Research map / Vision-language and multimodal / Multimodal LLMs

SatireDecoder: Visual Cascaded Decoupling for Enhancing Satirical Image Comprehension

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 1 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.

Proposes SatireDecoder, a training-free multi-agent framework (RAM, BLIP, Qwen2 agents) with chain-of-thought reasoning and temperature-based uncertainty analysis to improve MLLM comprehension of visual satire on YesBut.

Problem

Vision-language models struggle to comprehend purely visual satire: they fail to integrate local entity relationships with global context, leading to misinterpretation and hallucination.

Approach

A multi-agent visual cascaded decoupling (RAM for local tags, BLIP for global captions, Qwen2 for discrepancy analysis) feeds a three-step CoT prompt, and temperature sweeps are selected by minimizing uncertainty (negative Jaccard for tags, negative BERTScore for captions) to reduce hallucination, all without training.

Key results

• On YesBut, Qwen2.5-VL+SatireDecoder reaches an average NLG score of 0.360 and user-study average of 64.58% vs 54.33% for base Qwen2.5-VL.
• LLaVA with SatireDecoder improves user-study correctness from 25.67 to 62.33 (about 37 points).
• Removing uncertainty analysis raises Qwen2.5-VL+SatireDecoder CHAIR_i from 26.90 to 39.75 and drops user-study AVE from 64.58 to 57.17 (ablation reported on Qwen2.5-VL row).

Surprising findings

• Uncertainty analysis, not only the multi-agent component, drives much of the hallucination reduction, lowering CHAIR_i from about 55 to about 37 for LLaVA.

Limitations

• User study uses only 100 sampled images and three binary judges per criterion.
• Uncertainty is computed with temperature sweeps (0.2 to 1.0), adding inference cost.
• Reliance on RAM, BLIP and Qwen2 as agents limits what the decoupling can capture.

Experimental setup

Datasets: YesBut

Benchmarks: YesBut

Metrics: BLEU ROUGE-L METEOR BERTScore user study correctness/length/completeness/faithfulness CHAIR_i CHAIR_s

Baselines: MiniGPT-4 GPT-4 Kosmos-2 Gemini LLaVA LLaVA-NeXT Qwen-VL Qwen2.5-VL

Builds on: chain-of-thought prompting BERTScore CHAIR

Closest papers by meaning

• Cracking the Code of Juxtaposition: Can AI Models Understand the Humorous Contradictions (NeurIPS 2024 · similarity 0.86)
• SAVER: Mitigating Hallucinations in Large Vision-Language Models via Style-Aware Visual Early Revision (AAAI 2026 · similarity 0.85)
• Learning Interleaved Image-Text Comprehension in Vision-Language Large Models (ICLR 2025 · similarity 0.85)
• When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding (NeurIPS 2025 · similarity 0.85)
• HUMORCHAIN: Theory-Guided Multi-Stage Reasoning for Interpretable Multimodal Humor Generation (CVPR 2026 · similarity 0.85)
• Decoupling Contrastive Decoding: Robust Hallucination Mitigation in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.85)