PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Blind Compliance: Benchmarking Task Verification in OCR Reasoning

Yue Zhou, Yuan Wu, Yi Chang

arXiv:2609.00232v1cs.CV

TL;DR

OCR benchmarks often assume every task is valid and answerable, despite real-world failures involving unavailable evidence, contradictions, or missing variables. This paper introduces a human-verified benchmark and task-verification framework for deciding executability before answering, and finds blind compliance, diagnosis gaps, and prompt-induced over-refusal across evaluated systems.

  • Problem

    Existing OCR evaluations largely assume valid, answerable tasks, leaving models insufficiently evaluated on whether visual evidence, textual premises, and questions jointly define executable tasks.

  • Method

    The paper introduces OCR-grounded Task Verification and VeriOCRBench, a 1,800-sample benchmark with trap-injected invalid tasks, trap-free controls, VAF-grounded construction, and human auditing.

  • Results

    Evaluations of 15 MLLMs and two OCR-text baselines reveal blind compliance, diagnosis failures, and prompt-induced over-refusal; assisted prompting improves detection but can increase rejection of valid controls.

  • Takeaways & Limitations

    Robust OCR reasoning requires premise verification and grounded refusal in addition to perception and answer generation.

  • Takeaways & Limitations

    The benchmark is limited to single-turn OCR reasoning and eight invalidity types, excludes interactive and tool-augmented settings, and relies on semantic judgment for open-ended responses.

Abstract

from arXiv · show

Multimodal Large Language Models (MLLMs) have achieved strong performance on OCR-centric document understanding and text-rich visual reasoning benchmarks. Yet existing evaluations largely assume that every task is valid and answerable. In real-world OCR scenarios, this assumption often fails: questions may rely on illegible text, occluded evidence, nonexistent visual targets, contradictory premises, or missing variables. We study this reliability gap as OCR-grounded Task Verification: before answering, a model should determine whether the Image Premise (IP), Textual Premise (TP), and Question (Q) jointly define an executable task. We introduce VeriOCRBench, a 1,800-sample human-verified benchmark built from source images drawn from 8 OCR-related datasets and spanning 8 real-world image domains, with controlled, image-grounded diagnostic tasks. It contains 1,600 trap-injected invalid tasks across 8 trap types and four verification dimensions---Visual, Contextual, Factual, and Logical---plus 200 trap-free controls for measuring over-refusal. Built with a Visual Atomic Fact (VAF)-anchored pipeline and full human auditing, VeriOCRBench enables decoupled evaluation of task verification, root-cause diagnosis, and over-refusal. Evaluating 15 leading MLLMs reveals persistent blind compliance, diagnosis failures, and prompt-induced over-refusal, exposing a critical reliability gap in current OCR reasoning systems. The code is available at: https://github.com/zy001122/Beyond-Blind-Compliance.

1 Introduction

Existing OCR benchmarks generally assume tasks are valid and answerable, but real-world inputs can contain unavailable evidence, contradictions, nonexistent targets, or missing variables. VeriOCRBench addresses this gap by evaluating whether models verify task executability before answering, diagnose invalidity, and avoid rejecting valid tasks.

  • Existing OCR-centric benchmarks largely assume that image evidence is present, textual premises are consistent, and questions are well-posed.
  • Real-world OCR tasks may involve illegible or occluded evidence, nonexistent targets, incompatible contexts, contradictions, or missing variables.
  • Task Verification represents each task as T = (IP, TP, Q) and requires deciding whether it is executable before answering or issuing a grounded refusal.
  • VeriOCRBench contains 1,800 human-verified samples, including 1,600 trap-injected invalid tasks and 200 trap-free controls across 8 domains, 8 trap types, and four verification dimensions.
  • A VAF-anchored pipeline and full human auditing support visually grounded, plausible traps and benchmark quality.
  • Evaluations reveal blind compliance, diagnosis failures, and prompt-induced over-refusal, while milder verification prompts can reduce over-refusal for representative models.

2 Related Work

Prior text-rich visual benchmarks evaluate extraction, understanding, and reasoning, but the supplied related-work passages emphasize vulnerabilities when visual evidence is missing, contradicted, or insufficient.

  • Earlier benchmarks mainly assess document, scene-text, infographic, and chart understanding through extraction, information location, and text-grounded question answering.
  • More recent OCR-centric benchmarks extend evaluation toward broader OCR capabilities and reasoning over visually rich inputs.
  • OCR reasoning remains vulnerable when readable text, numbers, tables, or layout relations are unavailable, contradicted, or insufficient for solving the question.
  • Blind execution can produce confident but unsupported answers, while over-cautious behavior creates a separate reliability concern.

3 VeriOCRBench

VeriOCRBench defines executable OCR tasks through image, textual, and question premises, constructs invalid and valid controls across a structured trap taxonomy, and evaluates verification with complementary metrics and prompting settings.

  • 3.1 Definition of Task Verification: A text-rich visual reasoning task is represented as T = (IP, TP, Q), combining image evidence, user-provided context, and a reasoning question.
  • 3.1 Definition of Task Verification: A task is executable only when required evidence is available, IP is consistent with TP and Q, and their joint information suffices to answer Q.
  • 3.1 Definition of Task Verification: Invalid tasks require grounded refusal identifying unavailable evidence, contradictions, nonexistent targets, vague references, logical inconsistency, or missing variables.
  • 3.2 Taxonomy of Task Verification Traps: The benchmark organizes 8 trap types into Visual, Contextual, Factual, and Logical verification dimensions, including degradation, occlusion, domain mismatch, hallucinated targets, fact tampering, vague references, paradoxes, and missing variables.
  • 3.3 Benchmark Construction: Trap-free controls use zero-text-leakage tasks to measure over-refusal separately from invalid-task verification.
  • 3.3 Benchmark Construction: The dataset contains 1,800 human-verified samples from 8 OCR-related benchmarks and 8 image domains, including 1,600 invalid tasks and 200 controls.
  • 3.3 Benchmark Construction: The VAF-anchored pipeline extracts OCR evidence and bounding boxes, then uses those anchors for grounded task construction before human auditing.
  • 3.4 Evaluation Metrics: Evaluation reports Task Verification Rate, Diagnosis Accuracy, and Over-Refusal Rate under proactive and assisted prompting paradigms.

4 Experiments

Experiments evaluate 15 MLLMs and two OCR+LLM baselines on task verification, diagnosis, and over-refusal using VeriOCRBench. Results show pervasive blind compliance, a detection–diagnosis gap, and limited verification from OCR-only pipelines.

  • Experimental Setup: 15 MLLMs and 2 OCR+LLM baselines are evaluated on task verification, root-cause diagnosis, and over-refusal.Table 2 reports PTVR/ATVR, p-DA/a-DA, and p-ORR/a-ORR across trap-injected samples and trap-free controls.
  • Main Results: 15.38% PTVR for GPT-5 and 50.06% for Claude-4.5-Sonnet show pervasive blind compliance under proactive prompting.GPT-4o reaches 21.06%, most open-weight models remain below 30%, and PaddleOCR-VL-1.5 reaches 0.75%.
  • Main Results: 84.44% ATVR for GPT-5 and 96.25% for Gemini-3-Flash-NoThinking show that assisted prompting substantially improves invalid-task detection.The gains can coincide with high assisted over-refusal, including 79.50% for Gemini-3.1-Pro and 60.00% for Gemini-3-Flash-Thinking.
  • Main Results: 61.50% ATVR but 7.25% a-DA for Gemini-3.1-Pro demonstrates that detecting invalidity does not guarantee correct root-cause diagnosis.GPT-5, Claude-4.5-Sonnet, and Qwen3.5-397B-A17B achieve a-DA scores of 76.94%, 76.12%, and 80.50%, respectively.
  • Baseline Analysis: 70.06% ATVR but 53.56% a-DA for PP-OCRv5 + DeepSeek-R1-Distill-Qwen-32B shows that OCR+LLM pipelines provide partial but incomplete grounded verification.Text-only pipelines cannot fully capture visual degradation, occlusion, layout, or missing visual evidence.

5 Analysis & Discussion

Fine-grained analyses show that verification difficulty depends strongly on trap type and invalidity dimension, while aggregate detection can obscure diagnostic errors and over-refusal. Prompt ablations further show that milder verification instructions can reduce unnecessary refusal.

  • 5.1 Fine-Grained Trap Results: Trap difficulty is heterogeneous: models handle Domain Mismatch and Hallucinatory Target better than Occluded Target, Vague Reference, and Missing Variable.Localized and under-specified traps require identifying missing evidence, distinguishing visual anchors, or recognizing absent variables.
  • 5.1 Fine-Grained Trap Results: Contextual verification is comparatively strong, whereas Visual, Factual, and Logical verification remain uneven.Performance depends on where invalidity arises, including missing visual evidence, factual mismatch, or logical underspecification.
  • 5.2 Diagnosis and Over-Refusal: 54.25 points is Gemini-3.1-Pro’s diagnosis gap between 61.50% ATVR and 7.25% a-DA under assisted prompting.A reliable verifier should combine a small ATVR–a-DA gap with low a-ORR.
  • Milder Verification Instructions: 19.50% a-ORR for Gemini-3.1-Pro and 15.00% for Gemini-3-Flash-Thinking follow replacement of the original assisted prompt with a milder verify-if-needed prompt.The milder prompt consistently reduces over-refusal across all five evaluated models, while diagnosis changes are model-dependent.
  • Milder Verification Instructions: Effective verification prompts must balance invalid-task detection, correct rejection diagnosis, and avoidance of unnecessary refusal on valid inputs.The analysis characterizes stronger verification prompts as not always better because they can reduce blind compliance while inflating over-refusal.

6 Conclusion

The paper frames OCR-grounded Task Verification as deciding whether an OCR reasoning task should be answered before execution. VeriOCRBench evaluations reveal blind compliance, prompt-induced over-refusal, and a gap between invalid-task detection and root-cause diagnosis.

  • Task Verification asks whether an MLLM should answer an OCR reasoning task before executing it.
  • VeriOCRBench is a 1,800-sample human-verified benchmark covering eight trap types across four verification dimensions, with trap-free controls for over-refusal evaluation.
  • Evaluating 15 MLLMs and two OCR-text pipeline baselines reveals blind compliance under proactive prompting and prompt-induced over-refusal under assisted prompting.
  • The results show a gap between invalid-task detection and correct root-cause diagnosis.
  • Robust OCR reasoning requires premise verification and grounded refusal alongside perception and answer generation.

Limitations

VeriOCRBench is a controlled benchmark whose scope is limited to single-turn OCR reasoning and eight representative invalidity types across four verification dimensions. Its open-ended responses require semantic judgment, although validation and robustness checks support the evaluation.

  • The benchmark covers single-turn OCR reasoning and eight representative invalidity types across four verification dimensions.
  • It does not cover interactive settings, tool-augmented workflows, dynamically updated documents, or all domain-specific verification failures.
  • Because model responses are open-ended, evaluation relies on semantic judgment rather than exact-match scoring.
  • Structured judge outputs, human validation, and multi-judge robustness checks are used, but some nuanced cases may require finer-grained human evaluation.

A Formal Definition and Boundary Conditions

The paper represents an OCR reasoning task as an image premise, textual premise, and question, then defines validity through evidence availability, cross-modal consistency, and query solvability. VeriOCRBench operationalizes invalidity with eight trap types across four dimensions and includes valid controls to measure over-refusal.

  • A.1 Task Representation: A task is represented as T = (IP, TP, Q), combining an image premise, textual premise, and reasoning question.IP contains visual evidence; TP supplies assumptions or context; Q asks for an outcome based on IP and TP.
  • A.2 Task Validity Function: Task validity requires available visual evidence, cross-modal consistency, and sufficient joint information to answer the question.If any condition is violated, the task is invalid or unexecutable.
  • A.2 Task Validity Function: An ideal model answers valid tasks and produces a grounded refusal or critique for invalid tasks.Responses that follow a flawed premise without flagging the issue do not count as successful verification.
  • B Trap Taxonomy and Expected Model Behaviors: VeriOCRBench organizes eight trap types into four verification dimensions: Visual, Contextual, Factual, and Logical.Figure 5 assigns two trap types to each dimension.
  • B.1 Visual Dimension: Visual traps test unavailable or unreadable evidence, including image degradation and occluded targets.Models should avoid guessing degraded or hidden content.
  • B.2 Contextual Dimension / B.3 Factual Dimension: Contextual and factual traps test compatibility with image content, nonexistent targets, tampered attributes, and ambiguous references.Expected responses identify mismatches, absent targets, false values, or under-specified targets.
  • B.4 Logical Dimension: Logical traps involve contradictions or missing dependencies that prevent uniquely solvable reasoning.These include logical paradoxes and missing variables such as units, currency, scale, jurisdiction, or conversion basis.
  • B.5 Trap-Free Controls / C Representative Trap Examples: Trap-free controls are valid and answerable, allowing Over-Refusal Rate measurement alongside representative examples of invalid tasks.The controls prevent high verification scores from being achieved through generic refusal.

D Dataset Construction Details

VeriOCRBench is constructed through trap-specific generation grounded in visual atomic facts, followed by human auditing and structured evaluation of detection, diagnosis, and over-refusal.

  • Construction Pipeline: The pipeline constructs invalid tasks and trap-free controls with VAF grounding, zero text leakage, zero trap disclosure, and reasoning beyond direct extraction.Zero text leakage preserves dependence on image evidence rather than revealing answers in the textual premise or question.
  • Trap Taxonomy: Eight trap types span visual, contextual, factual, and logical failures, including degraded evidence, domain mismatch, fact tampering, paradoxes, and missing variables.Trap-free controls provide valid tasks for measuring over-refusal and preventing generic refusal policies from being rewarded.
  • Human Auditing: 1,800 samples underwent full human auditing against six criteria covering VAF accuracy, solvability, trap unsolvability, disclosure, alignment, and reasoning beyond extraction.Auditors verified that controls remained answerable while trap tasks contained clear invalidity without revealing the trap.
  • Dataset Composition: The benchmark contains 1,600 trap-injected invalid tasks and 200 trap-free controls, with Table 4 summarizing scale, taxonomy, domains, and evaluation setting.The dataset statistics are explicitly summarized in Table 4.
  • Evaluation Metrics: TVR measures invalid-task detection, while DA credits a sample only when the model both detects invalidity and identifies the correct trap mechanism.The assessment uses structured GPT-4o judging because exact matching is insufficient for open-ended responses; separate judges assess traps and over-refusal.

G Multi-Judge Robustness Analysis

The study tests judge robustness with multiple LLM judges and validates GPT-4o judgments against blinded human annotations. Agreement is consistently high, though not perfect.

  • Multi-Judge Analysis: 450 response-level units were stratified across nine task categories, with 50 samples per category and an even proactive–assisted split.The analysis covered all 15 evaluated MLLMs and excluded two OCR-text pipeline baselines.
  • Multi-Judge Analysis: Three judges—GPT-4o, Claude, and Gemini—were compared on rejection judgments and diagnostic correctness using unanimous agreement, Fleiss’ κ, and Krippendorff’s α.Diagnostic correctness was evaluated on the 400 trap-injected responses within the 450 sampled units.
  • Human Validation: Human evaluators applied the same response-level criteria as the automatic judge and were blinded to GPT-4o labels before adjudication.The criteria separately recorded premise rejection or correction and rationale alignment with the ground-truth trap logic.
  • Human Validation: 85.66% agreement (κ = 0.709) was achieved between GPT-4o and human labels for rejection judgments, while diagnostic correctness reached 83.40% (κ = 0.652).The comparisons used 286 total responses for rejection and 253 trap-injected responses for diagnostic correctness.

I.2 Paired Significance Tests

The analysis uses paired statistical tests and confidence intervals to distinguish supported trends from small, non-significant differences. Assisted prompting increases over-refusal across nearly all systems.

  • Significant Comparisons: Claude-4.5-Sonnet’s proactive verification advantage over GPT-4o and GPT-5 remains significant after Holm correction.Qwen3.5-397B-A17B also significantly improves assisted diagnosis over Qwen3.5-122B-A10B.
  • Statistical Testing: Wilson 95% confidence intervals and paired exact McNemar tests with Holm correction support the main metric trends and comparisons.Paired tests are appropriate because models are evaluated on the same benchmark instances.
  • Non-Significant Comparisons: 6.00 and 5.00 percentage-point ORR differences fall below corresponding 80%-power minimum detectable effects of 8.59 and 5.65 points.These sensitivity results motivate cautious interpretation of the two non-significant cross-system assisted-ORR comparisons.
  • Within-System ORR: Assisted ORR is higher than proactive ORR for all 17 systems, with 16 of 17 paired differences remaining significant after Holm correction.The only non-significant case is PaddleOCR-VL-1.5, whose ORR rises from 53.00% to 58.00%.

K.1 Fine-Grained Trap-Type Analysis

Verification difficulty varies sharply by trap type, and successful rejection does not guarantee grounded diagnosis. Assisted prompting generally helps, but can also disrupt diagnosis for some models.

  • Cross-Category Variation: Verification difficulty is highly trap-dependent across trap types, verification dimensions, and image domains.Figures 7, 8, and 3 provide proactive diagnosis, task-verification, and assisted-diagnosis breakdowns.
  • Trap-Type Difficulty: Assisted macro a-DA reaches 80.6% for Hallucinatory Target and 78.5% for Domain Mismatch, but falls to 29.0%, 41.4%, and 43.2% for Vague Reference, Occluded Target, and Missing Variable.The benchmark therefore tests subtle ambiguity, unavailable evidence, and missing-information failures beyond obvious false premises.
  • Detection Versus Diagnosis: ATVR is much higher than a-DA for several trap types, especially Occluded Target and Missing Variable, indicating detection without correct mechanism identification.Assisted prompting can increase invalid-task detection without consistently producing correct diagnostic explanations.
  • Prompt Effects: Most models improve with explicit verification instructions, whereas Gemini-3.1-Pro shows negative diagnosis shifts on several trap types.The result indicates that stronger verification instructions do not uniformly improve grounded diagnosis.

K.2 Fine-Grained Verification Dimension Analysis

Diagnosis accuracy varies across the four verification dimensions: Contextual verification performs best overall, while Factual and Logical verification are more difficult. Assisted prompting does not uniformly improve every dimension, and representative cases show both better contradiction detection and residual diagnosis and over-refusal failures.

  • Verification dimensions: Contextual verification achieves the highest macro performance overall, while Factual and Logical verification remain more challenging for many models.The reported pattern is associated with stronger difficulty in fine-grained grounding, ambiguity resolution, and logical consistency checking.
  • Verification dimensions: Contextual strength aligns with Domain Mismatch and Hallucinatory Target traps involving global incompatibilities between image and textual premise.
  • Assisted prompting: Assisted prompting does not uniformly improve diagnosis across models and verification dimensions; some systems gain substantially, while others improve less or degrade.The ∆DA analysis supports uneven effects of explicit verification instructions on grounded root-cause diagnosis.
  • Representative failures: GPT-4o detects a Fact Tampering contradiction under assisted prompting after following the falsified premise under proactive prompting.
  • Representative failures: GPT-5 rejects a Hallucinatory Target under assisted prompting but misdiagnoses the nonexistent target as illegible text.
  • Representative failures: Gemini-3.1-Pro answers a trap-free control proactively but rejects the same valid task with assisted prompting, illustrating over-refusal.
Loading 2609.00232v1…