PaperlayerView exact revision on arXiv

Source-linked AI summary

WildVision: Evaluating Vision-Language Models in the Wild with Human Preferences

Yujie Lu, Dongfu Jiang, Wenhu Chen, William Yang Wang, Yejin Choi, Bill Yuchen Lin

arXiv:2406.11069v1cs.CVcs.AIcs.CL

TL;DR

Current benchmarks and evaluation metrics do not fully capture human preferences for VLMs in real-world multimodal interactions. WildVision-Arena collects crowdsourced comparisons, and WildVision-Bench applies automatic judging to a curated sample. Its scores correlate strongly with Arena Elo ratings, while analysis of real-world chats identifies recurring model challenges.

  • Problem

    Existing benchmarks and reference-based or model-based metrics leave a gap in measuring human preferences across VLMs in real-world multimodal interactions.

    Source: §1 · p. 2

  • Method

    The authors collect human comparisons in WILDVISION-ARENA and curate safe, diversified data for WILDVISION-BENCH, where GPT-4o judges each VLM against Claude-3-Sonnet.

    Source: §1 · p. 2

  • Results

    WILDVISION-BENCH scores achieve a 0.94 Spearman correlation with WILDVISION-ARENA Elo ratings, and chat analysis identifies challenges in visual context, spatial reasoning, imagination, expert knowledge, hallucination, and safety.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Analysis of in-the-wild chats indicates future directions for advancing VLMs, and the authors plan to release chat and voting data for research.

    Source: §6 · p. 9

    The paper presents only some failure cases and plans to release additional cases periodically through its live platform.

    Source: §4 · p. 8

Abstract

from arXiv · show

Recent breakthroughs in vision-language models (VLMs) emphasize the necessity of benchmarking human preferences in real-world multimodal interactions. To address this gap, we launched WildVision-Arena (WV-Arena), an online platform that collects human preferences to evaluate VLMs. We curated WV-Bench by selecting 500 high-quality samples from 8,000 user submissions in WV-Arena. WV-Bench uses GPT-4 as the judge to compare each VLM with Claude-3-Sonnet, achieving a Spearman correlation of 0.94 with the WV-Arena Elo. This significantly outperforms other benchmarks like MMVet, MMMU, and MMStar. Our comprehensive analysis of 20K real-world interactions reveals important insights into the failure cases of top-performing VLMs. For example, we find that although GPT-4V surpasses many other models like Reka-Flash, Opus, and Yi-VL-Plus in simple visual recognition and reasoning tasks, it still faces challenges with subtle contextual cues, spatial reasoning, visual imagination, and expert domain knowledge. Additionally, current VLMs exhibit issues with hallucinations and safety when intentionally provoked. We are releasing our chat and feedback data to further advance research in the field of VLMs.

1 Introduction

Existing benchmarks and metrics do not fully capture human preferences in real-world multimodal interactions. WildVision-Arena and WildVision-Bench provide crowdsourced comparisons, a preference-aligned benchmark, and analysis of in-the-wild model failures.

1 Introduction

  • Existing benchmarks are constrained by simplicity and practicality, while reference-based and model-based metrics leave a gap from human preferences in real-world comparisons.The paper situates this gap in evaluations comparing many models’ capabilities in multimodal interactions.Source: §1 · p. 2
  • 0.94 Spearman correlation links WILDVISION-BENCH scores with WILDVISION-ARENA Elo ratings.The benchmark uses GPT-4o to compare each VLM with Claude-3-Sonnet.Source: §1 · p. 2
  • WILDVISION-BENCH is curated from safe, diversified WILDVISION-ARENA data for automatic evaluation against human preferences.GPT-4o judges each VLM against Claude-3-Sonnet.Source: §1 · p. 2
  • In-the-wild analysis identifies challenges in visual context, spatial reasoning and imagination, expert knowledge, hallucination, and safety.Lower-performing VLMs also struggle with fine visual details under resolution and contextual limitations.Source: §1 · p. 2
  • WILDVISION-ARENA hosts over 20 VLMs and collects 20,000+ multimodal conversations and 8,000+ votes for a live preference leaderboard.Users also provide fine-grained feedback on real-world chats.Source: §1 · p. 2

2 WILDVISION-ARENA: Ranking VLMs with Human Preference

WildVision-Arena ranks VLMs from crowdsourced preferences in multi-round image chats and analyzes user voting patterns and model standings.

2 WILDVISION-ARENA: Ranking VLMs with Human Preference

  • The section presents Arena chat data and analysis of human preferences as the basis for an online VLM leaderboard.Source: §2 · p. 3

2.1 Overview Design of WILDVISION-ARENA

  • Arena users compare models in multi-round chats over uploaded images, vote before model identities are revealed, and contribute to a leaderboard updated every few hours.The platform supports 20+ VLMs and keeps comparison generation hyperparameters the same.Source: §2.1 · p. 3

2.2 Statistics of Chat Data with Votings

  • Voting data center on Recognition, Descriptive, and Analytical questions and Entertainment, Urban, Expert, and People image domains, with growing interest in Interactive, Instructive, and Creative categories.Each voted chat is categorized using GPT-4v.Source: §2.2 · p. 3

2.3 Crowdsourced Human Preference on VLMs in the Wild

  • Across seven models, GPT-4o won 77% of its battles against second-ranked GPT-4V; expert-user agreement averaged 72.5% with Cohen’s Kappa of 0.59.The agreement analysis compared expert annotations with Arena users on 100 samples.Source: §2.3 · pp. 3–4

2.4 Model Ranking with Elo Rating in WILDVISION-ARENA

  • Elo models pairwise win probabilities, updates ratings by outcome minus expectation (α = 400), then uses Bradley–Terry maximum likelihood; ties are split into opposing wins.The paper adopts Bradley–Terry estimation for stable ratings because online Elo computation depends on comparison order.Source: §2.4 · pp. 4–5

2.5 WILDVISION-ARENA Leaderboard

  • By May 29, 2024, GPT-4o led overall and across categories and domains except Natural images; LLaVA-NEXT-34B led open-source models.Table 2 reports ratings and battle counts for 22 models; Figure 5 plots six models across question categories and image domains.Source: Table 2 · p. 4, §2.5 · p. 5, Figure 5 · p. 5

3 WILDVISION-BENCH: In-the-Wild Testbed for VLMs

WildVision-Bench is curated from Arena interactions for fast, real-world VLM evaluation; GPT-4o judging produces rankings closely aligned with Arena preferences.

3.1 Data Curation Pipeline

  • The benchmark’s public set contains 500 samples selected after unsafe-image filtering, deduplication, and diversity sampling; its hidden preview has 50 expert-annotated samples and is updated dynamically.Table 3 includes two expert-annotated cases, including example #61, where all models fall short.Source: §3 · p. 5, §3.1 · p. 6, Table 3 · p. 6

3.2 Automatic Evaluation on WILDVISION-BENCH

  • GPT-4V agrees more with Arena voters when model responses are distinguishable, but tie votes and cases where both models are bad are harder to judge.The paper notes that GPT-4V sometimes also falls short when both models are bad.Source: §3.2 · pp. 6–7, Figure 6 · p. 6
  • WildVision-Bench achieved the highest benchmark correlation with WildVision-Elo, with 0.94 Spearman correlation between its GPT-4o-judged rankings and the Arena leaderboard.GPT-4o compares each model’s answers against Claude-3-Sonnet; MMVet was the runner-up, followed by MMMU and MMStar.Source: §3.2 · p. 7, Figure 7 · p. 7

4 Analysis

The analysis uses real-world multimodal conversations to examine model performance and failure patterns. GPT-4V generally performs better on the validation examples, but compositional reasoning and subtle visual details remain challenging.

4 Analysis

  • In-the-wild conversations combine images and instructions from diverse sources with votes from varied users, offering material for examining real-world scenarios and improvement directions.The authors plan to release the chat and crowdsourced voting data.Source: §4 · p. 8
  • The analysis identifies failures involving insufficient background knowledge for GPT-4V and difficulty discerning subtle details for Gemini-Pro-Vision.The paper presents two documented failure instances and categorizes recurring errors across seven areas, including recognition, reasoning, hallucination, and safety.Source: §4 · p. 8
  • GPT-4V generally outperforms LLaVA-NEXT-34B and Gemini-Pro-Vision on validation examples, yet all three sometimes fail at compositional reasoning.Subtle visual details also challenge less capable models.Source: §4 · p. 8
  • The planned data release includes more than 20,000 multi-turn conversations and 8,000 human votes with reasons, alongside a live leaderboard and failure-case analysis.The authors also plan to open-source the Arena code and provide fast evaluation with WildVision-Bench.Source: §4 · p. 8

5 Related Work

Related work motivates live, in-the-wild VLM benchmarking and human-aligned evaluation. The paper builds on evaluator approaches and Arena-style pairwise comparisons to create a small benchmark aligned with crowd preferences.

5 Related Work

  • Prior work identifies needs for benchmarks that address contamination, robustness, difficulty, and real-world scenarios; the paper builds WILDVISION-ARENA for diversified, difficult, in-the-wild live benchmarking.The passage situates this need alongside benchmark work [101] [32] and research on contamination, robustness, and difficulty.Source: §5 · p. 9
  • Human alignment for open-ended vision-language evaluation has used mapping predictions to predefined choices and larger models as evaluators.Arena methods for LLM evaluation motivate the paper's use of pairwise comparisons against a reference model such as Claude-3-Sonnet.Source: §5 · p. 9
  • WILDVISION-BENCH is presented as a small-scale benchmark for fast pairwise evaluation against a reference model, aligned with crowdsourced preferences from WILDVISION-ARENA.The passage describes this approach as inspired by Arena evaluation methods.Source: §5 · p. 9

6 Conclusion

The conclusion presents WILDVISION-ARENA as a dynamic platform for comparing VLMs and describes sampling safe, diversified data for a benchmark whose automatic evaluation aligns with arena preferences. The appendix shows its interface and examples of tagged user data.

6 Conclusion

  • The authors introduce WILDVISION-ARENA for dynamic comparison across 20-plus VLMs, using 20,000-plus conversations and 8,000-plus votes to derive comparative insights and refine performance continuously.They sample safe, diversified data for WILDVISION-BENCH and apply automatic evaluation aligned with crowdsourced preferences.Source: §6 · p. 9

A User Interface

  • The arena interface lets users submit image-related questions, view multiple model responses side by side, and vote for the response that best fits their criteria.Figure 8 shows the platform interface.Source: A User Interface · p. 17, Figure 8 · p. 17

B Question Category and Image Domain

C Analysis of Failure Cases

The failure-case analysis groups in-the-wild errors across visual recognition, reasoning, spatial imagination, contextual understanding, expert knowledge, hallucination, and safety. Examples range from missed visual details to inappropriate handling of sensitive content.

C Analysis of Failure Cases

  • The analysis presents recurring VLM failure patterns from arena chats and organizes them into categories based on error types.Specific cases appear in Tables 9–13.Source: C Analysis of Failure Cases · p. 21, Table 9 · p. 22
  • Recognition errors include missed artistic or small text, unfamiliar objects or backgrounds, chart interpretation, and reflections.Examples include signs, screenshot text, and a helmet beside a motorcycle.Source: C Analysis of Failure Cases · p. 21
  • Visual reasoning failures include overlooking clothing changes, chess strategy, and reading analog clocks..Source: C Analysis of Failure Cases · p. 21
  • Spatial imagination errors appear in cube-folding tasks and visual riddles that require imagining a folded magazine..Source: C Analysis of Failure Cases · p. 21
  • Contextual understanding failures involve movie and video-game references as well as inappropriate responses to marked images..Source: C Analysis of Failure Cases · p. 21
  • Expert-domain failures include assessing architectural-code compliance and analyzing programming-code details..Source: C Analysis of Failure Cases · p. 21
  • Hallucinations include generating false content from subtle cues or nonexistent details, such as fictitious characters or obscure brand names..Source: C Analysis of Failure Cases · p. 21
  • Safety problems include harmful, biased, or inappropriate content and both excessive censorship and insufficient filtering of NSFW images.NSFW images are filtered out during benchmark curation.Source: C Analysis of Failure Cases · p. 21

D Data Analysis

The analysis reports voting activity and response-length effects in WildVision-Arena, then describes taxonomy and judging procedures used to organize and evaluate its data. Across these appendices, it specifies how questions and image domains are classified and how pairwise preferences are generated.

D Data Analysis

E.1 Taxonomy Annotation

  • GPT-4V annotates each sample with a question category and subcategory plus an image domain and subdomain from the specified taxonomies.Source: §E.1 · p. 28

E.2 VLM Voting

  • GPT-4V acts as an impartial pairwise judge, choosing the better response or labeling both wrong, equally satisfactory, or undecidable.Source: §E.2 · p. 29

E.3 WILDVISION-BENCH Evaluator

  • GPT-4o first generates its own answer, compares both assistants’ responses against it, and gives a five-level verdict from significantly better to significantly worse.Source: §E.3 · p. 30

F.1 Limitations

The authors identify limits in platform coverage and stress testing, alongside risks around model protections and data security. They also note that the crowd-sourced user base may not represent broader global perspectives.

F.1 Limitations

  • The platform omits some recently released models, potentially limiting the breadth of available insights.Source: §F.1 · p. 31
  • Stress testing is inadequate for rising query volumes, and the authors call for stronger security against data leaks while balancing third-party model protection with unbiased responses.Source: §F.1 · p. 31

F.2 Societal Impact

  • The crowd-sourced user base may overrepresent English-speaking users and people interested in multimodal research; interface changes aim to broaden participation.Source: §F.2 · p. 31

G.1 Dataset Documentation and Intended Uses

The paper provides access routes for interacting with the platform and its data. It describes ongoing data and code maintenance and states that the authors accept responsibility for rights violations in data collection.

G.1 Dataset Documentation and Intended Uses

  • Users can submit votes through Hugging Face Vision Arena, while the arena data is available for download and viewing.Source: §G.1 · p. 31

G.2 Maintenance Plan

  • The leaderboard updates every three hours, data will be continually updated, and platform and evaluation code are planned for release with community contributions welcomed.Source: §G.2 · p. 31

G.3 Author Statement

  • The authors accept responsibility for rights violations during data collection and say they will take appropriate action when needed.Source: §G.3 · p. 31
Loading 2406.11069v1…