PaperlayerView exact revision on arXiv

Source-linked AI summary

Mementos: A Comprehensive Benchmark for Multimodal Large Language Model Reasoning over Image Sequences

Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Gedas Bertasius, Mohit Bansal, Huaxiu Yao, Furong Huang

arXiv:2401.10529v2cs.CVcs.AIcs.CLcs.LG

TL;DR

Behavioral hallucinations in image-sequence reasoning lack dedicated study and evaluation metrics. Mementos benchmarks sequential reasoning, and evaluations find that current MLLMs struggle with object and behavior hallucinations, with behavior reasoning weaker than object reasoning.

  • Problem

    Behavioral hallucinations in image-sequence descriptions have received limited study, and existing work lacks a dedicated metric for evaluating them.

    Source: §4.2 · p. 9

  • Method

    Mementos comprises 4,761 image sequences and evaluates nine MLLMs using a GPT-4-assisted procedure that extracts and matches object and behavior keywords.

    Source: §1 · p. 1, §2.2 · p. 5, §5 · p. 9

  • Results

    All tested MLLMs struggle with behavioral and object hallucinations, and perform better on object metrics than behavior metrics.

    Source: §3.1.2 · p. 6, §5 · p. 9

  • Takeaways & Limitations

    The analyses identify interconnected object and behavioral hallucinations, co-occurring behaviors, and cumulative behavioral hallucinations as factors in reasoning failures.

    Source: §1 · p. 3

Abstract

from arXiv · show

Multimodal Large Language Models (MLLMs) have demonstrated proficiency in handling a variety of visual-language tasks. However, current MLLM benchmarks are predominantly designed to evaluate reasoning based on static information about a single image, and the ability of modern MLLMs to extrapolate from image sequences, which is essential for understanding our ever-changing world, has been less investigated. To address this challenge, this paper introduces Mementos, a new benchmark designed to assess MLLMs' sequential image reasoning abilities. Mementos features 4,761 diverse image sequences with varying lengths. We also employ a GPT-4 assisted method to evaluate MLLM reasoning performance. Through a careful evaluation of nine recent MLLMs on Mementos, including GPT-4V and Gemini, we find that they struggle to accurately describe dynamic information about given image sequences, often leading to hallucinations/misrepresentations of objects and their corresponding behaviors. Our quantitative analysis and case studies identify three key factors impacting MLLMs' sequential image reasoning: the correlation between object and behavioral hallucinations, the influence of cooccurring behaviors, and the compounding impact of behavioral hallucinations. Our dataset is available at https://github.com/umd-huang-lab/Mementos.

1 Introduction

Existing image-reasoning benchmarks focus on static, single-image information, leaving time-varying behaviors and events underassessed. Mementos benchmarks sequential-image reasoning and reveals frequent hallucinations and three factors associated with MLLM reasoning failures.

1 Introduction

  • Mementos contains 4,761 varied image sequences paired with human descriptions of primary objects and behaviors, targeting reasoning about positional changes, behavior, and logical connections.The sequences cover everyday life, robotics, and comic-style storyboards.Source: §1 · p. 1
  • The GPT-4-assisted evaluation extracts object and behavior keywords from generated and human descriptions, matches them, and uses domain-specific synonym graphs to improve matching.The authors report that the procedure compares reliably with human evaluation.Source: §1 · pp. 1–2
  • Behavioral hallucinations were more frequent than object hallucinations, indicating a marked weakness in inferring events from image sequences.The evaluation covered nine leading-edge MLLMs; GPT-4V examples show hallucinations across three domains.Source: §1 · p. 2, Figure 1 · p. 2
  • The analyses identify interconnected object and behavior hallucinations, co-occurring behaviors, and the cumulative effect of behavioral hallucinations as factors in reasoning failures.The paper presents quantitative analyses and case studies of these factors.Source: §1 · pp. 2–3

2 Mementos

Mementos evaluates MLLMs on sequential images using human-annotated descriptions of events, objects, and behaviors. Its data spans multiple domains and sequence structures, while its GPT-4-assisted evaluation extracts and matches object and behavior keywords against the annotations.

2 Mementos

  • The section introduces Mementos as a benchmark for testing MLLM reasoning under sequential-image input and describes its data, annotation, distribution, and evaluation protocol.The benchmark’s aim is to assess reasoning over image sequences rather than static scenes.Source: §2 · p. 3

2.1.1 Dataset Composition

  • Mementos comprises 4,761 image sequences, predominantly from Daily-life, Robotics, and Comics, with robotics and comic-style data adding embodied contexts and stylistic and episodic diversity.The section reports category counts in Table 1, but the supplied passage does not give the individual values.Source: §2.1.1 · p. 3, Table 1 · p. 3

2.1.2 Dataset Annotation

  • Each sequence has a ground-truth description of unfolding events, focusing on primary objects and their behaviors; Daily-life descriptions received manual review and cross-validation, while Robotics and Comics were manually annotated.Authors verified annotations across all categories.Source: §2.1.2 · p. 4

2.1.3 Dataset Statistics

  • 67.38% of validation sequences contain 4–14 frames and 31.90% exceed 15; robotics episodes often span 1–3 behaviors, while other data has episode lengths distributed across 4–17.Sequence length counts frames; episode length is estimated from behavior keywords extracted from descriptions.Source: Figure 2 · p. 4, §2.1.3 · p. 4, Figure 3 · p. 4

2.2 Evaluation Procedure and Metrics

  • The evaluation prompts MLLMs with an image sequence, extracts object and behavior keywords from their descriptions, and normalizes matches through synonym graphs before comparison with human annotations.Recall, precision, and F1 are calculated for objects and behaviors; recall reflects episode reasoning, while precision measures hallucination severity.Source: §2.2 · p. 5

3 Experiments

The experiments compare black-box and open-source MLLMs on Mementos and analyze their image-sequence reasoning failures. The results show stronger object than behavior reasoning and identify links between hallucinations, co-occurring actions, and longer sequences.

3.1.1 Models

  • The baseline includes nine black-box and open-source MLLMs; models without sequential-image capability use either a composite-frame c-input or sequential s-input setting, with Video-LLaMA-2 tested only on s-input.Video-LLaMA-2 and Chat-UniVi are designed for sequential images or video; other open-source models receive combined frames as a composite image.Source: §3.1.1 · pp. 5–6

3.1.2 Evaluation results

  • GPT-4V (s-input) leads overall, while MLLMs reason about objects better than behaviors; it ties Gemini (s-input) and LLaVA-1.5 in behavior precision.LLaVA-1.5 is the strongest open-source model, but its behavior inference is weaker than Gemini and GPT-4V.Source: Figure 5 · p. 6, §3.1.2 · p. 6
  • Under c-input, black-box and open-source performance does not differ significantly, and LLaVA-1.5 and mPLUG_Owl-v2 meet or exceed black-box models on many metrics.Source: Figure 5 · p. 6, §3.1.2 · p. 6

3.2 Analysis of Failure Reasoning

  • Object and behavioral hallucinations correlate weakly (0.1–0.4) for most MLLMs; hallucinated objects can cue related false actions, while black-box robotics correlations are negligible.Figure 6 illustrates a Daily-life case in which a hallucinated tennis court is followed by false racket-holding and tennis-playing behaviors.Source: §3.2 · p. 7, Table 3 · p. 8, Figure 6 · p. 8
  • In robotics, co-occurring behaviors can produce false action details despite correct outcomes, potentially degrading MLLM-based reward functions and causing robot behaviors to be mislearned.The example describes an arm grabbing a drawer handle before pulling it open, although the final outcome is accurately described.Source: §3.2 · pp. 7–8
  • In Daily-life sequences, GPT-4V and LLaVA-1.5 object and behavior recall decrease as episode length grows, consistent with hallucinations accumulating across frames.Source: §3.2 · p. 8, Figure 7 · p. 9

4 Related work

Prior MLLM benchmarks largely assess single-image reasoning, while Mementos evaluates behavioral hallucinations through image sequences. The related work identifies behavioral hallucination and its evaluation as underexplored.

4.1 Benchmarking in MLLMs

  • Earlier MLLM benchmarks emphasize single-image reasoning, whereas Mementos uses video-frame or comic sequences and unguided descriptions to assess behavioral hallucinations.The authors distinguish this from uniformly sampled video-frame visual QA, which guides models with questions.Source: §4.1 · p. 9

4.2 Hallucination in MLLMs

  • Behavioral hallucination—generated actions conflicting with image sequences—lacks dedicated study and a dedicated evaluation metric in prior work.The related work notes existing frameworks for mitigating hallucinations, but a research gap remains around behavioral hallucinations.Source: §4.2 · p. 9

5 Conclusion and Future Works

The paper presents Mementos, evaluates nine recent MLLMs, and reports that they struggle with hallucinations when describing image sequences. It identifies three factors in these reasoning failures and proposes directions for future work.

Conclusion

  • All nine evaluated MLLMs struggle with significant behavioral and object hallucinations when describing image sequences.The authors report that quantitative analyses and case studies identify three primary factors contributing to these reasoning failures.Source: §5 · p. 9

Future Works

  • Future dataset work proposes adding first-person navigation, sequential medical CT scans, and interactive gaming data to Mementos.The proposed diversification aims to cover a wider range of contexts and scenarios.Source: §5 · p. 9
  • Future work proposes more semantic evaluation, plus targeted strategies to reduce behavioral and object hallucinations.The authors suggest moving beyond predominant reliance on keyword matching and developing strategies informed by the three identified failure factors.Source: §5 · pp. 9–10

A Details of Open X-Embodiment Data Selection

The appendix directs readers to Table 4 for the selected Open X-Embodiment subsets and their video sampling numbers.

A Details of Open X-Embodiment Data Selection

B Human Evaluation

The human-evaluation comparison uses 200 randomly selected validation sequences and four representative MLLMs. GPT-4-assisted evaluation generally overestimates performance, but produces similar model rankings, with most metric differences between 1% and 4%.

B Human Evaluation

  • The comparison uses 200 randomly selected Val sequences and four top-performing models: GPT-4V, Gemini, Chat-UniVi, and LLaVA-1.5.Human evaluators manually extracted object and behavior keywords for AI-generated and human-annotated descriptions; six metrics were compared.Source: B Human Evaluation · p. 13
  • GPT-4-assisted evaluation extracts object and behavior keywords, with domain-specific synonym graphs refining keyword matching.The paper reports comparing these automated evaluation results with human evaluation.Source: B Human Evaluation · p. 13
  • Most GPT-4-assisted versus human-evaluation metric differences fall between 1% and 4%.GPT-4-assisted evaluation tends to overestimate performance, while relative rankings among models remain essentially unchanged.Source: B Human Evaluation · p. 13

C Prompt Details

This appendix section presents prompts used in the paper and includes a human-evaluation table. The passage identifies a prompt for merging questions and answers from Daily-life videos, but does not specify the complete prompt task.

C Prompt Details

  • The section presents the prompts used in the paper, including one for merging questions and answers from Daily-life videos.The passage cuts off before stating the full scope of that prompt.Source: C Prompt Details · p. 13
  • The available prompt description does not reveal the full wording or remaining prompt contents.The section text ends mid-sentence.Source: C Prompt Details · p. 13
  • Table 5 is identified as a human-evaluation table.Its caption does not provide evaluation procedures or results.Source: Table 5 · p. 14

D Case Study

The case study presents failure-reasoning examples across Daily-life, Robotics, and Comics domains. Captions attribute failures to object and behavioral hallucinations, co-occurring behaviors, and Snowball effects.

D Case Study

  • Across Figures 8–23, the paper presents failure-reasoning cases from different domains, with captions detailing specific failure reasons; shown examples include Daily-life and Robotics cases.Examples cite object hallucination and its correlation with behavioral hallucination, co-occurrence behavior, and Snowball.Source: D Case Study · p. 14, Figure 8 · p. 16, Figure 9 · p. 16
  • In Figure 12, after a behavioral hallucination, the model stops describing before the child pushes the box back along the hallway.The caption describes this omission as another outcome of the Snowball effect.Source: Figure 12 · p. 18
  • In Figure 14, the model misses the arm's move toward the disc after it lowers, then bases later description on its initial interpretation.The caption presents this as a Snowball effect; Figure 15 also describes an assumed cooking episode cutting off the pot-moving action.Source: Figure 14 · p. 20, Figure 15 · p. 21
Loading 2401.10529v2…