PaperlayerView exact revision on arXiv

Source-linked AI summary

PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence

GuangJian Team, Kaili Huang, Yongshuo Zhang, Bingtao Fu, Changjiang Jiang, Chenfan Qu, Chenfeng Zhang, Fangming Cui, Gaoyang Zhang, Jiangwei Xie, Jianshu Li, Jing Huang, Jingwen Bai, Mingqi Fang, Tao Fang, Weihong Zhang, Wenbo Du, Xiongfei Bai, Xuekang Zhu, Yinan Xia, Zhenming Wang, Jian Liu, Jingjing Liu, Xiang Qi, Weiqiang Wang

arXiv:2609.37712v1cs.CVcs.AI

TL;DR

Existing OCR systems often have uneven capabilities across recognition, parsing, and reasoning. PolyOCR combines a shared generative framework, large-scale quality-verified data construction, and competence-aware post-training. It reports strong, balanced performance across diverse OCR scenarios and competitive performance with specialized systems.

  • Problem

    Existing OCR systems either specialize in selected tasks or show uneven performance across OCR capabilities, making broad and balanced capability coverage a central challenge.

    Source: §1 · p. 2, §2.3 · p. 5

  • Method

    PolyOCR unifies OCR tasks in a shared generative framework, builds quality-verified multi-task supervision, and routes on-policy distillation according to teacher reliability and the teacher–student competence gap.

    Source: §1 · p. 3, §6 · p. 25

  • Results

    PolyOCR achieves strong and balanced performance across recognition, localization, structured document understanding, multilingual parsing, and information extraction, with competitive performance against specialized OCR systems.

    Source: §6 · p. 26

  • Takeaways & Limitations

    The evaluations support the feasibility of OCR foundation models that generalize across diverse text-containing visual environments while maintaining competitive performance with specialized systems.

    Source: §6 · p. 26

    A positive mean teacher–student competence gap alone does not establish useful supervision, and finite-rollout estimates may not reproduce under an independent draw.

    Source: §4.3 · p. 15

Abstract

from arXiv · show

Optical Character Recognition (OCR) is evolving from plain-text transcription toward general visual intelligence, requiring models to recognize, localize, and reason over textual information in complex visual environments. However, existing OCR systems often excel at only some tasks and struggle to balance recognition, parsing, and reasoning across scenarios. In this report, we present PolyOCR, a family of unified OCR foundation models of varying scales. PolyOCR combines a shared instruction-following framework with a large-scale data engine that converts heterogeneous visual resources into quality-verified OCR supervision. We introduce Competence-Guided Policy Optimization, which combines verifier-based Group Relative Policy Optimization with on-policy distillation through sample-wise routing based on teacher reliability and the teacher--student competence gap. We also introduce OCRBench v2.1, our revision of OCRBench v2 with manually verified annotation corrections and task-aligned scoring metrics. Extensive experiments across OCRBench v2.1, CC-OCR, in-house KIE Benchmark, OmniDocBench v1.6 and MDPBench demonstrate that PolyOCR achieves state-of-the-art or highly competitive performance.

1 Introduction

Existing OCR systems either specialize narrowly or acquire OCR capabilities unevenly, leaving unified recognition, localization, parsing, and reasoning difficult to balance. PolyOCR combines a shared generative framework, targeted data construction, and competence-aware post-training, and reports strong performance across diverse OCR scenarios.

1 Introduction

  • The 2B and 9B PolyOCR variants use a shared instruction-following, autoregressive framework for transcriptions, coordinates, structured records, documents, tables, formulas, translations, and open-form answers.The framework is built upon Qwen3.5 and supports knowledge sharing across recognition, localization, parsing, extraction, visual text understanding, and OCR reasoning.Source: §1 · pp. 2–3
  • Competence-Guided Policy Optimization combines verifier-based Group Relative Policy Optimization with on-policy distillation, adapting teacher supervision to teacher reliability and the evolving teacher–student competence gap.The authors describe this post-training framework as balancing heterogeneous task learning and preventing catastrophic forgetting.Source: §1 · p. 3
  • PolyOCR achieves state-of-the-art or highly competitive performance across OCRBench v2.1, CC-OCR, the in-house KIE Benchmark, OmniDocBench v1.6, and MDPBench.The model family reports balanced capabilities spanning text recognition, localization, document parsing, information extraction, visual text understanding, and OCR reasoning.Source: §1 · p. 3
  • A large-scale OCR data engine converts heterogeneous visual resources into quality-controlled, multi-task supervision through task-oriented construction, synthesis, and verification.A specialized PolyOCR-2.7B configuration is trained on an in-house corpus covering 21 language/script groups for diverse writing systems and acquisition conditions.Source: §1 · p. 3

2 Related Work

Prior OCR work spans specialized systems, document intelligence, OCR-oriented vision-language models, and data or training strategies, but the passages identify gaps in transfer, balanced capability coverage, and competence-aware optimization.

2.1 Specialized OCR Systems

  • Specialized OCR systems can perform strongly in target domains, but task-specific architectures and supervision fragment capabilities and limit transfer across OCR scenarios.The related work calls for a unified framework that learns heterogeneous OCR capabilities together.Source: §2.1 · p. 4

2.2 Document Intelligence Models

  • Document intelligence models recover page-level layout, reading order, structure, and semantics, while broader OCR capabilities across visual environments remain insufficiently unified.The cited passage contrasts page-centric emphasis with natural-scene perception, multilingual recognition, region-conditioned understanding, and OCR reasoning.Source: §2.2 · p. 4

2.3 OCR-oriented Vision-Language Foundation Models

  • General-purpose vision-language models offer a unified interface, yet OCR performance varies across fine-grained recognition, localization, extraction, and document reasoning.OCRBench and OCRBench v2 evaluations are cited as revealing substantial variation among OCR capability dimensions.Source: §2.3 · p. 5

2.4 OCR Data Construction and Capability-aware Post-training

  • OCR-oriented models broaden task coverage, but existing approaches differ in optimization focus and do not yet integrate capabilities with balanced supervision and adaptive optimization.Data construction remains costly and quality-sensitive; fixed sampling or static loss weighting may leave challenging or underrepresented capabilities insufficiently learned. Related approaches address outcomes, distillation, or broader task balancing, while the cited gap is explicit modeling of evolving teacher–student competence in OCR post-training.Source: §2.3 · p. 5, §2.4 · pp. 5–6

3 Data Construction

The data framework pairs broad curation across OCR capabilities with targeted synthesis for underrepresented tasks. It combines task-specific generation with verification and balancing, yielding approximately 60 million high-quality training instances.

3.1 Full-Spectrum OCR Data Collection

  • The full-spectrum corpus is organized around six benchmark-independent dimensions: text recognition, localization, document parsing, information extraction, visual text understanding, and OCR reasoning.Table 1 summarizes the organization of these capabilities in the training corpus.Source: §3.1 · p. 7, Table 1 · p. 9

3.2 Targeted Data-Construction Engine

  • Because naturally available datasets leave several capabilities underrepresented, the framework adds targeted synthetic and model-assisted supervision while preserving each task’s semantic, spatial, and output-format requirements.The targeted engine transforms heterogeneous visual resources into supervision for scene text, translation, KIE, document parsing, tables, and OCR reasoning.Source: §3.2 · p. 9

(1) Scene-Text and Translation Synthesis

  • Scene-text data combines cross-validated annotations of natural images with controllable text rendering, while bilingual translation samples retain verified regions and filter generated translations for consistency.Disagreeing scene-text annotations may receive human review; synthetic renderings are checked for visual-text consistency, and translation filters check symbols, entities, repetition, and length.Source: (1) Scene-Text and Translation Synthesis · p. 10

(2) Structured KIE & Document & Table Synthesis

  • Structured synthesis creates aligned KIE documents, selects diverse and complex pages for parsing, and builds full-page table data with localization and HTML annotations.The KIE pipeline preserves template semantics and layout; document annotations use independent parsers and correction, while table annotations use inter-model TEDS consistency and retain only pages whose tables pass checks.Source: (2) Structured KIE & Document & Table… · pp. 10–11

(3) Reasoning-oriented Annotation Synthesis

  • Reasoning supervision uses deterministic templates for character and instance counting, and validated model-generated questions, explanations, and answers for more complex OCR-VQA tasks.Generated examples cover document fields, key–value relations, tables, spatial reasoning, and other OCR-centric capabilities; invalid answers, unsupported reasoning, and formats are removed.Source: (3) Reasoning-oriented Annotation… · p. 11

3.3 Cross-Pipeline Quality Control

  • Quality control combines agreement-based review, structure-aware checks, and source-grounded validation, removing samples that fail the checks applicable to their task.Human adjudication is used for selected scene-text disagreements, while generated structures and reasoning outputs are checked against rendered images and source annotations.Source: §3.3 · p. 12
  • Approximately 60 million high-quality OCR training instances result from capability balancing and sampling, with hard examples mixed into broader coverage data without dominating the task distribution.Cross-pipeline quality control checks annotation agreement, structure and image consistency, and source-grounding; samples failing applicable checks are removed.Source: §3.3 · p. 12

4 Training

PolyOCR uses SFT to establish broad OCR capabilities, then refines them with CGPO, which combines verifier-based policy optimization and competence-routed on-policy distillation. Its SFT schedule allocates training across four task groups, while CGPO adapts teacher supervision to teacher quality and the teacher–student competence gap.

4.1 Training Pipeline

  • The two-stage pipeline uses SFT to establish broad OCR capabilities before CGPO refines them using verifier-based GRPO and adaptive on-policy distillation.The covered SFT tasks include recognition and localization, document parsing, table understanding, key information extraction, multilingual transformation, and OCR-centric visual question answering.Source: §4.1 · p. 12

4.2 Supervised Fine-Tuning

  • Across approximately 60 million SFT instances, TP, DS, OR, and GA receive 32.5%, 30.0%, 22.5%, and 15.0%, respectively.The three-phase schedule contains 20 million instances per phase; OR rises from 10.0% to 35.0% while GA remains at 15.0%, with early phases weighting TP and DS more heavily.Source: §4.2 · p. 13, Table 2 · p. 13

4.3 Competence-Guided Policy Optimization

  • CC-OCR scores are 81.72 for GRPO and 82.18 after adding fixed-weight OPD, supporting complementary supervision.GRPO remains active for outcome-level optimization while OPD adds token-level supervision along student-generated prefixes.Source: §4.3 · pp. 14, 16, 17
  • CGPO retains GRPO and routes OPD sample by sample using calibrated teacher quality and reproducible advantage over the student.OPD weight falls as student competence rises and the teacher’s advantage diminishes, while reliable guidance remains weighted by the teacher’s independently calibrated quality.Source: §4.3 · pp. 14, 16, 17
  • A positive teacher–student competence gap alone does not establish useful supervision, because low teacher quality or finite-rollout estimates may undermine its reliability.Routing therefore calibrates teacher quality and reproducibility of the teacher’s advantage; it operates at sample level, so local teacher errors may remain within high-reward samples.Source: §4.3 · pp. 15–16

5 Evaluation

PolyOCR is evaluated across OCR, document parsing, information extraction, multilingual parsing, and reasoning benchmarks using a refined OCRBench v2.1 protocol. Results show competitive cross-benchmark performance, with gains across several tasks and post-training comparisons.

5.1 Evaluation Protocol

  • The evaluation uses consistent benchmark versions, inference settings, and scoring procedures, with audited annotations and task-aligned criteria for OCRBench v2.1.The benchmark suite spans broad OCR capabilities, document parsing, multilingual parsing, and field-level information extraction.Source: §5.1 · pp. 17–18

5.2 Overall Comparison

  • PolyOCR-9B scores 80.42/78.37 on OCRBench v2.1 English/Chinese, 82.43 on CC-OCR, and 91.57 on OmniDocBench v1.6.The cross-benchmark comparison reports competitive and consistent performance across text-centric visual tasks.Source: §5.2 · p. 18, Table 3 · p. 19

5.3 OCRBench v2.1 Evaluation

  • PolyOCR-9B scores 80.42 and 78.37 overall on OCRBench v2.1 English and Chinese, respectively, ranking first in five of 13 language-specific categories.Its category-leading scores include English text recognition (82.99), text detection (70.84), element parsing (89.70), visual text understanding (87.58), and Chinese element parsing (86.86); some categories remain below the strongest baselines.Source: §5.3 · p. 19, Table 4 · p. 20

5.4 CC-OCR Evaluation

  • PolyOCR-9B scores 82.43 overall on CC-OCR and leads the Document Parsing and Key Information Extraction tracks with 71.71 and 94.02, respectively.The evaluation covers Multi-Scene OCR, Document Parsing, and KIE; it excludes the multilingual track.Source: §5.4 · p. 20, Table 5 · p. 21

5.5 OmniDocBench v1.6 Evaluation

  • On OmniDocBench v1.6, PolyOCR scores 91.57 for full-page processing and 95.52 with layout guidance, with formula CDM rising from 91.71 to 96.30 and table TEDS from 88.67 to 93.56.The layout-guided configuration delegates layout detection and reading-order prediction to PP-DocLayoutV3 before PolyOCR parses cropped regions.Source: §5.5 · p. 21, Table 6 · p. 22

5.6 In-house KIE Benchmark Evaluation

  • PolyOCR-9B achieves an overall KIE F1 of 93.60%, exceeding the reported scores of GPT-5.6-Sol (90.98%), Qwen3.6-35B-A3B (89.17%), and Qwen3.5-122B-A10B (89.03%).It leads on admission notices, while its passport result ranks second and its Xuexin certificate F1 is slightly below PolyOCR-2B.Source: §5.6 · p. 22, Table 7 · p. 23

5.7 MDPBench Evaluation

  • PolyOCR-2.7B scores 87.0% overall on MDPBench, 0.4 points above GPT-5.6-Sol, and exceeds the strongest OCR-specialized baseline by 3.7 points overall.It leads four of five summary measures, ties for the highest photographed-document score, ranks among the top two in 12 of 17 language/script groups, and leads six groups.Source: §5.7 · pp. 22–23, Table 8 · p. 24

5.8 Ablation Study of CGPO

  • CGPO leads the evaluated post-training strategies on all three benchmarks and exceeds cosine-decayed OPD by 2.19/2.06 OCRBench English/Chinese points, 0.11 CC-OCR points, and 1.45 OmniDocBench points.CGPO routes stronger OPD weights to samples where the teacher is estimated to offer reliable, non-redundant information; the routing surface disables OPD when calibrated teacher superiority is at most 0.5.Source: §5.8 · pp. 24–25, Table 9 · p. 25, Figure 5 · p. 25

6 Conclusion

PolyOCR combines capability-oriented data construction with competence-aware post-training in a unified framework for diverse OCR tasks. Evaluations report strong, balanced performance across OCR and document scenarios, while the authors identify further capability and robustness improvements as future work.

6 Conclusion

  • PolyOCR integrates real-world resources, controllable synthesis, and model-assisted annotation to provide quality-controlled supervision across heterogeneous OCR capabilities.CGPO coordinates outcome-based policy optimization and on-policy distillation according to the evolving teacher–student competence boundary.Source: §6 · pp. 25–26
  • Evaluations across OCRBench v2.1, CC-OCR, OmniDocBench v1.6, MDPBench, and the in-house KIE benchmark show strong, balanced performance across text, document, multilingual, and extraction tasks.The conclusion states that the results support the feasibility of OCR foundation models generalizing across diverse text-containing visual environments while remaining competitive with specialized systems.Source: §6 · p. 26
  • Future work targets stronger fine-grained text perception, localization, recognition, semantic understanding, reasoning, and robustness in complex real-world environments.Source: §6 · p. 26

B Implementation and Reproducibility Details

The appendix records shared model, input, SFT, and CGPO configurations for PolyOCR variants. Training combines multi-task supervised fine-tuning with a calibrated routing procedure for post-training.

B.1 Model and Input Configuration

  • PolyOCR variants share a task interface, data engine, and SFT/CGPO settings; the model and input configuration is recorded in Table 10.The appendix states that the image-area entry reflects the SFT launch-script environment setting.Source: Table 10 · p. 33, §B.1 · p. 33

B.2 Supervised Fine-Tuning Configuration

  • SFT jointly trains across recognition, localization, spotting, extraction, document and table parsing, formula recognition, translation, and OCR-centric reasoning.PolyOCR-2B and PolyOCR-9B share the described SFT configuration and three-phase, 60M-instance schedule.Source: §B.2 · pp. 33–34, Table 11 · p. 34

B.3 Competence-Guided Policy Optimization Configuration

  • CGPO starts from each variant’s SFT checkpoint and uses a frozen Qwen3.5-122B-A10B teacher, verifier-based competence estimates, and fixed calibration maps.Teacher competence uses eight offline generations per sample; student competence uses eight current rollouts per group, and calibration coefficients remain fixed during optimization.Source: §B.3 · p. 34, Table 12 · p. 35
  • The stopped-gradient OPD weight is zero when calibrated teacher superiority is at most 0.5; otherwise it scales the OPD term alongside the active GRPO objective.The reference-policy KL penalty is disabled, while the reverse-KL objective for teacher distillation is retained.Source: §B.3 · p. 35

C Task Interfaces and Output Normalization

PolyOCR uses autoregressive generation with task-appropriate representations and verifiers. Output normalization and optional LLM judging provide task-specific scoring, while the operational routing procedure combines GRPO and conditionally weighted OPD.

C Task Interfaces and Output Normalization

  • Task interfaces use autoregressive conditional generation with task-appropriate output representations and scalar verifier rewards.The appendix says exact serialization grammars, prompt templates, and normalization code will be released with the model and evaluator.Source: C Task Interfaces and Output… · p. 35, Table 13 · p. 36

C.1 Normalization Order

  • Verifier preprocessing parses outputs, validates schemas, canonicalizes representations, matches predictions to references, and assigns scalar scores.Parsing failures and out-of-schema outputs receive zero reward; canonicalization does not repair missing fields, hallucinations, invalid coordinates, or incorrect structure.Source: §C.1 · p. 36

C.2 LLM-Judge Controls

  • LLM judges score translation and selected open-ended VQA cases when deterministic surface matching is insufficient, returning a constrained score from 0 to 1.The described setup uses Gemini-3.5-Flash-Lite with deterministic decoding and logs and reruns judge failures under the same frozen configuration.Source: §C.2 · p. 36

D Operational Routing Procedure

  • For each sample, CGPO scores student rollouts, estimates competence and teacher superiority, computes GRPO and—when OPD weight is positive—reverse-KL loss, then updates the student.When calibrated teacher superiority is at most 0.5, OPD weight is zero; the teacher and calibration maps remain frozen.Source: D Operational Routing Procedure · pp. 36–37

E OCRBench v2.1 Evaluation Protocol

OCRBench v2.1 refines the original bilingual benchmark’s annotations and evaluation procedures while retaining its sample set and reporting scheme. The changes target annotation quality and metric consistency without changing benchmark composition, task distribution, or scope.

E OCRBench v2.1 Evaluation Protocol

  • OCRBench v2 is a bilingual, text-centric benchmark with 10,000 human-verified question-answer pairs spanning diverse visual scenarios.Its subtasks are reported as 30 subtasks: 21 English and 9 Chinese, aggregated into 13 language-specific task groups.Source: E OCRBench v2.1 Evaluation Protocol · p. 37
  • The audit targets annotation noise and format inconsistencies in references, as well as evaluation criteria that may not align with some tasks’ intended semantic objectives.Source: E OCRBench v2.1 Evaluation Protocol · p. 37
  • OCRBench v2.1 retains the original 10,000 samples and 30-subtask reporting scheme while incorporating manually verified annotation corrections and revised evaluation procedures.The refinement does not modify benchmark composition, task distribution, or evaluation scope.Source: E OCRBench v2.1 Evaluation Protocol · p. 37
  • All models in the report are re-evaluated using the OCRBench v2.1 protocol for fair comparison.Source: E OCRBench v2.1 Evaluation Protocol · p. 37

E.1 Annotation Cleaning

The OCRBench v2.1 audit manually revises question and answer fields to improve reference quality, format consistency, and instruction clarity. It preserves the benchmark’s samples, indices, task categories, bounding boxes, and evaluation configurations.

E.1 Annotation Cleaning

  • The revisions comprise format normalization, reference correction, and instruction refinement; a sample modified in both question and answer fields is counted in each applicable type.The type counts therefore slightly exceed the 2,449 unique revised samples.Source: Table 14 · p. 38, §E.1 · p. 38
  • 2,449 samples (24.5%) are revised, with no samples added or removed and original indices, task categories, bounding boxes, and evaluation configurations unchanged.All modifications are restricted to question and answer fields and manually inspected.Source: §E.1 · pp. 37–38
  • Format normalization standardizes structured annotations without changing semantics, including table HTML, JSON-like references, and text-grounding coordinate serialization.The changes remove non-essential table markup and convert chart-parsing and key-information-mapping references to valid JSON strings.Source: §E.1 · p. 38
  • Reference correction fixes inaccurate or incomplete answers, mainly in multiple-choice reasoning, key information extraction, and text translation.Examples include erroneous option labels, inaccurate field values, OCR text errors, currency formatting, and translation references; valid alternatives are enumerated when answers are semantically acceptable.Source: §E.1 · p. 38
  • Instruction refinement clarifies key-information-extraction requirements, while English and Chinese KIE prompts receive unified field specifications and explicit JSON-format requirements, respectively.These changes aim to make extraction metrics reflect identifying and organizing visual information rather than interpreting underspecified output formats.Source: §E.1 · p. 38

E.2 Metric Revisions

OCRBench v2.1 revises selected metrics to better align scoring with the intended OCR capabilities while retaining the original metrics and aggregation for other subtasks. The changes address content accuracy, visual correctness, semantic equivalence, and acceptable answer variation.

E.2 Metric Revisions

  • Other subtasks retain their original metrics and the official aggregation strategy, including the 13 task groups and English and Chinese overall scores.Retained metrics include exact matching or ANLS for VQA, TEDS for tables, IoU-based grounding and spotting, F1 for information extraction, and normalized L1 for counting.Source: §E.2 · p. 39
  • Formula recognition uses CDM to compare rendered formulas at the visual-token level, improving robustness to LaTeX syntax variation while remaining sensitive to visual correctness.The metric matches tokens with geometric verification and computes an F1 score.Source: §E.2 · p. 39
  • Document parsing is scored at the character, formula-rendering, and table-structure levels simultaneously.The revision replaces structural similarity alone with content-aware md2md evaluation, decomposing predicted and reference Markdown into typed elements.Source: §E.2 · pp. 39–40
  • Normalized edit distance replaces the original multi-metric aggregation for text recognition and full-page OCR, directly measuring character-level recognition accuracy.The original aggregation included BLEU, METEOR, token-level F1, and edit distance.Source: §E.2 · p. 40
  • Text translation is scored by an LLM-based judge for semantic equivalence, translation completeness, factual consistency, and target-language correctness.The sample-level score is in [0, 1]; for multiple acceptable references, the maximum score is used.Source: §E.2 · p. 40
  • Reasoning VQA supports multiple valid references, accepting a prediction that matches any enumerated answer while leaving single-answer evaluation unchanged.This addresses cases with multiple semantically acceptable answers.Source: §E.2 · p. 40

E.3 Comparability and Reproducibility

OCRBench v2.1 results are not directly comparable to original OCRBench v2 leaderboard scores because annotations and metrics changed. The evaluation is designed to be deterministic and reproducible, with an LLM-based translation evaluator as an exception to fixed offline metrics.

E.3 Comparability and Reproducibility

  • OCRBench v2.1 scores should not be directly compared with original OCRBench v2 leaderboard results because the protocol changes annotations and metrics for several subtasks.Source: §E.3 · p. 40
  • The v2.1 evaluation pipeline is designed to be deterministic and reproducible, with metrics computed offline using fixed procedures except for LLM-based translation evaluation.Source: §E.3 · p. 40
  • The translation judge uses Gemini-3.5-Flash-Lite with a fixed prompt shown in Figure 6.The figure identifies the prompt’s {predict} and {answer} placeholders as the model output and one reference translation, respectively.Source: §E.3 · p. 41, Figure 6 · p. 41

F Task Examples

The task examples cover diverse visual-text inputs and outputs, from reading and locating text to parsing structured content and answering visual questions.

F Task Examples

  • Extraction examples range from identifying fields in a Chinese sample form to matching nutrition values and extracting handwritten answers.A document-type prompt also asks for classification among listed categories.Source: F Task Examples · pp. 42, 46
  • Examples pair full-page OCR and OCR-VQA with answers that include text or its bounding-box location.One prompt requests all page text; another asks for a book title, while plane-text and answer-extraction examples request bounding boxes.Source: F Task Examples · pp. 42–43
  • Chart and table tasks ask models to convert visual information into nested dictionaries or reconstruct table structure.Other examples request formula transcription and mathematical question answering from an image.Source: F Task Examples · pp. 42, 45, 46
  • Additional examples ask for normalized text-region coordinates and interpretation of ASCII art.The localization prompt specifies coordinates normalized from 0 to 1000 by image dimensions.Source: F Task Examples · pp. 44–45
  • Visual question examples include diagram and reasoning questions, as well as questions about historical events and app messages.Other prompts ask whether a chart value is present and how many times a character appears.Source: F Task Examples · pp. 42, 44, 46
Loading 2609.37712v1…