100%

DocVQADocVQA

#1 für Luna · ICDAR 2026#1 for Luna · ICDAR 2026

  • ICDAR DocVQA — seit 2020 · populärster Track der Konferenzsince 2020 · the conference's most popular track
  • Edition 20262026 edition — bislang anspruchsvollste Ausgabethe toughest edition yet
  • Top-5 schwerste CompetitionTop-5 hardest competition weltweit — härteste im Document Understandingworldwide — #1 in document understanding
  • Tiefes ReasoningDeep reasoning · Planung · mehrstufige Inferenz über reale Dokumente· planning · multi-step inference over real documents
  • Microsoft · Google · Alibaba · Anthropic · OpenAI · Meta
The 8 DocVQA 2026 document domains — business reports, engineering drawings, scientific publications, comics, maps, posters, slides, infographics
8 Dokument-Domänen · 240 Fragen — alle von Menschen erstellt & validiert8 document domains · 240 questions — all human-written & validated
Spatial · temporal · multi-hop Reasoning — Evidenz über Text, Tabellen, Abbildungen & Seiten verteiltSpatial · temporal · multi-hop reasoning — evidence spread across text, tables, figures & pages

DocVQA 2026 Challenge ↗DocVQA 2026 Challenge ↗ · ICDAR 2026 · Wien ↗ICDAR 2026 · Vienna ↗

Warum das schwer ist — ein BeispielWhy this is hard — a worked example

Räumliches + mehrstufiges Reasoning über ein reales Dokument · Kategorie mapsSpatial + multi-hop reasoning over a real-world document · maps category

Zoom: legend
5Namen in der Legende nachschlagen.Look up the name in the legend.
Zoom: stand 16
4Der Hauptallee nach Norden folgen → letzter Stand Nr. 16.Follow the main avenue north → final stand no. 16.
Zoom: stand 30
2Startpunkt lokalisieren: Stand 30.Locate the starting point: stand 30.
Zoom: compass rose
3Norden bestimmen — die Kompassrose.Establish north — the compass rose.
California's 86th Annual State Fair — 1940 illustrated map (Sacramento)
👆 Klick auf ein Bild für Zoom👆 Click any image to zoom
1Q: I'm at stand number 30, if I keep going north in the main avenue, what is the name of the last stand that I will encounter?
GT: Stand 16 · City of Sacramento Water Department
✗Gemini-3-proMain Entrance
✗GPT-5.4Farm and Road Machinery
✗Qwen3-VL-32B-ThinkingNo answer after 8192 thinking tokens
✓LunaCity of Sacramento Water Department
SchrittStep 1 / 6→ blättert Schritt für Schritt durch die Reasoning-Kette→ steps through the reasoning chain
🎮 Der DocVQA Playground startet auf Slide 8 — ungeduldig? Direkt hinspringen →🎮 The DocVQA Playground starts on slide 8 — impatient? Skip ahead →

Luna: architecture beats scale

Luna = two layers on top of any foundation model — IDA reads · DCA thinks

ApproachDocVQA 2026
Single frontier model (Gemini-3.1)~29 %
Mixture-of-Experts (4 top models)~40 %
Luna (IDA + DCA on top)~60 %
IDA — ReadingDCA — Thinking + Acting
Pixels → KnowledgeKnowledge → Result
+ 7 pp on DocVQA+ 13 pp on DocVQA
Value Chain — Frontier Models vs. Luna (IDA + Frontier + DCA)
Both layers are complementary to the frontier model — they add performance on top, they don't replace it. Every frontier-model upgrade lifts the baseline; IDA and DCA stack on top.

Paper: DCA at DocVQA 2026 — Technical Report ↗

Layer 1 · IDA — Reading

Pixels → structured knowledge · Planet AI's heritage · 10+ years of research · 7 ICDAR awards

AspectDetail
What IDA doesConverts pixels into structured knowledge — exact tables, correct numbers, layout preserved
Why Luna needs IDADeterministic precision where VLMs hallucinate numbers, dates, totals
In production atIBM · Unicredit · adesso · German federal agencies
CapabilitiesLayout-aware OCR · table extraction · handwriting · multi-page documents · 30+ languages

References: planet-ai.com/ida-en · awards

Layer 2 · DCA — Thinking as Teamwork

Distributed Cognitive Architecture — four pillars turn foundation models into one cooperating team

PillarWhat it does
Agent TeamworkSpecialized agents cooperate, hand off and cross-check each other — like regions of the neocortex
Memory & Context ManagementState persists and updates across steps — far beyond any single context window
Convergence — without hallucinationsAnswers are iterated and cross-verified until they are stable, instead of one-shot guessing
Learning from FeedbackThe team improves with every solved case — no model retraining needed

Papers: DCA — Theory I: Atomic Agents · Fractal Composition · Convergence ↗ · DCA — Foundations: Predictive Hierarchies · Thalamo-Cortical Convergence ↗ · Tech Report: DCA at DocVQA 2026 ↗

DCA in practice — Agent Architecture

Luna Agent Architecture — Reading (IDA) + VLM Agents + Reasoning Agent

Reading agents · VLM agents · Reasoning agent — orchestrated by Luna. Fractal composition: agents orchestrate agents — same blueprint at every level, from a single question to a multi-document workflow.

DocVQA 2026 — Ranking

DocVQA 2026 Ranking ICDAR 2026 — Certificate of Achievement · First Place DocVQA2026 · Luna / Planet AI

Live-Leaderboard, out-of-competition Einträge entfernt · Quelle:Live leaderboard, out-of-competition entries removed · Source: rrc.cvc.uab.es — offizielle Winner-Bekanntgaberrc.cvc.uab.es — official winner announcement

🎮 DocVQA Playground

Die Übersicht: 8 Kategorien · je 10 Fragen · Scores beider SystemeThe overview: 8 categories · 10 questions each · both systems' scores

➜ Der interaktive Teil startet auf den nächsten Slides — eine Kategorie pro Slide, jede Frage klickbar. Die Kurzanleitung steht über jeder Tabelle.➜ The interactive part starts on the next slides — one category per slide, every question clickable. A quick guide sits above each table.

KategorieCategory DokumenttypDocument type N Gemini-3.1 (Baseline) ✓ Luna ✓
business_reportGeschäftsberichte (NVIDIA, Alphabet …) — bis zu 181 SeitenAnnual & business reports (NVIDIA, Alphabet …) — up to 181 pages102 (20.0%)7 (70.0%)
comicsComic-Seiten — visuelles Szenen-VerständnisComic pages — visual scene understanding103 (30.0%)6 (60.0%)
engineering_drawingTechnische Zeichnungen — Stücklisten, Bauteil-BezügeEngineering drawings — part lists, component references101 (10.0%)7 (70.0%)
infographicsInfografiken — Symbole, Diagramme, PiktogrammeInfographics — symbols, charts, pictograms107 (70.0%)8 (80.0%)
mapsKarten — räumliche Orientierung & ReasoningMaps — spatial orientation & reasoning100 (0.0%)6 (60.0%)
science_paperWissenschaftliche Paper — Abbildungen, Captions, FormelnScientific papers — figures, captions, formulas102 (20.0%)6 (60.0%)
science_posterWissenschaftliche Poster — dichte Layouts, PlotsScientific posters — dense layouts, plots104 (40.0%)9 (90.0%)
slidePräsentations-Slides — Finanzkennzahlen, DiagrammePresentation slides — financial figures, charts104 (40.0%)8 (80.0%)
GesamtTotal8023 (28.8%)57 (71.2%)
Luna gewinnt in jeder Domain · +42.5 pp absolut — schaffst du mehr als 28.8%?Luna wins in every domain · +42.5 pp absolute — can you beat 28.8%?

business_report (10)

Baseline 20% · Luna 70%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
business_report_1_q1In Fiscal 2025, by how many dollars does NVIDIA’s TSR value exceed the Nasdaq-100 Index TSR value? Do not provide units.2048.88✓ 2048.88✓ 2048.88
business_report_1_q2 ★What is the title of the bullet point that is in the same position (starting from the bottom) as the element in the pre…Agent Blueprints✗ Unknown✓ Agent Blueprints
business_report_2_q1What was the highest increment dollars YoY on the share repurchases? Provide the result in millions and only the number.9633✗ Unknown✓ 9633
business_report_3_q1In one page there is a pictogram of a stroller, in the page before, there is a pictogram with a bush. What is the title…child education✗ Industry-leading reward✗ Flexible Work
business_report_3_q2 ★What was the percentage difference between 2024 and 2023 in unionization rates?4%✗ Unknown✓ 4%
business_report_4_q1What word is used to describe the Class A and Class C stock?volitale✗ Unknown✓ volitale
business_report_4_q2 ★How is the Google Subscriptions broken down? Respond with the first words up to the first comma of the bullet points. S…consumer subscriptions, platforms, devices, …✗ consumer subscriptions, platforms, …✓ consumer subscriptions, platforms, devices, …
business_report_4_q3How are the Google Cloud revenues broken down? Respond with the first words up to the first comma of the bullet points.…Google Cloud Platform, Google Workspace, …✓ Google Cloud Platform, Google Workspace, …✓ Google Cloud Platform, Google Workspace, …
business_report_4_q4Into which items is the Google Network revenue broken down?Google advertising✗ AdMob, AdSense, Google Ad Manager✗ AdMob, AdSense, Google Ad Manager
business_report_4_q5What are the main components of the R&D expenses? Respond with the first words up to the first comma of the bullet poin…depreciation, employee compensation expenses for engineering …✗ depreciation, employee compensation expenses, third-party …✗ depreciation, employee compensation expenses for engineering …

comics (10)

Baseline 30% · Luna 60%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
comics_1_q1How many times do people get in the head in Nyoka and the Witch Doctor's Madness?4✗ [chain-of-thought, ends with 7]✗ 7
comics_2_q1What was the last state that the man murdered in front of the diner visited?Indiana✗ Missouri✓ Indiana
comics_2_q2What character was wearing a leopard-skin coat on the final story? Answer with the surname.Milanoff✓ Milanoff✓ Milanoff
comics_2_q3How many years after Paul's death did Madeleine attempt to kill Pietro?5✗ Unknown✓ 5
comics_2_q4Which color is the dress of Benny Mickson's wife while they are in south dakotaolive green / green / dark green✗ Green✗ Green
comics_3_q1How many faces does Ted have in reality? Answer with a number1✓ 1✓ 1
comics_3_q2How does Landor refer to this creation the second time he does?Tiger-man✗ MY LITTLE PET✗ My little pet
comics_3_q3By which entity is the comic drawn by Rem Brant created?U.S. Marines✓ U.S. MARINES✓ U.S. Marines
comics_4_q1What object is thrown to the Music Master before he enters the pipe?wrench / spanner / monkey wrench …✗ Wrench✗ A wrench
comics_4_q2 ★What is the name of the comic featuring a man riding a horse that has a human head?Hortense✗ [chain-of-thought, ends with Jingle Jangle]✓ Hortense

engineering_drawing (10)

Baseline 10% · Luna 70%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
engineering_drawing_1_q1In which direction can you see the eight temperature sensing manifold?D✓ D✓ D
engineering_drawing_1_q2How many clamps are used in total?27✗ Unknown✓ 27
engineering_drawing_1_q3Which labeled parts are used at the end of the tube with identifying number 1901105? List the part identification numbe…1901069-002, 1901094, AN 910-2, 1016315-001✗ [chain-of-thought, ends with 124, 123]✗ 144, 141
engineering_drawing_2_q1How many matrix assy are used in the fixed memory module assembly?22✗ 21✓ 22
engineering_drawing_2_q2 ★Which is the identifying number of the part that is placed before the screw jacking?1004546-1✗ 13✓ 1004546-1
engineering_drawing_3_q1How many insulation sleeving shrink fit are used?68✗ 44, 2✓ 68
engineering_drawing_3_q2What is the maximum length between E1 and part 1897336-001? Do not provide the unit.8.50✗ [chain-of-thought, ends with 4 FEET 6 INCHES]✓ 8.50
engineering_drawing_3_q3What is the note number under the PARTIAL VIEW schematic?8✗ [chain-of-thought, ends with 8]✓ 8
engineering_drawing_4_q1What is the distance between the left side and the LOGIC MODULE inscription? Provide the unit as well3.7 in✗ [chain-of-thought]✗ 3.70 inches
engineering_drawing_4_q2What is the width of the piece? Respond with the unit as well0.598 in✗ [chain-of-thought]✗ 8.850 / 8.830 inches

infographics (10)

Baseline 70% · Luna 80%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
infographics_1_q1 ★In which year was made the media whose symbol resembles the most to a robot face?1987✗ 1965✓ 1987
infographics_1_q2What is the date of discovery of the element for which the difference between the years of use for the two types of usa…Unkown✗ Unknown✗ 1976
infographics_2_q1Assuming you get 15 USD/hour in Angola what's the number of coffees you could buy for the price of an ipod? Answer usin…Unknown✗ [chain-of-thought]✗ 181
infographics_2_q2What is the cost in USD of gasoline of the city depicted with a brown person?1.94 USD✓ 1.94 USD✓ 1.94 USD
infographics_2_q3Which city is depicted with a brown pyramid-like shaped that is not a building?Tokyo✓ Tokyo✓ Tokyo
infographics_2_q4Which city is depicted with a large round object?Osaka✓ Osaka✓ Osaka
infographics_2_q5Which one is the last city on the list to not have the iPod statistic depicted?Libreville✓ Libreville✓ Libreville
infographics_2_q6Which one is the last city on the list to not have the iPod statistic depicted?Libreville✓ Libreville✓ Libreville
infographics_2_q7Which city is represented with the largest element (not real sized)?Hong Kong✓ Hong Kong✓ Hong Kong
infographics_2_q8Which city is depicted with the largest amount of green in its symbolic illustration?Geneva✓ Geneva✓ Geneva

maps (10)

Baseline 0% · Luna 60% · Δ +60 pp

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
maps_1_q1 ★If I'm standing at the Pantheon and looking toward the Colosseum, which hill will I see most clearly to the right of it?Capitolinus✗ MONS COELIO✓ Capitolinus
maps_1_q2I'm at the Panteon, then I walk in the direction of the Capitolinvs, before reaching it I turn left, and after a bit mo…P.Popoli✗ P. POPOLI✗ P. S. LAVRETII
maps_2_q1By which kind of road are Colchester and Yantic connected?Macadam & Gravel✗ Paved & Hard Surfaced Roads✓ Macadam & Gravel
maps_2_q2Which is the most populated town in the E-10 coordinates?Wareham✗ Unknown✓ Wareham
maps_2_q3What is the milage between Taunton and Dedham? Do not provide the unit.27✗ 22✗ 22
maps_2_q4From Worcester I take highway 140 towards Taunton, I take the second macadam & gravel road that I encounter, continuing…Woonsocket✗ Franklin✗ Bellingham
maps_2_q5If I follow highway 109 from Pittsfield to Northampton, how many towns do I cross (without counting start and ending lo…7✗ 6✓ 7
maps_3_q1Im at stand number 30, if I keep going north in the main avenue, what is the last stand that I will encounter?16✗ Industrial Palace✓ 16
maps_3_q2I am at the entrance (stand number 1), if i go straight until I encounter a fountain, I go left past a garden, then tur…4✗ 12✗ 23
maps_3_q3I start at stand 33 going straight inside of the state fair, pass an stadium, turn right and continue walking along the…42✗ 32✓ 42

science_paper (10)

Baseline 20% · Luna 60%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
science_paper_1_q1How much data (in Millions) do they use to ablate the Perception Encoder? Respond with the number.17✗ Unknown✓ 17
science_paper_1_q2What family of models mentioned in the paper can be used to make comparisons that involve native resolution?qwen✗ [chain-of-thought]✗ Vision encoders
science_paper_1_q3When performing the spatial alignment, what is the type of similarity that is used on the first plot when reporting for…Unknown✗ pairwise feature cosine similarity✗ cosine similarity
science_paper_1_q4In the page before the "Related Work", what is the last citation number used?76✓ 76✓ 76
science_paper_1_q5In the page before the "Related Work", what is the first citation number used? Exclude the ones present in figure capti…99✗ [chain-of-thought]✓ 99
science_paper_1_q6What are the MLLM Results utilizing QwenLM3 8B and InternViT2.5 0.3B when applied to the DocVQA task?Unknown✓ Unknown✗ 61.1
science_paper_1_q7What did the authors of CLIPer identified? Answer with a quote of the paperearly layers in CLIP possess good spatial …✗ Unknown✓ early layers in CLIP possess good spatial …
science_paper_2_q1 ★In the figure with peaks (that seem like frequency peaks). There is a row with a clearly visible last peak. What is the…making✗ localization✓ making
science_paper_3_q1What is the last citation number that is used in the page before the "Comparison with SOTA Models"?18✗ 226✓ 18
science_paper_3_q2Which datasets do the authors consider internal data?Textbooks-QA, Arxiv2Markdown✗ Arxiv2Markdown, Internal TextbookQA Dataset, …✗ Arxiv2Markdown, Textbooks-QA

science_poster (10)

Baseline 40% · Luna 90%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
science_poster_1_q1What is the percentage score improvement from Baseline to TexTok in rFID-50k for the ImageNet 512x512 case with 128 tok…30.2%✓ 30.2%✓ 30.2%
science_poster_1_q2What is the percentage score improvement from Baseline to TexTok in the ImageNet 512x512 case with 128 tokens?30.2%✓ 30.2%✓ 30.2%
science_poster_1_q3Which model has the fastest inference speed for ImageNet 256x256?TexTok-32 + DiT-XL✓ TexTok-32 + DiT-XL✓ TexTok-32 + DiT-XL
science_poster_1_q4There are yellow bounding boxes in one figure located near the center. What is the column that has the biggest ones?128✗ Unknown✓ 128
science_poster_1_q5On the cat example, what is the last word highlighted on the 4th last sentence?green✗ Unknown✓ green
science_poster_2_q1What is the biggest model that the authors include at the table? Respond with the param count in millions.540000✗ [chain-of-thought]✓ 540000
science_poster_2_q2Which benchmark is to the right of the one that has a 3.4 increment in performance?WinoGrande✗ TydiQA✗ TydiQA
science_poster_2_q3 ★Looking at the two bar plots, from one that more closely resembles a Gaussian distribution, provide the second set of c…ARC-C✗ Unknown✓ ARC-C
science_poster_2_q4On the results table, what is the only number in red that does not have any more red numbers in its row? Do not include…24.7✗ Unknown✓ 24.7
science_poster_2_q5In the results table, what is the largest number highlighted in blue that is located highest in the table? Do not inclu…79.0✓ 79.0✓ 79.0

slide (10)

Baseline 40% · Luna 80%

👆 Frage anklicken → Deep Dive mit Original-Dokument & allen Antworten👆 Click any question → deep dive with original document & all answers 🎯 Erst selbst lösen, dann vergleichen🎯 Solve it yourself first, then compare ★ = Reasoning-Panel mit Hinweisen★ = reasoning panel with hints
Question IDTaskTargetGemini-3.1 (Baseline)Luna
slide_1_q1What is the correct number of total tuna releases?201425✓ 201425✓ 201425
slide_1_q2How much did the percentage of SKJ increase between Releases and Recoveries?10%✗ -1%✗ 1%
slide_2_q1 ★Assuming Q3 revenue is distributed by customer in the same way as Q3 backlog, and costumers are equally distributed bet…23✗ Unknown✓ 23
slide_2_q2In the slide mentioning the path to first launch, which is the last lower case word in the paragraph closest to the che…to✗ program✓ to
slide_2_q3In the page before FINAL HIGHLIGHTS AND OUTLOOK, what is the last word written?mount✓ mount✓ mount
slide_3_q1On the big table shown, what is the mean QTE_EXPENSES_EVAL?Unknown✓ Unknown✓ Unknown
slide_3_q2What is the last word written on the page before the one that mentions "agricultural statistics as helpful"?takeaways✓ Takeaways✓ takeaways
slide_3_q3What is the ID of the table input two rows above the input with the fourth largest cost/area ratio?32050020877✗ Unknown✓ 32050020877
slide_3_q4Provide the name of the equation that has a sum element on the denominatorUnknown✗ [chain-of-thought]✗ SDG 2.3.1
slide_3_q5Considering the labor productivity of 2021 and the labor units available, what is the total labor productivity in $PPP …3120✗ Unknown✓ 3120

Luna gewinntLuna wins

First Place · 60% ANLS · deutlicher Abstand zum Runner-up (46,25%)First Place · 60% ANLS · significant gap to runner-up (46.25%)

Offizielle Winner-Bekanntgabe ↗Official winner announcement ↗ · Technical Report (Zenodo) ↗Technical Report (Zenodo) ↗

NAV
← → · T Theme · L Hell · G Sprache · 0 Reset · F Vollbild← → · T Theme · L Light · G Lang · 0 Reset · F Full
01 / 12
SYS