Research map / Vision-language and multimodal / Multimodal LLMs

HUMORCHAIN: Theory-Guided Multi-Stage Reasoning for Interpretable Multimodal Humor Generation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 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.

Proposes HUMORCHAIN, a theory-guided multi-stage LLM pipeline (image description, plausibility/incongruity judgment, four humor-strategy generators, safety check) plus a LoRA-tuned Qwen3-VL-4B humor discriminator with a classifier head that triggers regeneration, evaluated on humorous captioning against prompting baselines and CLoT.

Problem

Existing humorous image captioning systems produce fluent but literal or unfunny descriptions, lacking explicit modeling of humor mechanisms grounded in humor theory.

Approach

Operationalize Incongruity-Resolution, Benign Violation, Superiority, and Relief theories as a staged reasoning chain that routes each image to one of four humor strategies (Absurdity, Contrast_irony, Emotion_analogy, Object_analogy), with a human-preference-trained discriminator closing a generate-evaluate-refine loop.

Key results

• HUMORCHAIN (I) achieves the top Arena Elo of 1554.60 and BT 3.57, ahead of the best prompting baseline F (Elo 1528.84) and external CLoT (Elo 1464.06).
• In pairwise human comparisons, I beats OxfordTVG-HIC captions with an 0.860 win rate (n=1007) and CLoT with 0.683 (n=794).
• The LoRA + classifier-head discriminator reaches precision 0.670 versus 0.523 for the un-tuned baseline; closed-source LLM judges reach precision of only about 0.46-0.47 due to over-positive labeling.

Surprising findings

• Closed-source LLMs (Gemini-2.5-Flash, GPT-4.1, Claude-3.5-Haiku) label 88-97% of captions as humorous, yielding low precision, so a larger model does not improve discrimination.
• Few-shot and CoT prompting alone give little gain over zero-shot in the single-title humor score (combined strategies sit around 0.38-0.40).

Limitations

• Theory coverage is incomplete; culturally embedded or reference-dependent humor may be underrepresented.
• Humor is subjective and the discriminator is biased toward majority-appealing captions rather than individual preferences.
• Failure cases arise from inadequate image descriptions and non-incongruous images with limited humor cues.

Experimental setup

Datasets: Meme-Image-No-Text Oogiri-GO OxfordTVG-HIC

Benchmarks: Meme-Image-No-Text Oogiri-GO OxfordTVG-HIC

Metrics: win rate hard win rate Arena Elo Bradley-Terry human humor score precision CLIPScore EA-Rev GM-Rev Distinct-1

Baselines: Zero-shot Few-shot Rule-Based prompting CoT + rule-guided CLoT OxfordTVG-HIC

Builds on: Chain-of-Thought CLoT Chain-of-Humor OxfordTVG-HIC

Closest papers by meaning

• From Perception to Punchline: Empowering VLM with the Art of In-the-wild Memes (NeurIPS 2026 · similarity 0.92)
• Humor in AI: Massive Scale Crowd-Sourced Preferences and Benchmarks for Cartoon Captioning (NeurIPS 2024 · similarity 0.90)
• Cracking the Code of Juxtaposition: Can AI Models Understand the Humorous Contradictions (NeurIPS 2024 · similarity 0.90)
• Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps (ICLR 2025 · similarity 0.89)
• Talk Funny! A Large-Scale Humor Response Dataset with Chain-of-Humor Interpretation (AAAI 2024 · similarity 0.88)
• Let's Think Outside the Box: Exploring Leap-of-Thought in Large Language Models with Creative Humor Generation (CVPR 2024 · similarity 0.86)