AI learns cartoon humor with new Incongruity-Resolution Supervision framework

Researchers have developed a new framework called Incongruity-Resolution Supervision (IRS) to improve multimodal humor understanding in AI models. This framework breaks down humor comprehension into identifying incongruities, resolving them, and aligning with human judgments. When applied to 7B, 32B, and 72B parameter models, IRS enhanced performance on the New Yorker Cartoon Caption Contest benchmark, with the largest model achieving a 76.10% ranking score, surpassing both non-expert human performance and existing multimodal baselines. AI

IMPACT This framework could lead to more sophisticated AI systems capable of understanding nuanced human communication like humor.

RANK_REASON Academic paper detailing a new framework and benchmark results. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• 32B models
• 72B models
• 7B models
• arXiv
• Hatice Merve Vural
• Incongruity-Resolution Supervision

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →