PaperlayerView exact revision on arXiv

Source-linked AI summary

MemeTrace: A Multimodal LLM-Based Agent for Safety-Aware Meme Provenance Analysis

Bo Shao, Yihan Ma, Yiting Qu, Michael Backes, Yang Zhang

arXiv:2610.19267v1cs.AIcs.CLcs.CY

TL;DR

Memes can convey implicit harmful meanings that depend on cultural and current context, while common moderation studies focus on image-level classification rather than provenance. MemeTrace generates structured provenance reports using multimodal reasoning and external evidence retrieval; on Meme100, it outperforms a VLM-only baseline and remains robust under visual perturbations.

  • Problem

    Memes can carry unsafe content whose implicit meaning depends on cultural context, while classification-focused studies lack broader provenance analysis.

    Source: §1 · p. 1

  • Method

    MemeTrace uses a multimodal LLM agent and evidence-driven reasoning to generate structured provenance reports from meme images.

    Source: §1 · p. 2, §3.2 · p. 3

  • Results

    MemeTrace substantially outperforms the VLM-only baseline on Meme100, with higher human-evaluation scores and strong robustness under visual perturbations.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Meme provenance frames meme analysis as structured reconstruction of origin, spread, and cultural context rather than classification alone.

    Source: §1 · p. 2, §6 · p. 8

    MemeTrace’s MemeDB is limited in size and does not comprehensively cover meme-template diversity, especially newly emerging or non-English variants.

    Source: Limitations · p. 8

Abstract

from arXiv · show

Online memes are often carriers of unsafe content such as disinformation and propaganda. Their rapid evolution in response to news events and strong dependence on contextual information make it difficult for platforms and moderators to detect and moderate them effectively. Moreover, existing moderation studies typically classify memes as safe or unsafe based on individual images, lacking the ability to perform broader provenance analysis, i.e., where a meme originates, how it spreads, what cultural symbolism it represents, whether it poses potential risks to society, etc. To bridge this gap, we introduce MemeTrace, a multimodal LLM-based agent that automatically traces the provenance and potential risks of emerging memes. We evaluate MemeTrace on Meme100, a dataset of 100 real-world memes, using both human annotation and automatic metrics. The results demonstrate that MemeTrace substantially outperforms the VLM baseline. More importantly, MemeTrace achieves higher accuracy in detecting unsafe memes, especially those involving disinformation and sensitive content, and remains robust under visual perturbations.

1 Introduction

Memes’ implicit meanings and dependence on current cultural context make safety assessment difficult, while classification-focused approaches provide limited provenance analysis. MemeTrace addresses this with structured, evidence-driven reports and outperforms a VLM-only baseline on real-world memes.

Motivation

  • Memes combine visual and textual cues with cultural references and implicit intent, making harmful meanings difficult to interpret from an image alone.Humor, irony, and satire can mask harmful intent, and rich background knowledge may be needed to assess it.Source: §1 · p. 1

Task and approach

  • The paper introduces meme provenance as a task that produces structured reports across 11 fields, from origin and spread to cultural context and safety flags.MemeTrace uses a multimodal LLM agent to analyze a single image through evidence-driven, step-by-step reasoning rather than one-shot VLM inference.Source: §1 · p. 2

Evaluation

  • The evaluation uses Meme100, a dataset of 100 real-world memes, and proposes automatic report-quality metrics that correlate highly with human scores.The metrics are intended to support scalable evaluation of generated reports.Source: §1 · p. 2

Results

  • MemeTrace substantially outperforms the VLM-only baseline in interpreting real-world memes, including implicit hateful content, and demonstrates robustness under common image perturbations.The evaluation also finds strong backbone transferability, suggesting that gains come from the agent design rather than a specific VLM.Source: §1 · p. 2

2 Related Work & Problem Statement

Prior meme research commonly studies narrow classification tasks or individual analytical dimensions, often without real-world evidence and broader online context. The paper frames meme provenance as a structured report spanning semantic, social, psychological, and safety dimensions.

Related Work

  • Earlier meme studies often classify content as hateful, offensive, or harmful using multimodal features, focusing on narrow, single-dimensional tasks.Other work examines genealogy, evolution, or sociopolitical ideologies, but still treats one analytical dimension at a time.Source: §2 · p. 2
  • These approaches rely on static models and lack real-world evidence or broader online context, limiting their analysis of memes’ backgrounds.The related-work discussion also notes that newer LLM-agent approaches remain focused on single-meme classification rather than systematic background modeling.Source: §2 · pp. 2–3

Problem Statement

  • The problem is to generate an interpretable report for a target meme image across 11 fields covering meaning, spread, culture, psychological influence, and safety.The listed fields include About, Structure, Origin, Spread, Culture, Usage, Target Audience, Emotional Impact, Cognitive Load, Persuasive Technique, and Safety Flag.Source: §2 · p. 3

3 The MemeTrace Agent

MemeTrace combines multimodal perception, external retrieval, evidence assessment, and report synthesis in a confidence-routed reasoning loop. Its retrieval branches gather and assess context before the system produces a structured provenance report.

3.1 Design Rationale

  • MemeTrace is designed to gather current online information, interpret diverse sources, and verify their reliability before synthesizing a provenance report.This design responds to the limits of static VLMs, which cannot incorporate up-to-date knowledge.Source: §3.1 · p. 3

3.2 System Overview

  • The agent takes a single meme image and coordinates perception, retrieval, and reasoning through a closed-loop Think–Act–Evaluate process.Its modules support end-to-end provenance analysis rather than one-shot VLM report generation.Source: §3.2 · p. 3

3.3.1 Image Loading & OCR Extraction

  • The OCR module extracts overlaid English text and uses a multimodal LLM to correct errors and remove irrelevant output such as watermarks.The cleaned text supplies clues to downstream provenance analysis.Source: §3.3.1 · p. 4

3.3.2 Local Database Retrieval

  • Local database retrieval searches MemeDB for the top three similar meme candidates using image and metadata representations indexed for similarity search.MemeDB contains more than 21,000 confirmed memes from KnowYourMeme.Source: §3.3.2 · p. 4

3.3.3 Reverse Image Search (RIS) Branch

  • Reverse image search retrieves related webpages, prioritizes domains with a scoring algorithm, and has an evidence checker summarize and score crawled sources.Usefulness scores dynamically update domain rankings.Source: §3.3.3 · p. 4

3.3.4 Open-Web Search (Tavily) Branch

  • Tavily search uses broad exploratory queries when reverse image search is insufficient and targeted queries when evidence is partial, uncertain, or inconsistent.Its quality scores and summaries guide subsequent queries and help avoid low-quality search directions.Source: §3.3.4 · p. 4

3.3.5 Confidence Evaluator & Routing Policy

  • After each tool call, the confidence evaluator checks evidence sufficiency and conflict, routing the agent to further retrieval or report synthesis.When needed, it proceeds from reverse image search to exploratory Tavily search and then iterative targeted queries until evidence is sufficient or the iteration limit is reached.Source: §3.3.5 · pp. 4–5

3.3.6 Report Synthesis

  • Once evidence is sufficient or the iteration limit is reached, GPT-4o integrates the collected evidence into a structured provenance report.The report follows the schema defined in Section 2.Source: §3.3.6 · p. 5

4 Experimental Setup

The experiments use a diverse real-world meme dataset, compare MemeTrace with GPT-4o-based baselines, and assess report quality through expert ratings and report-derived automatic metrics.

4.1 Evaluation Datasets

  • Meme100 contains 100 classic and newly emerged memes collected across social platforms, while derived subsets support ablation, backbone-transfer, and robustness tests.Robustness subsets apply Gaussian blur, border cropping, and UI overlays at levels chosen to preserve human recognizability while affecting OCR, visual matching, and retrieval.Source: §4.1 · p. 5

4.2 Baseline

  • The GPT-4o VLM baseline generates reports from the meme image without external evidence, while a second baseline uses GPT-4o’s built-in text-based web search.Both baselines use the same report-generation prompt and temperature zero.Source: §4.2 · p. 5

4.3 Human Evaluation

  • Three in-domain experts independently rate each report field for informativeness and correctness on a five-point Likert scale.Ratings are averaged across annotators, and two external annotators conduct a blinded re-evaluation.Source: §4.3 · pp. 5–6

4.4 Automatic Report Quality Metrics

  • Automatic report-quality metrics measure numerical specificity, named-entity richness, and certainty and context using features computed from report text.The metrics include counts of numerals, dates, named entities, hedge expressions, and platform mentions.Source: §4.4 · p. 6

5 Results

On Meme100, MemeTrace earns higher human ratings than the VLM-only baseline, with results also covering report-quality metrics, module ablations, visual robustness, and safety assessment.

5.1 Main Results

  • MemeTrace scores highest in every evaluation field, particularly about, origin, and spread, and its reports use more specific timelines and contextual entities.Numbers per 100 words: 2.088 vs. 0.631; year mentions: 0.880 vs. 0.168; explicit dates: 0.380 vs. 0.020. Platform mentions and hedging decrease as claims become more selective and evidence-grounded.Source: §5.1 · pp. 6–7, Table 1 · p. 7
  • MemeTrace’s adjusted score rose from 4.128 to 4.841 (+17%) over the VLM-only baseline, with informativeness and correctness gaps significant (p < .001).Informativeness increased from 4.050 to 4.838, and correctness from 4.205 to 4.843.Source: §5.1 · p. 6, Table 2 · p. 7
  • Automatic metrics correlate with human judgments, reaching Pearson r = 0.78 and Spearman ρ = 0.94.Source: §5.1 · p. 7

5.2 Ablation Studies

  • The full system scores 4.853; removing Tavily drops the adjusted score to 4.227, the largest decline, while removing RIS or LocalDB yields 4.317 or 4.388.Tavily contributes most to information-intensive fields such as about and origin; disabling RIS weakens information about meme propagation.Source: §5.2 · p. 7

5.3 Robustness Results

  • Visual perturbations lower adjusted scores by about 0.56–0.62, roughly 12%, while about, origin, and structure remain relatively stable.UI overlays cause the smallest drop; blur primarily reduces correctness, while border cropping reduces both informativeness and correctness.Source: §5.3 · pp. 7–8, Table 4 · p. 8

5.4 Safety-Flag Analysis

  • For safety_flag, MemeTrace exceeds the VLM-only baseline in informativeness (4.777 vs. 4.483) and correctness (4.670 vs. 4.400); disinformation appears in 44% of reports.Sensitive content appears in 16% of reports; political and geopolitical memes show the highest disinformation rate (≈50%).Source: §5.4 · p. 8

6 Conclusion

The paper frames meme provenance as a new task and presents MemeTrace as a system for structured, multi-perspective analysis. On Meme100, it reports stronger human ratings than the VLM-only baseline, aligned automatic metrics, and robustness under visual perturbations.

6 Conclusion

  • Meme provenance is introduced as a task for reconstructing memes’ origin, spread, and cultural context through structured reports.Source: §6 · p. 8
  • On Meme100, MemeTrace achieves substantially higher human informativeness and correctness scores than the VLM-only baseline.Source: §6 · p. 8
  • Automatic report-quality metrics align strongly with human evaluations, and the Tavily, RIS, and LocalDB modules make complementary contributions, with Tavily contributing most.Source: §6 · p. 8
  • The authors describe structured reports as useful research artifacts for studying meme evolution and diffusion, and as contextual support for platform moderators assessing harmful content.Source: §6 · p. 8

Limitations

The authors identify constraints in dataset coverage, system cost, robustness testing, and automatic report-quality metrics.

Limitations

  • MemeDB is limited in size and does not comprehensively cover meme-template diversity, especially newly emerging or non-English variants.Source: Limitations · p. 8
  • MemeTrace requires multiple LLM calls and online retrievals, resulting in relatively high costs.Source: Limitations · p. 8
  • Evaluation focuses mainly on natural perturbations, and systematic red-teaming tests have not yet been performed.The authors note potential source pollution from search-optimized or low-quality sites and maliciously injected or templated webpage content; domain scoring cannot fully prevent it.Source: Limitations · pp. 8–9
  • Automatic report metrics remain imperfect: date and numeric-mention frequencies indicate information density, not factual accuracy.Source: Limitations · p. 9

Ethical Considerations

The study relies on publicly accessible data and reports safeguards for platform compliance, privacy, and human evaluation.

Ethical Considerations

  • Data collection uses publicly accessible platform data, and the agent follows platform terms of service, robots.txt, and rate limits.Source: Ethical Considerations · p. 9
  • Because collected evidence is public, the authors state that the study introduces no additional privacy risks.Source: Ethical Considerations · p. 9
  • Human evaluation used content warnings and an option to skip, without crowd workers or sensitive annotator data.Logs may include offensive material and are restricted to offline research use rather than online dissemination.Source: Ethical Considerations · p. 9

LLM Usage Considerations

The appendices describe MemeTrace’s report fields and retrieval components, including how stored memes and external sources are selected, assessed, and summarized. The manuscript also states that authors inspected all LLM-assisted outputs.

LLM Usage Considerations

A Meme Provenance Report Fields

  • The 11 report fields cover semantics, spreading dynamics, cultural context, psychological influence, and safety.Three fields are inspired by KnowYourMeme entries; the others extend the format to culture, practical uses, and potential risks.Source: A Meme Provenance Report Fields · p. 11, Table 5 · p. 12

B MemeDB Coverage and Technical Details

  • MemeDB stores structured metadata for confirmed KnowYourMeme entries; CLIP retrieval returns three candidates, and GPT-4o retains those scoring above 8/10 for provenance reasoning.The database is based on 21,305 entries manually reviewed by KnowYourMeme editors; Figure 5 depicts its coverage.Source: B MemeDB Coverage and Technical Details · p. 11, Figure 5 · p. 11

C Domain Scoring Algorithm Details

  • The domain-scoring algorithm ranks Google reverse-image-search URLs, and only the top five are crawled.It uses a confidence-bound lower bound with a decaying exploration bonus to balance prioritizing high-quality domains with giving new domains exposure.Source: C Domain Scoring Algorithm Details · pp. 11–12

D RIS Evidence Checker Details

  • The RIS evidence checker cleans noisy webpage text, creates cited summaries, and scores each source’s usefulness from 0 to 1.Those scores update domain-quality records, allowing rankings to adapt as source quality changes.Source: D RIS Evidence Checker Details · p. 12, Figure 6 · p. 13

E Tavily Quality Checker Details

  • The Tavily quality checker filters noise, preserves relevant details, and returns cited summaries with usefulness scores from 0 to 1.The evaluator uses the summary and score to guide subsequent actions.Source: E Tavily Quality Checker Details · p. 13

F MemeTrace Implementation Details

MemeTrace uses fixed implementation settings for its model comparisons, while Meme100 combines event-driven memes with older classic templates. The dataset contains 100 memes collected from public platforms and Know Your Meme.

F MemeTrace Implementation Details

  • Model comparisons replace GPT-4o with Gemini-Pro-1.5 or Qwen2.5-VL-72B-Instruct while keeping other settings fixed; the main experiments use temperature zero and a five-iteration limit.The system is implemented in LangGraph, and execution logs are saved as Markdown files to enable auditing.Source: F MemeTrace Implementation Details · p. 13, Figure 7 · p. 13

G Meme100 Dataset Construction

H Alignment Between Automatic and Human Evaluation

The paper examines whether automatic report metrics align with human evaluations and reports annotator agreement. Entity-based metrics align most strongly, while the main experiment shows low rating variance and moderate ICC reliability.

H Alignment Between Automatic and Human Evaluation

  • Entity-based metrics, especially unique named-entity counts, align most strongly with human scores, while temporal and numerical measures show moderate positive correlations.Hedging frequency and platform enumeration are negatively correlated with human evaluation; Figure 9 displays Pearson and Spearman correlations.Source: H Alignment Between Automatic and… · p. 14
  • The authors conclude that automatic metrics can support scalable assessment of report quality.Source: H Alignment Between Automatic and… · p. 14

I Inter-Annotator Agreement

J External Blinded Re-Evaluation and Search-Augmented Baseline

In an external blinded comparison, MemeTrace outperforms both the VLM-only and search-augmented baselines on the 20-meme subset. The appendices also report qualitative cases, safety-flag visualizations, retrieval confidence, execution profiles, and costs.

J External Blinded Re-Evaluation and Search-Augmented Baseline

  • MemeTrace wins on 18 of 20 memes; the search-augmented baseline scores 4.594 versus 3.964 for the plain baseline.External annotators rescored three systems with identifiers removed and report order randomized; the authors attribute MemeTrace’s gains to specialized retrieval modules rather than generic web search alone.Source: J External Blinded Re-Evaluation and… · pp. 14–15, Table 8 · p. 15

K Per-Field Human Evaluation Results

L Qualitative Case Studies

  • Across three cases, MemeTrace provides richer, more contextually grounded analyses than the VLM baseline, retaining key details under moderate UI overlay interference.The cases cover a Florida traffic-accident meme, a penguin-tariff satire, and a Last Supper parody; Figures 10–12 show the comparisons.Source: L Qualitative Case Studies · p. 15, Figure 10 · p. 17, Figure 11 · p. 18

M Safety-Flag Analysis

N Trace-Log Visualizations

O.1 Execution Profiles

  • MemeTrace averages 18.46 processing steps per meme, including 11.54 multimodal LLM calls, 3.92 external search calls, and approximately three local operations.The majority of latency and API cost arises from LLM-based iterative reasoning.Source: §O.1 · p. 16

O.2 Cost and Latency Analysis

  • Processing averages $0.1259 and approximately 111 seconds per meme, versus tens of minutes for manual expert analysis.The authors note that this latency is higher than that of lightweight baselines.Source: §O.2 · p. 16

O.3 Practical Usage Scenarios

  • The authors propose MemeTrace as a selective fallback for challenging or ambiguous cases when lightweight models or rule-based systems lack confidence.Source: §O.3 · p. 16
Loading 2610.19267v1…