From Shallow Humor to Metaphor: Towards Label-Free Harmful Meme Detection via LMM Agent Self-Improvement

AI summary

Overview

Research area: Multimodal content moderation and computer vision, specifically harmful meme detection using Large Multimodal Models (LMMs).

Technical level: Intermediate. The paper assumes familiarity with vision-language models, prompting, and contrastive learning, but its central ideas are conceptually accessible.

One-sentence scope: The paper introduces ALARM, a framework that detects harmful memes without any labeled training data by having an LMM agent mine detection knowledge from easy-to-judge memes and apply it to harder, subtler ones.

What This Paper Is About

Harmful memes are increasingly used to spread discrimination online, but existing detectors require large volumes of manually labeled examples and go stale quickly as new meme formats and topics emerge. The paper asks whether a model can learn to spot subtle, metaphorical, or implicit hate speech by studying only the obvious, "shallow" cases it can already judge confidently — with zero human annotations, no fine-tuning, and no weight updates. ALARM answers yes, and in some settings outperforms fully supervised methods.

Key Contributions

1. The first label-free harmful meme detection framework. ALARM requires no annotated data, no gradient-based training, and no model weight optimization, allowing it to adapt immediately to newly emerging harmful memes by drawing references from the unlabeled data itself.

2. A Confidence-based Explicit Meme Identification mechanism. The framework uses the LMM's own prediction probabilities to separate easy, explicit memes from subtle ones and automatically assign pseudo-labels to the high-confidence segment, verified by a preliminary experiment showing confidence correlates positively with accuracy.

3. A Pairwise Learning Guided Agent Self-Improvement paradigm. Explicit memes are reorganized into contrastive harmful/benign pairs via multimodal retrieval, and an LMM agent distills fine-grained case experiences from them, then iteratively purifies those experiences into a compact set of generalized detection references using four atomic operations (ADD, UPVOTE, DOWNVOTE, EDIT).

4. Strong empirical validation across three datasets and two LMM backbones. ALARM surpasses label-driven baselines on FHM and ToxiCN, matches or exceeds them on MAMI, and improves its Qwen2.5-VL-72B and GPT-4o backbones by an average of 6.35 percentage points in accuracy.

Main Findings

• Label-free can beat label-driven. ALARM achieves 75.80/75.79 (Accuracy/Macro-F1) on FHM and 79.21/72.51 on ToxiCN with Qwen2.5-VL-72B, edging out the best supervised baselines such as ExplainHM (75.60 on FHM) and MR.HARM (75.40 on FHM). On MAMI it reaches 85.50/85.50 with GPT-4o, the highest score in the comparison table.

• Large gains over the underlying model. Enabling ALARM on a raw LMM backbone yields an average accuracy improvement of 6.35%, showing the self-improvement loop adds real capability rather than just reformatting prompts.

• Confidence is a reliable proxy for correctness. The preliminary experiment across ten probability thresholds on FHM and MAMI shows that as the confidence threshold rises, sample counts drop but prediction accuracy rises, justifying the use of high-confidence memes as pseudo-labeled training material with minimal noise.

• Few-shot methods lag behind. Few-shot baselines such as OPT-30B, OpenFlamingo-9B, and even the prior agent-based LoReHM remain below label-driven performance, because they draw only on a handful of annotated demonstrations and ignore the wealth of unlabeled memes.

• Every component matters. Ablation results show removing reference refinement causes the largest drop (FHM accuracy falls from 75.80 to 69.60), followed by removing confidence-based selection (to 70.78), experience gathering (to 72.00), and retrieval-based pairing (to 72.40). This confirms that the pipeline's value comes from the whole chain, not any single step.

• Random selection of "easy" memes fails. Replacing confidence-based selection with random selection degrades performance substantially, because random picks include subtle memes that receive incorrect pseudo-labels and inject noise downstream.

• Minimal tuning required. The confidence selection ratio τ is fixed at 0.2 and the reference set capacity L at 15 across all three datasets, and experiments run on a single NVIDIA L40s GPU.

Methodology in Plain English

The pipeline has three stages.

First, every unlabeled meme is passed to an LMM with a simple prompt asking whether it is harmful or benign. The model's output probabilities are used to rank the memes. The top 20% by confidence are assumed to be "explicit" — obvious cases the model handles reliably — and are given pseudo-labels based on whichever category scored higher. The rest are treated as the subtle set needing help.

Second, the pseudo-labeled explicit memes are split into harmful and benign groups. Using CLIP text and image embeddings, each pseudo-harmful meme is matched with its most semantically similar pseudo-benign counterpart, producing contrastive pairs that differ in harmfulness but look alike. This is the key trick: similar-looking pairs force the agent to articulate what actually makes something harmful.

Third, an LMM agent examines each pair using two-step chain-of-thought reasoning — first describing both memes, then explaining what differentiates them. These case notes accumulate into a pool of experiences. The agent then reads through the experiences one by one and maintains a running reference list of at most 15 high-level detection rules, adding, editing, upvoting, or downvoting entries depending on how much each new case supports them. This converts thousands of scattered examples into a short, portable set of guidelines.

At inference, the subtle memes are classified by the same LMM agent, now supplied with the self-generated reference list as guidance. Nothing is trained, and no gradients are computed — the references act as a kind of "verbal gradient" that reshapes the agent's behavior.

Why This Matters

This work challenges a core assumption in content moderation research: that effective detection requires large annotated datasets. It reframes the problem as one of knowledge distillation from the model's own confident predictions, and shows that this is competitive with — sometimes better than — supervised training. It matters particularly because harmful content changes faster than annotation pipelines can keep up, and because labeling hate speech is expensive, requires domain expertise, and can expose annotators to distressing material.

Real-world applications:

• Social media moderation at scale. Platforms could deploy ALARM-style detection on new trending topics without waiting for a labeled dataset, using the model's own confident judgments as a starting point.
• Non-English and low-resource markets. The strong results on the Chinese ToxiCN dataset suggest the approach transfers to languages where labeled moderation data is scarce.
• Rapid response to breaking events. New hate campaigns often erupt around news events; ALARM can derive references from the same batch of memes it needs to classify, without a training cycle.
• Ad tech and brand safety. Advertisers need to avoid placing content next to harmful material, and a label-free detector can be pointed at new inventory streams immediately.
• Research on agentic self-improvement. The reference-refinement mechanism is domain-general and could transfer to other subjective classification tasks such as misinformation or toxicity in text.

Industry relevance: the approach targets trust-and-safety teams at platforms and vendors, where labeling budgets, annotation latency, and content drift are constant operational pains. The ability to run on open-weight models like Qwen2.5-VL-72B also matters for organizations that cannot send moderation data to closed APIs.

Future Directions

• Verifying the distilled references. The paper does not report human evaluation of the self-generated references, so it is unclear whether they are interpretable, stable, or auditable enough for regulated moderation settings.
• Reducing backbone dependence. The confidence-based selection step still requires the open-source Qwen2.5-VL-72B model, limiting use in environments that rely only on closed APIs or smaller local models.
• Robustness under adversarial evasion. As detection improves, harmful actors adapt; testing ALARM against deliberate obfuscation, coded language, and image manipulation is a natural next step.
• Finer-grained harm taxonomy. The current setup is binary (harmful vs. benign) and does not distinguish categories such as racism, sexism, or religion-based attacks, which is what most moderation policies actually require.
• Longitudinal and cost evaluation. Inference with a 72B model over large meme streams is expensive, and the paper does not report how reference sets hold up over time as topics drift.

Target Audience

Researchers and practitioners in multimodal machine learning, computational social science, and online trust and safety. It is most valuable to those working on content moderation systems who are frustrated by annotation costs, and to researchers interested in agentic self-improvement and label-free learning. Readers with a background in vision-language models and prompt engineering will get the most from the methodology sections, while policy and platform operations readers will find the experimental framing and dataset results directly relevant.

Authors’ abstract

The proliferation of harmful memes on online media poses significant risks to public health and stability. Existing detection methods heavily rely on large-scale labeled data for training, which necessitates substantial manual annotation efforts and limits their adaptability to the continually evolving nature of harmful content. To address these challenges, we present ALARM, the first lAbeL-free hARmful Meme detection framework powered by Large Multimodal Model (LMM) agent self-improvement. The core innovation of ALARM lies in exploiting the expressive information from "shallow" memes to iteratively enhance its ability to tackle more complex and subtle ones. ALARM consists of a novel Confidence-based Explicit Meme Identification mechanism that isolates the explicit memes from the original dataset and assigns them pseudo-labels. Besides, a new Pairwise Learning Guided Agent Self-Improvement paradigm is introduced, where the explicit memes are reorganized into contrastive pairs (positive vs. negative) to refine a learner LMM agent. This agent autonomously derives high-level detection cues from these pairs, which in turn empower the agent itself to handle complex and challenging memes effectively. Experiments on three diverse datasets demonstrate the superior performance and strong adaptability of ALARM to newly evolved memes. Notably, our method even outperforms label-driven methods. These results highlight the potential of label-free frameworks as a scalable and promising solution for adapting to novel forms and topics of harmful memes in dynamic online environments.