Understanding Multi-Agent Reasoning with Large Language Models for Cartoon VQA
Tong Wu, Thanet Markchom
TL;DR
This work targets cartoon-based VQA, where stylisation and narrative context challenge standard LLMs trained on natural images. It introduces a modular three-agent framework (visual, language, critic) that decomposes perception, reasoning, and verification, and evaluates it on Pororo and Simpsons datasets using open-ended prompts. The study shows that visual grounding and critic-based validation contribute to robustness in a domain with exaggerated cues and episodic content, with effects varying by dataset. By comparing a cartoon-tailored GPT-4o-mini visual description agent against BLIP-2, the paper provides insights into how explicit cartoon grounding improves multimodal inference and answer quality. The findings inform the design of multi-agent architectures for stylised VQA and multimodal reasoning beyond natural images.
Abstract
Visual Question Answering (VQA) for stylised cartoon imagery presents challenges, such as interpreting exaggerated visual abstraction and narrative-driven context, which are not adequately addressed by standard large language models (LLMs) trained on natural images. To investigate this issue, a multi-agent LLM framework is introduced, specifically designed for VQA tasks in cartoon imagery. The proposed architecture consists of three specialised agents: visual agent, language agent and critic agent, which work collaboratively to support structured reasoning by integrating visual cues and narrative context. The framework was systematically evaluated on two cartoon-based VQA datasets: Pororo and Simpsons. Experimental results provide a detailed analysis of how each agent contributes to the final prediction, offering a deeper understanding of LLM-based multi-agent behaviour in cartoon VQA and multimodal inference.