Research map / Multi-agent systems and games / LLM agents

MAMA-Memeia! Multi-Aspect Multi-Agent Collaboration for Depressive Symptoms Identification in Memes

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the RESTOREx meme dataset with LLM and human explanations and MAMA-Memeia, a multi-agent multi-aspect discussion framework (GPT-4o, Claude 3.5 Sonnet, Gemini-2.0-flash) for fine-grained depressive symptom identification in memes.

Problem

Identifying fine-grained PHQ-9 depressive symptoms expressed in memes, which involve sarcasm, irony and figurative language that is hard for models to interpret.

Approach

A training-free multi-agent framework where three closed-source multimodal LLMs each receive a different CAT-derived knowledge aspect (depression, emotional, cultural), then iteratively discuss predictions, confidences and explanations over rounds before a confidence-weighted consensus vote.

Key results

• MAMA-Memeia reaches 72.73 macro-F1 and 72.45 weighted-F1, a 7.55% macro-F1 gain over the previous SOTA (65.18).
• Best single-agent setup (Claude 3.5 Sonnet with raw explanations) reaches 68.39 macro-F1; GPT-4o vanilla reaches 60.00.
• Multi-aspect prompting alone raises GPT-4o and Gemini-2.0-flash substantially over vanilla prompting; removing aspect-specific prompting drops MAMA-Memeia to 71.88 macro-F1.

Surprising findings

• Claude 3.5 Sonnet with vanilla explanations outperforms human-written explanations on macro-F1 (68.39 vs 65.00).
• Human-annotated explanations score lower than several LLM-generated explanation setups.

Limitations

• Relies on closed-source API models with limited transparency and reproducibility.
• Gemini-2.0-flash over-predicts and GPT-4o under-predicts, with explanation length possibly a factor; mechanism not fully analyzed.
• Some test-set labels are noisy; the Lack of Energy class was removed and non-meme images filtered, reducing dataset size by about 20%.

Experimental setup

Datasets: RESTORE RESTOREx

Benchmarks: RESTOREx

Metrics: macro-F1 micro-F1 weighted-F1

Baselines: BERT MentalBERT BART MentalBART ViT ResNet EfficientNet CLIP VisualBERT LLaVA 1.5

Builds on: RESTORE Cognitive Analytic Therapy competencies PHQ-9

Closest papers by meaning

• MoodAngels: A Retrieval-augmented Multi-agent Framework for Psychiatry Diagnosis (NeurIPS 2025 · similarity 0.83)
• MSE-Adapter: A Lightweight Plugin Endowing LLMs with the Capability to Perform Multimodal Sentiment Analysis and Emotion Recognition (AAAI 2025 · similarity 0.82)
• MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.82)
• Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.81)
• MuMA-ToM: Multi-modal Multi-Agent Theory of Mind (AAAI 2025 · similarity 0.81)
• DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis (AAAI 2026 · similarity 0.81)