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

AI summary

Overview

Research area: Multimodal large language model (LLM) agents for content moderation and online meme analysis.

Technical level: Intermediate. The paper assumes familiarity with vision-language models, retrieval-augmented pipelines, and agent orchestration frameworks, but the system design itself is described modularly and readably.

Scope: The paper defines a new task called meme provenance — generating a structured analytical report about a meme's meaning, origin, spread, cultural context, psychology, and safety risks — and introduces MemeTrace, a multimodal LLM agent that produces such reports, evaluated on a new 100-meme dataset.

What This Paper Is About

Existing meme moderation research mostly treats each image in isolation, asking a closed-set question such as "hateful vs. benign," which gives little insight into why a meme is harmful and generalizes poorly to newly emerging memes that depend on fresh news and cultural context. The authors instead formalize meme provenance analysis: given a single meme image, produce a structured report covering 11 fields (About, Structure, Origin, Spread, Culture, Usage, Target Audience, Emotional Impact, Cognitive Load, Persuasive Technique, and Safety Flag). To do this, they build MemeTrace, an agent that gathers and verifies external evidence before writing the report, and they evaluate it against a GPT-4o vision-language baseline using both expert human ratings and automatic report-quality metrics.

Key Contributions

1. A new task and report schema. The paper introduces meme provenance as a research task and defines an 11-field provenance report format, grouped into semantics, spreading dynamics, cultural context, psychological influence, and safety assessment. Three fields (about, origin, spread) are inspired by KnowYourMeme encyclopedia entry structure; the rest extend it.

2. The MemeTrace agent. A multimodal LLM-based agent using local database retrieval, reverse image search, and targeted web search to synthesize structured provenance reports, coordinated through a Think–Act–Evaluate loop and a confidence evaluator that routes between retrieval tools.

3. The Meme100 dataset. A curated set of 100 real-world memes collected from Reddit, Twitter/X, 4chan, Facebook, TikTok, YouTube, and Instagram using an event-driven strategy, mixing classic templates from KnowYourMeme with newly emerged memes tied to high-impact events (e.g., the 2024 U.S. election, the Russia–Ukraine conflict). MemeTrace significantly outperforms the VLM baseline on this set.

4. Automatic report-quality metrics. A suite of metrics computed directly from report text (numerical specificity, entity richness, certainty and context) that correlate strongly with human scores, enabling scalable evaluation.

Main Findings

• Human evaluation on Meme100: MemeTrace raises informativeness from 4.050 to 4.838, correctness from 4.205 to 4.843, and the combined adjusted score from 4.128 to 4.841 — an overall improvement of 17%. All three gaps are significant at p < .001.

• Per-field dominance: MemeTrace achieves the highest scores across all 11 evaluation fields compared with the baseline, with particularly strong performance in about, origin, and spread. Safety_flag informativeness scores 4.777 vs. 4.483 and correctness 4.670 vs. 4.400.

• Automatic metrics favor MemeTrace: It uses 2.088 numbers, 0.880 year mentions, and 0.380 explicit dates per 100 words, versus 0.631, 0.168, and 0.020 for the baseline. Hedge expressions drop from 1.42 to 0.64 per 100 words, and platform mentions drop from 1.162 to 0.653 per 100 words. Entity coverage rises for persons (3.22 vs. 1.96), organizations (2.72 vs. 1.49), GPEs (1.83 vs. 1.29), events (0.11 vs. 0.01), locations (0.11 vs. 0.05), and facilities (0.13 vs. 0.02).

• Metrics track human judgment: The strongest correlations between automatic metrics and human scores reach Pearson r = 0.78 and Spearman ρ = 0.94.

• Ablation on a 20-meme subset: The full system scores an adjusted 4.853. Removing Tavily causes the largest drop to 4.227, followed by RIS at 4.317 and LocalDB at 4.388. This indicates open-web evidence matters most.

• Backbone transferability: With the same agent settings, Gemini 1.5 Pro reaches 4.451 and Qwen2.5-VL-72B reaches 4.575, both well above the VLM-only baseline of 3.961 but below default GPT-4o-based MemeTrace. The authors do not rank Gemini against Qwen because n = 20 cannot separate two closely performing backbones.

• Robustness to visual perturbations: Adjusted scores are 4.247 under Gaussian blur (σ = 2.5), 4.297 under UI overlay (≈12% top coverage), and 4.229 under border cropping (15%) — a decrease of roughly 0.56–0.62, about 12% relative to the full system. UI overlay causes the smallest drop; Gaussian blur primarily harms correctness; border cropping reduces both informativeness and correctness.

• Safety-flag patterns: Disinformation is the most frequent risk, appearing in 44% of reports, followed by sensitive content at 16%. Political and geopolitical memes show the highest disinformation rate (≈50%), while technology-related memes are rarely flagged as sensitive. A co-occurrence matrix shows strong coupling between disinformation and fake misattribution.

• MemeDB scale: The local database holds 21,305 confirmed KnowYourMeme entries as of June 2025.

Methodology in Plain English

Instead of asking a vision-language model to write a report in one shot, the authors built an agent that works in a loop of thinking, acting, and evaluating. Given a meme image, it first reads the overlaid text with the EasyOCR library and then uses a multimodal LLM to clean up that raw output, removing noise such as watermarks.

Next, the agent checks a local database called MemeDB, which holds 21,305 confirmed memes from KnowYourMeme. Each meme's image and metadata are encoded with CLIP into a shared embedding space, concatenated, and indexed with FAISS. The agent retrieves the top-3 similar candidates, and GPT-4o then scores their semantic relevance from 0 to 10, keeping only candidates above a threshold of 8 so that visually similar but semantically unrelated matches are filtered out.

A confidence evaluator acts as the system's brain. After each tool call, it judges whether the collected evidence is sufficient and trustworthy across the 11 required fields. If not, it triggers a reverse image search branch, which uses Google RIS, ranks candidate URLs with a domain-based scoring algorithm, crawls only the top-5, and passes them to a RIS evidence checker that uses GPT-4o to produce structured summaries with inline citations and usefulness scores from 0 to 1. Those scores feed back into the domain rankings, so the system adapts to source quality over time.

If evidence is still incomplete, the agent moves to the Tavily open-web search branch, which runs in two modes: an Element-Extraction Mode for broad exploratory search over salient visual and textual elements (especially proper nouns), and an Evaluator-Query Mode for targeted completion of specific gaps. Tavily returns up to five results with snippets and source links, and a quality checker summarizes them with citations and scores from 0 to 1. The evaluator refines queries iteratively until evidence is sufficient or a maximum iteration limit (default 5) is reached, at which point a report synthesis module uses GPT-4o to assemble the 11-field report.

For evaluation, the authors compared against a GPT-4o VLM baseline that uses the same report-generation prompt but no external evidence, plus a stronger search-augmented baseline (GPT-4o with native web search). Temperature was set to zero for reproducibility. Three in-domain experts rated every report field on informativeness and correctness using a five-point Likert scale, producing 7,920 rating items per annotator and 23,760 total human-assigned scores across the three annotators, with ICC ranging from 0.502 to 0.600. A separate blinded re-evaluation by two external annotators who are not co-authors covered the search-augmented baseline as well. The system was implemented with the LangGraph framework and run on a MacBook Pro (M4 Max, 36 GB RAM).

Why This Matters

Impact on research: The paper reframes meme analysis from closed-set classification to open-ended provenance reporting, which requires dynamic external knowledge rather than static model weights. It also offers a dataset and a validated set of automatic metrics for a task where human evaluation is expensive.

Real-world applications:

• Content moderation: Platform moderators can get contextual and cultural background about a meme that traditional classifiers cannot supply, helping them judge manipulative or misleading content.
• Dataset annotation: The structured multi-perspective reports turn raw memes into annotated research artifacts that support downstream studies of meme evolution and diffusion.
• Misinformation triage: Because disinformation is the most frequently flagged risk (44% of reports) and co-occurs strongly with fake misattribution, the reports can surface misreferenced sources.
• Safety risk surfacing: By examining a meme across multiple dimensions at once, the agent can reveal layers of risk that a single safety label would hide.

Industry relevance: Social platforms, trust-and-safety teams, and fact-checking organizations have a direct interest in interpretable, evidence-grounded meme assessments that degrade gracefully under the image distortions common in reshared screenshots — precisely the perturbations this paper tests. The authors note the system requires multiple rounds of LLM calls and online retrievals, with an efficiency, latency, and cost analysis in the appendix.

Future Directions

1. Expand the knowledge base. MemeDB is limited in size and does not comprehensively cover meme template diversity, especially newly emerging or non-English variants — a constraint the authors list as a primary limitation.

2. Address cost and latency. The agent depends on multiple rounds of LLM calls and online retrieval, which the authors identify as relatively expensive; making its efficiency and deployment cost practical at scale remains open.

3. Conduct systematic red-teaming. Current evaluation covers natural perturbations only. Untested vulnerabilities include source pollution from search-engine-optimized or low-quality websites and maliciously injected or templated content hidden in webpages. Domain scoring mitigates but does not prevent information pollution.

4. Improve automatic metrics. The authors acknowledge their metrics are imperfect: for example, the frequency of dates or numeric mentions reflects information density rather than factual accuracy.

Target Audience

Researchers and practitioners in multimodal LLM agents, content moderation, and computational social science, particularly those working on hateful meme detection, misinformation, or provenance and lineage analysis of online media. It is also useful for trust-and-safety engineers evaluating whether agentic retrieval pipelines can outperform single-shot vision-language models on context-dependent tasks, and for dataset builders interested in event-driven collection strategies and human-validated automatic evaluation metrics.

Authors’ abstract

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.