---
title: "YesBut V2 Exposes How AI Still Can't Understand Contradictory Humor"
subtitle: "A new benchmark of 1,262 juxtaposed comic panels exposes how badly today's vision-language models handle contradiction-based humor and comparative reasoning."
url: "https://alphasignal.ai/news/yesbut-v2-exposes-how-ai-still-can-t-understand-contradictory-humor"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-15T00:00:00.000Z"
updated: "2026-08-23T22:16:41.353Z"
type: "Paper"
topics:
  - "Benchmarks"
  - "Image"
subtopics:
  - "Vision Language"
upvotes: 2247
source_url: "https://arxiv.org/abs/2503.23137"
comments: 0
truncated: true
---

# YesBut V2 Exposes How AI Still Can't Understand Contradictory Humor

> A new benchmark of 1,262 juxtaposed comic panels exposes how badly today's vision-language models handle contradiction-based humor and comparative reasoning.

Humor built on contradiction is a stress test for machine reasoning. A comic that shows a hopeful **YES** panel next to a deflating **BUT** panel forces a reader to hold two situations in mind, spot the mismatch, and decode the social point being made. A team from Case Western Reserve University and The Hong Kong Polytechnic University argues that this is exactly where current vision-language models fall down, and they have built a benchmark to prove it.

Their paper, [When 'YES' Meets 'BUT'](https://arxiv.org/abs/2503.23137), introduces [YesBut V2](https://vulab-ai.github.io/YESBUT-v2/), a benchmark aimed squarely at contradictory-humor comprehension. It contains 1,262 comic images drawn from diverse multilingual and multicultural contexts, each carrying annotations that capture different aspects of narrative understanding. That is a roughly 3.6x expansion over the original 349-image version released at NeurIPS.

## Why juxtaposition breaks VLMs

Reading a YesBut comic is not really an image-captioning problem. The authors identify three primary challenges: accurate visual interpretation of each panel, nonlinear reasoning that integrates and compares key elements across panels to detect the humor-creating contradiction, and social and cultural comprehension of the norms that make the joke land. Miss any of those layers and the punchline is gone.

Experiments across a wide range of models show they still struggle with contradictory humor, and tend to make common errors including visual perception failures, key element omission, and hallucination in narrative understanding. Augmenting social knowledge and training through textual data distillation both improve performance.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [When 'YES' Meets 'BUT': Can Large Models Comprehend Contradictory Humor Through Comparative Reasoning?](https://arxiv.org/abs/2503.23137)
- [YESBUT-v2 Benchmark Project Page](https://vulab-ai.github.io/YESBUT-v2/)
- [YESBUT_Benchmark_V2 Dataset](https://huggingface.co/datasets/zhehuderek/YESBUT_Benchmark_V2)
- [Cracking the Code of Juxtaposition (YesBut v1)](https://vulab-ai.github.io/YESBUT_Homepage/)