Survey maps multimodal LLM struggles with visual humor and cultural reasoning

Illustration accompanying: Computational Humor with Multimodal LLMs: Methods, Datasets, Evaluation, and Challenges

A new survey maps the landscape of computational humor across multimodal LLMs, identifying why systems struggle with memes, cartoons, and comics that rely on cultural context and non-literal reasoning rather than pixel-level scene understanding. The work organizes the field through a capability hierarchy spanning recognition, interpretation, and generation, while documenting the industry shift from task-specific fusion architectures toward large-model approaches built on multimodal alignment. This matters because humor understanding remains a hard frontier for evaluating genuine reasoning and cultural grounding in foundation models, and the benchmark taxonomy here will likely shape how teams measure progress on this underexplored capability.

Modelwire context

Explainer

The survey's core contribution is not the humor benchmarks themselves, but the explicit framing of humor as a test for cultural grounding and non-literal reasoning. Most prior work treated humor as a narrow NLP task; this positions it as a diagnostic for whether models actually understand context beyond surface patterns.

This connects directly to the MeetingToM benchmark from earlier this month, which exposed how current evals miss latent social dynamics beneath overt signals. Humor operates the same way: a meme's joke lives in what's unstated, the cultural reference, the ironic gap between image and text. Both papers argue that foundation models can pattern-match surfaces while failing at the interpretive layer that requires genuine reasoning about human intent and shared context. The MIRA-Ev clinical benchmark from the same week reinforces this pattern, showing that correctness without interpretability masks whether models ground reasoning or just chase correlations.

If major labs release humor-specific evals using this survey's taxonomy within the next six months, and if those benchmarks show current multimodal models scoring below 60% on interpretation tasks (not just recognition), that confirms humor remains a legitimate reasoning frontier rather than a solved problem misclassified as hard.

Coverage we drew on

MeetingToM: Evaluating Multimodal LLMs on Theory-of-Mind Reasoning in Multi-Party Meetings · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsMultimodal LLMs · Memes · Cartoons · Comics

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Computational Humor with Multimodal LLMs: Methods, Datasets, Evaluation, and Challenges”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.