PaperlayerView exact revision on arXiv

Source-linked AI summary

Fine-Grained Multi Image Object Hallucination Benchmark

Joonki Min, Chaeyun Kim, Hyungwook Choi, Yejin Kim, Kihyun Kim, Yohan Jo, Joonseok Lee

arXiv:2608.30653v1cs.CVcs.AIcs.LG

TL;DR

Object hallucination remains difficult to diagnose in multi-image MLLMs because existing benchmarks are largely single-image or high-level. MIOH provides a fine-grained benchmark spanning object-centric tasks, reasoning patterns, and adversarial pressures, and its evaluation finds integration-stage limitations across 29 models.

  • Problem

    Existing benchmarks do not systematically diagnose how visual complexity and multi-image reasoning demands trigger object hallucination.

  • Method

    MIOH crosses four object-centric tasks with three multi-image reasoning patterns under three controlled adversarial pressures.

  • Results

    Across 29 models, including GPT-5 and Gemini-2.5-Pro, MIOH reveals integration-stage breakdowns in maintaining object representations across multiple images.

  • Takeaways & Limitations

    MIOH provides a controlled framework for analyzing multi-image object hallucination and targeting improvements in multi-image understanding.

Abstract

from arXiv · show

Multimodal Large Language Models (MLLMs) are increasingly deployed in multi-image scenarios requiring complex reasoning across visual contexts. However, current MLLMs remain fundamentally limited by object hallucination-generating plausible yet factually inconsistent descriptions about objects. Existing benchmarks, designed primarily for single-image settings or providing only high-level multi-image assessments, cannot systematically diagnose how visual complexity and reasoning demands trigger hallucination. To address this gap, we introduce MIOH, a fine-grained multi-image object hallucination benchmark that systematically evaluates object hallucination across four foundational tasks (existence, counting, attribute, position) through three multi-image reasoning patterns (comprehensive, comparative, selective) under three controlled adversarial pressures (visual context scale, perceptual difficulty, contextual bias). Through evaluation of 29 models, we reveal that even state-of-the-art systems like GPT-5 and Gemini-2.5-Pro exhibit distinct failure patterns across different reasoning patterns and tasks. Our evaluation reveals that hallucination stems not merely from perceptual failures but from integration-stage limitations when maintaining object representations across multiple images. MIOH provides a controlled framework for analyzing multi-image object hallucination and serves as a critical evaluation tool for developing more reliable multimodal AI systems.

1. Introduction

MIOH addresses the limited diagnosis of object hallucination in multi-image MLLMs by combining fine-grained object-centric tasks, reasoning patterns, and controlled visual challenges.

  • Multi-image MLLMs must recognize individual-image content and integrate information across diverse visual inputs.
  • Object hallucination produces plausible but factually inconsistent descriptions about queried objects.
  • Existing hallucination benchmarks mainly use single-image binary questions focused on existence and counting.
  • General multi-image benchmarks assess overall reasoning but lack controlled manipulation of visual factors for hallucination diagnosis.
  • MIOH crosses existence, counting, attribute, and position tasks with comprehensive, comparative, and selective reasoning patterns.
  • MIOH varies visual context scale, perceptual difficulty, and contextual bias to analyze hallucination under visual complexity and reasoning demands.

2. Related Work

Prior work studies multi-image reasoning and object hallucination separately, leaving fine-grained multi-image hallucination insufficiently covered.

  • Early MLLMs faced cross-image reasoning challenges from limitations in visual token processing and inter-image semantic modeling.
  • Recent multi-image systems use large-scale image-text training data to support multi-image understanding.
  • Object hallucination has been linked to training-data biases, vision-encoder limitations, modality-alignment deficits, and weak context attention.
  • Existing general multi-image benchmarks evaluate reasoning capabilities but do not specifically target object hallucination.
  • Specialized hallucination benchmarks predominantly use single-image binary or captioning settings, with limited attributes and spatial relations.

3. Overall Design of our MIOH Benchmark

MIOH organizes multi-image hallucination evaluation around object-centric tasks, reasoning patterns, and adversarial pressures, with representative examples and validated benchmark data.

  • Overall Design: MIOH crosses four object-centric tasks with three reasoning patterns under varying adversarial conditions for fine-grained diagnosis.
  • Multi-Image Reasoning Patterns: Comprehensive reasoning aggregates information across all images into a holistic judgment.
  • Multi-Image Reasoning Patterns: Comparative reasoning identifies differences between specific images while maintaining separate scene representations.
  • Multi-Image Reasoning Patterns: Selective reasoning retrieves the image matching criteria while filtering irrelevant information from candidate images.
  • Diagnostic Value: The three patterns diagnose distinct failures, including aggregation, comparison, retrieval, localization, and object-recognition bottlenecks.
  • Object-Centric Tasks: The benchmark evaluates existence, counting, attribute, and position, covering presence, enumeration, property binding, and spatial-relation binding.
  • Adversarial Pressures: Adversarial factors test integration capacity, feature extraction for small or occluded objects, and susceptibility to misleading co-occurrence priors.
  • Implementation Details: The final benchmark contains 3,484 questions across 11,732 images, validated by three independent annotators.

4. Benchmark Results and Discussion

Across 29 models, MIOH shows that multi-image hallucination varies by reasoning pattern, task, and adversarial pressure, with integration across images emerging as a major failure source.

  • Overall Performance: 29 models achieve only 36.1% overall average accuracy; Gemini-2.5 Pro reaches 64.4% and GPT-5 63.1%.Top open-source results include Qwen2-VL-7B at 49.1% and MiniCPM-V-2.6 at 48.5%.
  • Reasoning Patterns: Performance varies significantly across Comprehensive, Comparison, and Selection reasoning patterns, beyond the narrow coverage of existing benchmarks.The benchmark evaluates these patterns across existence, counting, attribute, and position tasks.
  • Reasoning Patterns: Selection averages 34.2%, while Attribute under Selection reaches only 26.7%, reflecting failures to bind object-attribute pairs to the correct image.Models may recognize that an object-attribute pair exists in the image set but fail to localize it to a specific image.
  • Task-Specific Vulnerability: Existence is the strongest task at 49.4%, whereas Counting falls to 25.4%, with Attribute at 32.4% and Position at 37.3%.Counting failures involve both aggregating counts across images and locating an image with a specified number of objects.
  • Impact of Adversarial Pressures: Adversarial constraints reduce performance across tasks: Attribute drops from 37.5% to 31.3% under HP, and Position from 46.9% to 35.2% under HN.Counting’s marginal decline is attributed to a floor effect because baseline accuracy is already near random chance.
  • Impact of Adversarial Pressures: With eight images, Existence falls to 30.0%, a relative drop exceeding 50% from the easy baseline, despite stronger robustness from GPT-5 and Gemini 2.5 Pro.Their substantial performance drops show that scaling multi-image reasoning remains a critical challenge.
  • Accuracy-Robustness Trade-off: Higher baseline accuracy does not guarantee adversarial robustness: Qwen2-VL-7B and Qwen2.5-VL-7B show robustness drops of -27.2% and -21.8%.Model size correlates moderately with easy-question performance but shows virtually no correlation with robustness.
  • Multi-Image Context as a Hallucination Amplifier: Single-image decomposed evaluation consistently outperforms simultaneous multi-image analysis across models and scales on Existence.The controlled comparison uses identical visual content and indicates that cross-image integration amplifies hallucination beyond simple error accumulation.

5. Conclusion

MIOH evaluates multi-image object hallucination across object-centric tasks, reasoning patterns, and adversarial pressures, revealing integration-stage limitations in current MLLMs.

  • MIOH evaluates object hallucination across four object-centric tasks, three reasoning patterns, and three adversarial pressures.
  • Evaluation of 29 models, including GPT-5 and Gemini-2.5-Pro, reveals fundamental limitations in multi-image integration.
  • Hallucination stems not merely from perceptual failures but from integration-stage breakdowns when models maintain object representations across images.
  • MIOH provides a foundation for targeted improvements and a tool for evaluating more reliable multimodal AI systems.

A. Detailed Benchmark Construction

The benchmark combines curated data, structured metadata, rule-based multi-image question generation, adversarial examples, and validation to produce a balanced evaluation set.

  • Data Sources: COCO-ReM, PACO, and SVG address annotation gaps in masks, attributes, and spatial relations needed for reliable benchmark construction.SVG uses filtering, region-mask verification, and high relation density to support position-based questions.
  • A.2. Metadata Construction: Metadata organizes objects by task-specific properties, difficulty levels, and image identifiers for efficient question retrieval.Filtering uses bounding-box size, occlusion thresholds, and image resolution; difficulty classes distinguish clear, hard-positive, and hard-negative cases.
  • A.3. Question Generation: Rule-based templates generate comprehensive, comparative, and selective multiple-choice questions across four core tasks.For counting, templates sum object counts across selected images and include incorrect answers plus “None of the above.”
  • A.4. Adversarial Benchmark Design: Perceptually difficult positives are selected using object size, mask area, occlusion, and image position, while misleading negatives exploit co-occurrence patterns.These strategies target objects that are difficult to perceive or contextually suggested despite being absent.
  • A.4. Adversarial Benchmark Design: Difficulty increases as the proportion of hard-positive or hard-negative examples approaches all N images in a question.
  • A.4.4. Quality Assurance and Validation: After generating over 26,000 questions and manual review by three annotators, approximately 20,000 verified questions remained.Review covered ground-truth correctness, clarity, answer-option validity, and annotation consistency; persistent ambiguities were removed.
  • A.4.4. Quality Assurance and Validation: Validation still found systemic noise in position and counting questions, including ambiguous spatial relations, missed instances, and miscounts.
  • Final Benchmark: Stratified sampling produced a final benchmark of 3,484 questions across 11,732 images, balanced by tasks, reasoning patterns, and adversarial pressures.

B.1. Variation Analysis

Model variability differs across tasks, reasoning patterns, and adversarial pressures, with spatial and selective reasoning varying most while counting and large contexts remain consistently difficult.

  • Reasoning Patterns: Selection has the highest reasoning-pattern variance (σ = 0.153), indicating differing model ability to identify a specific target.
  • Adversarial Pressures: Increasing Number of Images has the lowest adversarial-pressure variance (σ = 0.079), comparable to Counting.The passage associates this low variance with uniformly poor performance under extreme context pressure.

B.2. Cross-dimensional Interaction Analysis

Cross-dimensional analysis shows that context scale and combined perceptual pressures drive substantial degradation, while robustness and model-category gaps depend on the evaluated dimension.

  • Degradation Analysis: Increasing Number of Images causes catastrophic degradation across dimensions, especially for Comparison reasoning and Existence tasks.
  • Degradation Analysis: Combined HP+HN pressure produces larger performance drops than either pressure alone, indicating additive difficulty.
  • Degradation Analysis: Comprehensive reasoning is more robust under adversarial pressure, while Counting degradation reflects floor effects because its Easy baseline is already low.
  • Model Comparison: Commercial models have their largest performance advantages over open-source models in Counting and Comparison tasks.

C.1. Qualitative Examples from MIOH benchmark

MIOH examples evaluate object-centric capabilities across four tasks, three question types, and increasing visual difficulty. The examples pair task-specific reasoning demands with pressures such as scale, occlusion, and contextual bias.

  • Task coverage: MIOH examples cover existence, counting, position, and attribute tasks across multifaceted multi-image evaluation.The examples test object verification, counting, spatial relations, and property binding.
  • Existence Tasks: Existence questions use selective retrieval to identify which image contains a target object, including difficult cases such as a bench in a rainy scene.
  • Counting Tasks: Counting questions use comprehensive reasoning to aggregate object occurrences across images, including heavily occluded or cluttered objects.Examples include counting elephants across four frames and sandwiches distributed across difficult visual scenes.
  • Position and Attribute Tasks: Position questions use selective retrieval to test whether specified spatial relations hold, while attribute questions use comparative reasoning to bind properties to objects.Examples include dog–cat proximity, person–umbrella proximity, and subtle mug-color distinctions.
  • Reasoning patterns and difficulty: The benchmark combines comprehensive, comparative, and selective questions while varying visual scale, occlusion, and contextual bias.This progression supports multifaceted evaluation from collective understanding to image-specific retrieval and cross-image differences.

C.2. Failure Case Examples of Commercial Models

Failure examples show GPT-5 and Gemini-2.5-Pro hallucinating under perceptual and multi-image reasoning pressures despite strong baseline performance. Errors include missed ambiguous objects, fine-grained category confusion, and incorrect forced selections when abstention is available.

  • Overview: GPT-5 and Gemini-2.5-Pro exhibit hallucination patterns under visual adversarial pressures despite strong baseline performance.The reported cases concern failures to ground visual evidence correctly within multi-image contexts.
  • Attribute-Comprehensive: Both models fail to verify dark gray bowls across all target images in an Attribute-Comprehensive task involving ambiguity or occlusion.The example is characterized as perceptual blindness or aggregation failure in comprehensive verification.
  • Counting-Comparative: In Counting-Comparative questions, models confuse wine glasses with similar water goblets or tumblers, producing incorrect frequency comparisons.The error reflects weakness in fine-grained classification under perceptual similarity.
  • Position-Selective: Position-Selective questions with a None of the above option trigger hallucinated object identities or violations of the requested spatial relation.Examples include identifying a zebra as a cow and selecting a dog–cow pair with incorrect positioning.
  • Benchmark examples: MIOH benchmark examples span existence, counting, position, and attribute questions with answer choices that include abstention options.The cited examples include bench and keyboard existence, elephant counting, and sandwich counting prompts.
Loading 2608.30653v1…