall accuracy exhibit fundamentally different failure patterns under real-world document acquisition conditions.
— Independent third-party benchmark on production documents (8 cases, 463 fields, 3 runs) revealing multimodal VLM layout weakness: space-ocr 91.1% accuracy vs Mistral OCR 70.0% vs Mistral VLM 73.5% on tables with merged cells and hierarchical structure.
— Vendor SOTA multimodal embedding model (3B–8B parameters) achieving 63.42 NDCG@10 on ViDoRe V3 benchmark for visual document retrieval; explicitly designed for enterprise documents combining text, tables, charts, figures.
— Production deployment metrics from named customers: HomeLight automated 90% of document workflows, Checkr processes millions at 95–100% accuracy with 60–100% human-review reduction. RealDoc-Bench benchmark 0.847 Adjusted F1 on 1,500 real-world documents.
— Large-scale independent benchmark (370 documents, 4,869 pages, 14 systems, 67 document types) revealing critical multimodal model limitations on long documents (87.9%→27.9% F1) and weak evidence grounding (<50% F1 word-level).
— NVIDIA's open-weight parser, under 1B parameters, adds chart-aware parsing, chart-to-table conversion and improved table structure recovery. It publishes no benchmark scores.
— Independent IFAC study of six VLMs on an industrial manual that mixes diagrams, graphs and tables. The models are strong on images and schematics but degrade notably on quantitative elements.
— Peer-reviewed journal synthesis of multimodal document parsing strategies (perception-cognition-action lens) for scientific documents; validates reliability frameworks and layered defenses for production deployment.
— Production-ready technical tutorial demonstrating distributed document parsing with Docling and Ray Data on enterprise Kubernetes; handles thousands of complex PDFs with streaming execution for wall-clock performance.
— Peer-reviewed ACL 2026 Industry track showing high OCR accuracy does not ensure downstream RAG success; structural/semantic errors from layout misunderstanding cause retrieval failures despite low character-error rates.
— Salesforce CDP Summer 2026 release integrating Docling for complex table extraction; major enterprise platform adoption signal demonstrating production quality and cost-effectiveness at scale.
— Technical research validating visual encoding quality as primary differentiator; controlled experiments show 13.2-point benchmark gap from encoder alone, with pixel-level reconstruction improving low-resolution performance 16.7 points.
— ACL 2026 benchmark evaluating 21 leading MLLMs on real financial OCR tasks; GPT-4o achieves only 46.01% overall, documenting sharp performance degradation in multilingual settings—critical evidence of frontier model limitations.
— Alibaba releases 0.8B-parameter end-to-end multimodal model achieving 96.58 on OmniDocBench v1.6, first to exceed traditional pipeline methods; single-pass Markdown generation with text, formulas, tables, visual regions.
— Market analysis showing frontier labs (Anthropic, OpenAI, Google) stopped publishing quantified DocVQA/ChartQA scores after March 2026, indicating benchmark saturation and potential vendor opacity on model performance.
— Industry synthesis of Q2 2026 trends: weekly frontier model releases, OCR commoditization (Mistral $7/10K pages vs AWS $15), agentic routing architectures, and production readiness gaps despite capability acceleration.
— SIGIR 2026 workshop paper documenting six-category taxonomy of silent failures in multimodal models (modality shortcuts, phantom grounding, provenance hallucination); shows surface accuracy overestimates trajectory-level correctness.
— Docling integrated as Model Context Protocol server with 63k+ GitHub stars; demonstrates adoption as standard interface for agentic AI document processing, enabling first-class capability in AI agent tool interoperability.
— ACL 2026 paper presenting hierarchical multimodal retrieval framework addressing routing failures and evidence fragmentation in large document sets; 12.9% retrieval recall improvement on industrial ODQA benchmarks.
— Major LLM vendor (Mistral) releasing production-ready multimodal document AI model via Microsoft Foundry with OCR, vision reasoning, and multilingual support—strong signal of ecosystem maturity.
— Peer-reviewed benchmark directly assessing frontier multimodal models on realistic professional PDF tasks, revealing significant capability gaps and documenting failure modes.
— Artificio's market analysis explicitly covers multimodal document understanding as a mainstream 2026 trend, cites Gartner on agentic adoption (67% of initiatives), provides specific use cases for complex document handling, and honestly assesses implementation barriers (40% underperform ROI due to integration, not accuracy).
— ACL 2026 paper integrating visual cues into KG-based RAG, addressing multimodal reasoning over long-form domain-specific content. Consistent improvements on textual and multimodal benchmarks.
— Technical deep-dive into Docling-Rust (docling.rs), the production-ready Rust port addressing Python deployment bottlenecks, with performance claims and multi-threaded GUI implementation.
— Peer-reviewed ACL 2026 Findings survey comprehensively reviewing MLLM-based document understanding methods, training paradigms, challenges, and roadmap.
— Production case study: named org (Capestart) migrated clinical PDF extraction from OCR+LLM to direct multimodal PDF understanding (Claude Sonnet Base64 input). Achieved 95.6% accuracy with visual grounding across tables, charts, and layout-complex documents. Demonstrates mature deployment and architecture patterns.
— Major data platform (Databricks) launching managed document parsing as SQL function, extracting structured content with layout metadata, confidence scores, and bounding boxes — signals ecosystem broadening beyond specialist vendors.
— Production regression: Mistral Document AI 2512 suddenly failing with 422 errors for previously-working parameters, multiple users affected, indicates adoption and real deployment friction.
— Production deployment of contextual RAG technique on complex documents (PDFs, insurance, real estate) with measurable retrieval improvements.
— Practitioner analysis of Q2 2026 OCR/parsing landscape showing benchmark saturation, multi-page parsing advances, and shift toward agentic document processing. Names 10+ specific models with performance metrics.
— Production architecture guide with measured deployment outcomes showing 15–35% accuracy improvement for multimodal RAG in enterprise knowledge bases.
— CHI 2026 peer-reviewed research addressing core multimodal challenge: chart data extraction using MLLMs. Proposes human-centered progressive learning framework with 7B model achieving state-of-the-art accuracy on diverse chart types without visible labels.
— Market inflection analysis documenting cost shift from traditional IDP to vision-LLM approaches: Gemini Flash extraction at $0.17/1K pages vs Textract $1.50/1K, signaling vendor consolidation and platform transition toward multimodal foundation models.
— Comprehensive guide to agentic IDP architectures: agentic systems achieve 90%+ automation rates vs 60–70% traditional; 98% classification accuracy and 77% cost reduction reported in healthcare deployments; 8–40 seconds per-page latency documented.
— Google Cloud announced document understanding support in Gemini Enterprise Agent Platform: 330 customers each processing over one trillion tokens in past year; supports up to 3,000 files/prompts with variable sequence length tokenization for improved latency.
— KDD 2026-accepted paper identifying critical bottleneck: retrieval over heterogeneous multimodal knowledge is difficult, cross-modal alignment challenging, existing retrievers poorly suited to multimodal corpora—signals architectural immaturity in production RAG deployment.
— Technical comparison of production multimodal document services: Mistral OCR 4 tops OlmOCRBench leaderboard, supports 170 languages, available for self-hosting; Google Document AI excels on specialized processors and compliance breadth, GCP-native only.
— Peer-reviewed multimodal benchmark (7,262 items, 27 frontier LVLMs evaluated): identifies Visual Logic as systematic weakness across current models, providing negative signal on capability maturity and need for improved evaluation methodology.
— Practitioner analysis documenting critical gap between extraction metrics (99%+ cell accuracy) and real-world financial correctness: column slips, locale errors, header merging remain invisible to TEDS/grid similarity scoring, revealing production maturity constraint.
— AWS Bedrock product GA for Claude multimodal vision capabilities: extracts insights from documents, processes diagrams, reads charts. Named enterprise use cases in legal (parse documents and answer questions), insurance (claims/policy analysis), operations (extract from emails/business documents). 1M token context window.
— Production-document benchmark across logistics, healthcare, finance, real estate (1,500 layout samples, 1,359 Q&A prompts): demonstrates agentic parsers (Extend, LlamaParse) achieve 91-96% Q&A accuracy vs traditional OCR platforms 70.5% (AWS Textract), revealing infrastructure preference for layout-preserving agentic approaches.
— Multimodal document understanding deployment on legal contracts (CUAD, 510 contracts, 249K instances): reveals 51–56% aggregate hallucination masking typed gaps (numeric/obligation claims 65–74%, temporal 29–35%); multi-agent debate reduces fabrications 45%, achieves parity with commercial APIs at 4B parameters.
— RSGI independent study (87 respondents, 60 firms, April-June 2026): 68% deploying multimodal AI agents for legal documents; 21% running 50+ agents; 11 hours/week time savings; 44% revenue increase, 53% profitability gain among firms tracking outcomes—signals production-scale adoption in regulated vertical.
— IBM general availability of Docling as managed enterprise service (40M total downloads, 500k daily), with hardening work for production resilience and independent user validation: Singapore financial institution reported improved accuracy and 2× parsing speed over open-source baseline.
— Peer-reviewed hallucination mitigation framework for MLLMs using external visual evidence retrieval and reliability scoring. Empirical results: improves accepted prediction accuracy 85.84%→88.88% (89% coverage) and reduces hallucination rate 14.16%→11.12% on ImageNet-100.
— CVPR 2026 fine-grained hallucination detection system with VisionHall dataset (6.9k manually annotated MLLM outputs, 20k synthetic samples) classifying errors across six categories at phrase-level. Outperforms GPT-4o and Llama-3.2 on detection/editing, enabling actionable error correction in production document pipelines.
— Production study of 480M verified AI outputs across legal, financial, healthcare: multi-model verification reduced factual errors from 8.3% to 3.2% (61% reduction), demonstrating concrete reliability improvement for document-centric workflows in regulated sectors.
— Benchmark revealing persistent gap: layout detection models struggle to generalize to operational institutional documents despite academic benchmark strength—documents real-world generalization barrier directly relevant to production multimodal document understanding maturity.
— Microsoft Build 2026 announcement of Azure Content Understanding combining Document Intelligence with LLM reasoning. Named production deployments: DataSnipper (Excel integration), FinHero (LLM-powered evolution), Wolters Kluwer (tax workflows)—demonstrates enterprise adoption and ecosystem maturity.
— Production contract intelligence system achieving 99% accuracy (vs 55% rules-based predecessor) using layout-aware smart chunking and dual semantic+structural analysis; demonstrates dramatic ROI improvement from OCR+rules to multimodal AI architecture.
— Peer-reviewed comparative study of LayoutLMv3, Donut, and Qwen3 on RVL-CDIP benchmark: specialized multimodal Transformers outperform LLM-based approaches on layout-intensive documents; image dominates text signal—provides architectural guidance for document type classification.
— ICML 2026 peer-reviewed research identifying hallucination cascade failure mode in multimodal models during interactive use; models progressively neglect visual grounding across conversational turns—identifies critical limitation in document-based interactive scenarios.
— Inference-time framework for fact-level hallucination repair in multimodal document processing with convergence guarantees; tested across image-to-text, image+text-to-text, audio-to-text—demonstrates practical progress toward reliable production systems.
— Enterprise strategy synthesis by third-party consulting firm citing Gartner: 80% of enterprise software will be multimodal by 2030; McKinsey: highest-ROI AI involves ≥2 input modes; insurers using multimodal for first-notice-of-loss reduced claim cycle 35% and fraud accuracy improved—signals mainstream enterprise readiness.
— Production deployment guide comparing three open-source document intelligence frameworks with detailed GPU requirements and throughput benchmarks; demonstrates mature open-source alternatives to commercial platforms—signals ecosystem breadth and self-hosting viability.
— Independent benchmark of 10 providers (traditional OCR + VLMs) on 1,000 real documents: VLMs matched/exceeded traditional OCR especially on charts, handwriting, complex fields; methodology and datasets released for reproducibility—provides empirical VLM viability signal.
— IBM open-source vision-language model specialized for document extraction (charts, tables, KVPs) with 94.2% VAREX zero-shot accuracy; 4B parameters; integrates with Docling pipelines—demonstrates lightweight frontier-level performance for production extraction stacks.
— Peer-reviewed innovation improving multimodal chart extraction via VLM self-ensembling with 23% relative improvement; introduces WB-ChartExtract benchmark (7× more complex than existing); demonstrates practical advancement in extraction reliability.
— Practitioner guidance articulating production hybrid patterns: vision LLMs win on semantic tasks, layout-dependent work, charts/figures; traditional OCR wins on cost-per-page at scale; hybrid OCR+vision LLM pipelines now standard in 2026 enterprise deployments.
— Peer-reviewed benchmark with 10,000 diverse receipts evaluating MLLMs on hierarchical extraction tasks. Reveals critical 'Analyst-Calculator Dichotomy': GPT-5.4 excels at semantic reasoning but achieves only 37.82% on numerical tasks; visual grounding bottleneck documented across multiple items.
— Domain-specific benchmark of 11 MLLMs on 2,878 financial PDFs; no model exceeds 65% accuracy vs 79.41% human expert. Documents 'Analyst-Calculator Dichotomy' and visual grounding bottleneck in cross-page multi-chart synthesis—signals performance gaps in expert-domain document understanding.
— Peer-reviewed research proposing Evidence-Carrying Multimodal Agents (ECA) to prevent hallucination-to-action conversion in document/screenshot processing. Achieves 0% unsafe-action rate (Wilson 95% CI: 2.67%) vs 100% for naive agents; directly addresses production maturity gap.
— KPMG AI Adoption Index shows Claude climbed from seventh to second place (Q1 2025→Q1 2026); vendors in legal tech, financial services, healthcare cite superior document handling on messy real-world inputs and lower hallucination rates vs GPT-4 as adoption driver.
— Peer-reviewed benchmark revealing 'Attribution Hallucination': models produce correct answers while citing incorrect document regions. Gemini-3.1-Pro achieves only 76% Strict Attributed Accuracy despite higher answer correctness; identifies critical reliability gap in production systems.
— Industrial benchmark of 5,990 Alibaba real-world documents across 12 commercial domains reveals critical failure: text-only configurations outperform multimodal configurations (Qwen2.5-VL-7B, Qwen3-VL), indicating models fail to leverage visual modality—contradicts multimodal value proposition.
— Named enterprise deployment: Goldman Sachs developed autonomous agents with Anthropic for client vetting and KYC workflows, extracting entities, determining document sufficiency, and making judgment calls within reasoning boundaries. Production evidence at Fortune 500 scale.
— Framework for high-fidelity OCR datasets and multilingual document understanding spanning 82 languages. DPO-based adaptation improves in-domain +1.9% and out-of-domain +1.8% without language degradation; directly addresses low-resource language gap in global enterprise deployment.
— Google announced general availability of multimodal RAG in Gemini API (May 5, 2026) using Gemini Embedding 2 for unified vector space across text, images, audio, video, and PDFs. Eliminates preprocessing and signals ecosystem maturity via major vendor product release.
— Comprehensive technical guide on production multimodal RAG architecture (OCR, layout analysis, LLM reasoning). References Gartner forecast ($2.09B IDP market by 2026, 13% CAGR). Covers real-world deployments in financial services, insurance, healthcare, and legal with architectural patterns.
— Major open-source document conversion toolkit (55.8k GitHub stars) using vision-language model (Granite-Docling-258M) to convert PDFs, DOCX, images into structured data, detecting tables, formulas, reading order, OCR—core multimodal document understanding infrastructure.
— Amazon Science benchmark for evaluating vision LLMs on long-context document understanding; tests ability to maintain grounded understanding across large document contexts beyond short OCR snippets; signals maturity of evaluation infrastructure.
— Conference presence at Red Hat Summit (May 11–13, 2026) with four sessions covering production multimodal document understanding use cases (aviation maintenance manuals, banking/insurance claims at scale, RAG on Kubernetes, Ray Data pipelines).
— Amazon Science research paper proposing schema optimization to address extraction unreliability. Problem: treating JSON schemas as static contracts leads to suboptimal extraction, hallucinations, and unreliable agent behavior. Solution: dynamic schema optimization for improved extraction fidelity.
— Research-backed analysis introducing InduOCRBench, a benchmark specifically designed to evaluate OCR robustness in RAG systems. Demonstrates OCR paradox: high character-level accuracy (CER/WER) doesn't guarantee effective document understanding, showing semantic/structural understanding gap that multimodal approaches address.
— Production case study: Sun Finance deployed AWS Textract + Amazon Bedrock + Rekognition for identity verification. Measured outcomes: accuracy 79.7%→90.8%, cost reduction 91%, processing time 20 hours→<5 seconds. Went live January 22, 2026. Demonstrates hybrid OCR+LLM+vector-search pattern at production scale.
— Production research on industrial KYC document extraction using VLMs; multistage pipeline with page-level retrieval achieves 87.27% accuracy on 120 production documents across 3000+ pages; demonstrates page retrieval as dominant factor in multimodal extraction.
— Production deployment case study: global investment firm reduced portfolio analysis from one week to hours using Claude multimodal agents to extract images from 100+ page PDFs and interpret visual content (maps, topography).
— Frontier-model benchmark for extraction and grounded reasoning on complex documents; Qwen3.6-35B leads at 89.9%; active use in model comparison tables signals evaluation maturity and sustained competitive differentiation.
— Industry positioning covering Abbyy's VLM-focused strategy and DocLang working group (IBM, Red Hat under Linux Foundation) signals ecosystem consolidation toward multimodal and agentic AI architectures with standardization effort underway.
— Real-world comparative test of GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro on 120 financial documents: headline accuracy 87-89% masks field-level error patterns; confidence calibration failures; production architecture requires model-specific tuning per use case.
— Experienced practitioner (10+ years ML on document systems) recommends hybrid OCR+LLM architecture with separation of concerns; highlights LLM limitations (no character-level confidence, numerical precision errors, layout sensitivity); emphasizes human-in-the-loop necessity.
— Practitioner analysis revealing benchmark-to-production gap: 97% accuracy masks 30% line-item corruption in real invoices; OCR accuracy degrades 60-80% on academic papers and handwritten documents; structure recovery independent from text accuracy. Critical negative signal on maturity.
— Swiss consulting firm case study: logistics deployment achieving 75% error reduction and 12× cycle time improvement (48 hours→4 hours) using multimodal AI agent for automated delivery note extraction and validation—concrete ROI signal.
— Production architecture decision-making: Azure Document Intelligence ~10× faster and cheaper per call than GPT-5.4 vision; hybrid strategy uses DI for primary detection, reserves vision-LLM calls for edge cases—demonstrates real deployment trade-offs in evaluation.
— Thoughtworks elevates Docling to Trial tier (April 2026), signaling mainstream practitioner adoption of open-source multimodal document processing with production-scale assessment: 'performs well on large files containing text, tables, images with strong quality-to-cost balance for agentic RAG.'
— Independent benchmarking (April 10, 2026) ranking multimodal models on document understanding (OfficeQA Pro: 45% weight) and academic reasoning. Shows rapid model evolution and documents enterprise capability differentiation among frontier and open-weight models.
— Discussion of production challenges in multimodal systems including document processing failure modes and integration patterns.
— Community-developed visual inspection tool for Docling document processing pipelines (April 5, 2026) indicates ecosystem maturity; enables quality validation and debugging before feeding extracted chunks into RAG systems.
— Benchmarking report shows SOTA multimodal models score below 50/100 on OCRBench v2 for layout perception and complex element parsing. Manual exception-handling costs (~$4.83/page) dominate API costs, indicating adoption bottleneck.
— Comparison of 5 competing document parsing platforms with named healthcare RCM deployment: 120K prior authorization pages daily scaled to 240K with 60%→90%+ accuracy using ADE with zero-data-retention HIPAA compliance.
— Multi-industry deployments across financial services, media, healthcare document AI: fund analytics pipeline 98% accuracy, Hollywood VFX script breakdown 90% time reduction, music metadata validation 48h→30m. Documents the 95% demo/60% reality gap and architectural building blocks that generalize across domains.
— Microsoft Azure AI Search official documentation (GA, March 30/updated April 6, 2026) for multimodal document understanding: native ingestion of text and images, shared vector embeddings, handling embedded diagrams in PDFs for RAG applications.
— AIIM/Deep Analysis survey (600 enterprises): 61% still paper-dependent despite 78% operational AI, 66% IDP tool replacement rate. Documents vendor demo accuracy (98% clean PDFs) vs. real-world (70% on client documents with scans, handwriting, layout variation).
— EACL 2026 peer-reviewed study on real business documents: image-only MLLMs match OCR+traditional pipelines, validating multimodal-only architecture for enterprise use and reducing infrastructure complexity.
— Industry metrics across financial services, insurance, lending: mortgage processing ($8.5-28K monthly savings at 200 apps/month), 6-18 month payback, 96-99% accuracy threshold required for financial document ROI at scale.
— Benchmark documenting 'severe hallucination issues in SOTA models regarding detailed visual perception,' particularly in fine-grained grounding tasks; directly relevant to document field extraction and table reading quality.
— Agentic extraction API achieving 99.16% accuracy (5,286/5,331) on DocVQA with parse-once-query-unlimited and visual grounding, enabling unlimited downstream queries without reprocessing; 18 of 45 errors are true parsing failures.
— CVPR 2026 benchmark (15,234 pages, 11 languages) documenting 'pronounced performance degradation on low-resource Southeast Asian languages,' signaling critical adoption barrier in multilingual contexts.
— 1,777-document benchmark (20 models): below 4B parameters, schema compliance is dominant bottleneck (depressing accuracy 45-65pp); fine-tuning yields +81pp gains; layout-preserving text outperforms pixel-level visual cues.
— Healthcare production deployment processing tens of thousands of claims weekly across 9 markets with 95% document classification accuracy, 87% field-level extraction, and 300× efficiency gain (300min to 1min per claim) using compact Qwen VL and privacy-constrained architecture.
— Systematic error analysis of 4,000+ examples reveals modality gap: math tasks degrade 60+ points when text rendered as images. Self-distillation raises GSM8K from 30.71% to 92.72%, addressing core document QA limitation.
— Amazon Science framework enabling closed-box LLMs to generate field localization without external OCR or fine-tuning, solving privacy/cost constraints in business-critical document processing applications.
— Textract holds 2.3% IDP market share; Associa production deployment across 48M documents improved unknown-document accuracy from 50% to 85%, demonstrating real-world scaling and ROI in high-volume enterprise environments.
— Systematic production failure: custom models copied from Dev to Test resource return HTTP 500 InternalServerError; issue persists across older and newly copied models, indicating resource-level backend inconsistency and deployment pipeline constraints.
— UiPath releases Field Groups, Monetary Quantity support, and Document Understanding API v2 (Preview), enabling hierarchical field organization and taxonomy-driven extraction, signaling vendor product maturation and ecosystem advancement.
— Real-world POC planning for extracting Certificate of Analysis data from 12K–15K documents with >90% accuracy target, human review for <0.9 confidence fields, structured workflow design demonstrating production-stage multimodal document deployment.
— Empirical benchmark (Multimodal Finance Eval) evaluating six VLMs on French financial documents, revealing critical limitations: 34-62% accuracy on charts, error propagation in multi-turn dialogue (50% accuracy regardless of model size), indicating brittleness in interactive financial analysis.
— LlamaIndex agentic workflows process 500M+ documents with 90%+ automation; Convr delivers 97% accuracy on commercial insurance with 30-second processing; Mistral OCR at 2K pages/minute ($1 per 1K pages), demonstrating agentic system maturation and market-scale adoption.
— Allegis Global Solutions production deployment combining UiPath Document Understanding with Agent Builder and Maestro to handle constantly changing invoice formats, achieving 80-90% success rates in real-time adaptation.
— Independent benchmark comparing OCR and multimodal LLM accuracy across 300 documents shows Azure Document Intelligence leading at 96% on printed text, with analysis demonstrating SOTA multimodal LLMs now provide viable alternative to traditional OCR.
— Production issue report documenting Azure Document Intelligence extraction requests hanging indefinitely without timing out in January 2026, causing application downtime and reinforcing January 2026 vendor platform reliability constraints.
— Oracle Cloud Infrastructure Document Understanding action integrated into Oracle Integration Cloud, supporting extraction from invoices, receipts, passports, healthcare IDs, and custom documents with prebuilt and custom AI models.
— Document Force announces new basic models with 200%+ improvement on GPQA performance metrics and cost reduction up to 1/3 through prompt caching, signaling international vendor innovation in multimodal document AI.
— SANER 2026 conference paper from Fujitsu proposes multimodal LLM method for automatic design document review, achieving high accuracy in structural recognition but identifying persistent challenges in semantic-level interpretation of complex diagrams.
— September 2025 survey of 465 organizations finds 64.5% have AI in production but only 38.1% rate document data as excellent for AI use, with 76.6% storing 25-75% data in documents and 82.8% planning document automation investment, revealing infrastructure readiness as Q4 adoption bottleneck.
— Production outage report: Azure Document Intelligence hanging indefinitely without errors on custom classification and extraction models in November 2025, affecting real deployments and reinforcing Q4 vendor platform reliability constraints.
— Indico Data's agentic AI platform processes over 1 million pay stubs daily for US mortgage provider with 90%+ extraction accuracy after replacing hyperscaler solution (70% accuracy), signaling Q4 shift toward agentic automation and vertical vendor specialization.
— Production emergency report of Azure Document Intelligence hanging indefinitely on prebuilt-read model in West Europe, causing 100% processing failure on medical journal PDFs (previously <1 second, now timing out), rendering healthcare application unusable and signaling regional/model-specific reliability failures.
— User-reported production outage of Azure Document Intelligence (September 18-19) causing 30+ minute processing delays on 50-page PDFs in East US region, highlighting persistent vendor service reliability constraints affecting production deployments.
— SER Group survey reports 78% organizational adoption of intelligent document processing but reveals critical implementation gaps: 61% workflows still rely on paper, 48% expect paper volumes to rise, most deployments remain rule-based rather than true AI-driven, signaling adoption headroom constrained by integration friction.
— Oracle Cloud Document Understanding Version 2.0 announces multilingual support and Label Studio integration, signaling third-tier vendor entry and continued platform maturation across enterprise cloud ecosystems.
— Comprehensive arXiv survey of MLLM-based visually-rich document understanding covering methods, training paradigms, datasets, and challenges, synthesizing research maturity and highlighting efficiency, generalizability, and robustness as key advancement frontiers.
— ACL 2025 Findings peer-reviewed survey systematically reviewing text-rich image understanding MLLMs, covering timeline, architecture, performance benchmarks, and future directions, confirming research consensus on field maturity and rapid evolution.
— AWS announces support for superscripts, subscripts, rotated text, and improved low-resolution document handling, signaling continued platform refinement and vendor commitment to document understanding capabilities.
— Production outage report of Azure Document Intelligence in US East region (June 17, 2025) with multiple users experiencing InternalServerError and service failures on both custom and prebuilt models, highlighting vendor reliability and availability constraints.
— Vendor case studies detail multimodal deployments including pharmaceutical company automating lab/trial reports, insurtech achieving 99.7% accuracy on supplier forms, and banking integrations reducing approval cycles, demonstrating enterprise adoption momentum.
— Peer-reviewed PLOS ONE paper presents MDKG-RL model for multimodal archival retrieval achieving 0.85 MRR, 0.88 NDCG, 92.4% entity linking accuracy with 38.2% faster response time, advancing technical foundations for document understanding systems.
— RMIT University independent evaluation of Textract AnalyzeExpense API identifies strengths in total detection but reveals limitations in vendor name/date extraction, language handling, and image quality sensitivity, signaling real-world adoption constraints.
— 3.5-hour tutorial presented at IJCAI 2025 systematically covers MLLM-driven document understanding frameworks, key tasks (layout analysis, KIE, DocVQA), benchmarks, and hands-on labs, signaling formal knowledge dissemination and academic-practitioner alignment.
— CVPR 2025 paper introducing Docopilot native document-level VLM and Doc-750K dataset with 758K QA pairs, outperforming Gemini-1.5-Pro on MMLongBench-Doc and achieving inference latency improvements over RAG-based approaches.
— ACL 2025 Findings survey analyzing multimodal RAG systems including document understanding as key application domain, covering retrieval methodologies, fusion strategies, and innovations signaling research community engagement and maturity.
— Production issue report of Azure Document Intelligence freezing intermittently with trained custom models, indicating reliability and scalability limitations in real deployments on constrained service plans.
— ICCV 2025 Workshop paper introducing Document Haystack benchmark with 400 document variants and 8,250 questions for evaluating VLMs on long, visually complex documents with needle-in-haystack retrieval challenges.
— ICLR 2025 research introducing BigDocs dataset with 7.5M multimodal samples across 30 document tasks, showing up to 15.14% improvement on document benchmarks and surpassing proprietary models by 25.8% on BigDocs-Bench.
— Production deployment processing 50,000 invoices monthly across 15 countries achieved 85% processing time reduction (to 2 minutes per invoice), 95% error reduction, and $2.1M annual savings with ROI in 7 months.
— Baidu PaddlePaddle research presenting novel MLLM with synthetic Chinese document dataset (477k samples) achieving SOTA on English benchmarks while outperforming open-source and commercial models on Chinese document understanding tasks.
— Developer analysis revealing Azure Document Intelligence remains in preview despite rebrand, with GA API only accessible via legacy 'azure-ai-formrecognizer' library, indicating SDK maturity and adoption friction.
— Production deployment reporting intermittent HTTP 400 errors with Azure Document Intelligence AnalyzeDocumentAsync, indicating service reliability issues and input validation robustness constraints during real-world use.
— ACL 2024 research identifying critical enterprise adoption barriers: data limitations (under-represented tasks), model issues (calibration, licensing), and evaluation gaps in field-level performance and reading order comprehension.
— Bloomberg and UNC Chapel Hill research on multimodal RAG framework handling 40,000 pages across 3,368 documents with sub-2-second retrieval latency and 36.5% F1 on open-domain VQA, demonstrating advanced scalable architecture.
— Federal agency MLaaS platform deployment by Precise Software Solutions (AWS Advanced Tier Partner) using Textract, achieving four-fold productivity improvement in document processing workflows.
— Named edtech company Stride deployed UiPath Document Understanding at scale, replacing manual enrollment document processing with 79% classification automation, 58% extraction automation, achieving $67K cost savings and 72% throughput improvement.
— ACL 2024 peer-reviewed research introducing fine-grained and coarse-grained knowledge distillation approach for form document understanding, demonstrating technical advancement that outperforms existing baselines in handling complex visually-rich document structures.
— Practitioner report of production on-premises UiPath Document Understanding deployment encountering dynamic ML skill failures, pipeline cascading failures with dataset growth, and infrastructure scaling issues, indicating operational maturity constraints in real deployments.
— Named enterprise Deltek deployed multimodal RAG solution using Amazon Textract and Bedrock for government document Q&A, demonstrating production adoption with specific technical implementation in collaboration with AWS Generative AI Innovation Center.
— NPJ Digital Medicine peer-reviewed study reveals critical limitations in GPT-4V for medical document understanding, showing gaps between reported expert-level accuracy and actual reliability in handling complex medical documents, signaling domain-specific maturity constraints.
— IBM Research open-source multimodal document processor with 55.8k stars and production-ready capabilities for PDF understanding, table extraction, image classification, and gen AI integrations, signaling ecosystem maturity and adoption across developers.
— Stanford research introducing WONDERBREAD benchmark with 2,928 workflow demonstrations evaluating multimodal FMs on business process management tasks; finds models can document workflows at 88% recall but struggle with validation (F1 < 0.3), signaling capability gaps.
— JPMorgan research on layout-aware multimodal document understanding using OCR tokens and bounding boxes instead of image encoders; demonstrates novel architecture for handling formatted documents with disentangled spatial attention, signaling continued industry innovation.
— ICLR 2024 peer-reviewed evaluation of 10 open-source large multimodal models (3B-80B parameters) revealing major flaws in hallucinations, compositionality, and explainability; confirms scaling alone does not resolve these limitations.
— Amazon announces Bedrock adoption by tens of thousands of customers including named enterprises (NYSE processing thousands of pages of regulations, Ryanair automating crew manuals, Netsmart targeting 50% reduction in health records management time).
— CTO technical analysis of production-scale Azure Document Intelligence limitations: mandatory polling (75-90% of processing time), rate limits (15 TPS POST, 50 TPS GET) preventing horizontal scaling, requiring workarounds for high-volume deployments.
— UiPath internal deployment of Document Understanding for accounts payable automation; production instance processing ~1,000 invoices monthly with 716 automations achieving 70,677 hours freed in Q4 and $59M cumulative cost avoidance.
— Adobe launches generative AI-powered Document Assistant in Reader and Acrobat beta, enabling Q&A, summarization, and intelligent citation across multiple document formats, signaling major vendor ecosystem expansion.
— GitHub issue reporting persistent 404 Resource Not Found errors in Azure Document Intelligence SDK, indicating deployment and integration challenges in vendor tooling at the time.
— Multiple named organizations (Change Healthcare, Symbeo, NHS BSA) deploying Textract for document understanding in production, with metrics showing 68% automation rate and processing acceleration (3 minutes reduced to under 1 minute).
— Academic benchmark introducing layout-level retrieval granularity for multimodal documents, addressing gaps in existing benchmarks for page and table/figure retrieval tasks across 313 documents.
— Enlyft reports Amazon Textract used by 1,729 companies with 0.7% ML market share, spanning Information Technology (33%), Computer Software (17%), Financial Services (8%), indicating enterprise-wide adoption across sectors.
— Benchmark with 851 samples for multimodal understanding of multi-hundred-page documents with texts, figures, and tables, plus retrieval-aware tuning framework achieving 4.6% improvement over baseline models.