Research map / Vision-language and multimodal / Multimodal LLMs
When ‘Yes’ Meets ‘But’: Can AI Comprehend Contradictory Humor in Comics?
Tuo Liang, Zhe Hu, Jing Li, Hao Zhang, Yiren Lu, Yunlai Zhou, Yiran Qiao, Disheng Liu, Jierui Peng, Jing Ma, Yu Yin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale up model size within a family | Qwen2-VL / LLaVA-OneVision · YesBut-V2 Title matching | accuracy: 74.72 -> 81.25 (Qwen2-VL 7B to 72B) | Improved |
| use human oracle literal descriptions instead of VLM-generated ones for LLM reasoning | LLMs · YesBut-V2 deep reasoning | accuracy | Improved |
| use descriptions from stronger VLMs (GPT-4o, Qwen2-VL-72B) instead of LLaVA-Next-13B | LLMs · YesBut-V2 deep reasoning | accuracy | Improved |
| add two-stage decomposition with model-generated description (ablation) | VLMs · YesBut-V2 Title matching | accuracy | Hurt |
| add two-stage decomposition with model-generated description (ablation) | LLaVA-OneVision-72B · YesBut-V2 Title matching (single image) | accuracy: 77.10 -> 58.56 | Hurt |
| add two-stage decomposition with model-generated description (ablation) | Qwen2-VL-72B · YesBut-V2 Underlying Symbolism (single image) | accuracy: 79.98 -> 76.66 | Hurt |
| add decomposition on LLaVA-OneVision-0.5B (ablation) | LLaVA-OneVision-0.5B · YesBut-V2 Title matching (single image) | accuracy: 38.53 -> 46.44 | Improved |
| replace composite image by split multi-image panels (ablation) | VLMs · YesBut-V2 deep reasoning | accuracy: LLaVA-OneVision-7B symbolism 67.64 -> 65.45 | Hurt |
| finetune language component on 20,000 synthetic text-only comic descriptions (ablation) | LLaVA-Next-7B · YesBut-V2 Underlying Symbolism | accuracy: +8.56 percentage points | Improved |
| finetune language component on synthetic text-only data (ablation) | LLaVA-Next-13B · YesBut-V2 Underlying Symbolism | accuracy: +4.89 percentage points | Improved |
| add annotated social knowledge to prompt (ablation) | VLMs · YesBut-V2 deep reasoning | accuracy | Improved |
| add translated embedded text (ablation) | GPT-4o · YesBut-V2 Title matching | accuracy: 77.42 -> 88.73 | Improved |
Closest papers by meaning
- Cracking the Code of Juxtaposition: Can AI Models Understand the Humorous Contradictions (NeurIPS 2024 · similarity 0.95)
- HUMORCHAIN: Theory-Guided Multi-Stage Reasoning for Interpretable Multimodal Humor Generation (CVPR 2026 · similarity 0.88)
- Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps (ICLR 2025 · similarity 0.87)
- SatireDecoder: Visual Cascaded Decoupling for Enhancing Satirical Image Comprehension (AAAI 2026 · similarity 0.86)
- From Perception to Punchline: Empowering VLM with the Art of In-the-wild Memes (NeurIPS 2026 · similarity 0.86)
- MARVEL: Multidimensional Abstraction and Reasoning through Visual Evaluation and Learning (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
