(C) BoingBoing
This story was originally published by BoingBoing and is unaltered.
. . . . . . . . . .



Study reveals 700% error increase when AI encounters cat-related text [1]

['Gail Sherman']

Date: 2025-07-31

Even the most advanced AI models can easily be confounded by cats.

Reasoning language models are large language models that are trained to break problems down into smaller chunks. They solve problems in steps and revisit and revise previous steps. This iterative process allows reasoning models to solve more complex problems. However, they are easily confused by cats, according to a new study.

The study found that injecting "query-agnostic adversarial triggers – short, irrelevant text that, when appended to math problems, systematically mislead models to output incorrect answers." Some models were up to 700% more likely to get an answer wrong when subjected to a CatAttack.

Screenshot: Cats Confuse Reasoning LLM: Query-Agnostic Adversarial

Triggers for Reasoning Models

A human solving a math problem would easily recognize and disregard these off-topic triggers, but even state-of-the-art reasoning language models were vulnerable to CatAttack. In addition, substantially more computing power was used even when the correct solution was reached. Non-cat related adversarial triggers were equally effective.

In their conclusion, the researchers emphasize that, "This work underscores the need for more robust defense mechanisms against adversarial perturbations, particularly, for models deployed in critical applications such as finance, law, and healthcare."

Previously:

• Frustrated feline swipes at cat show judge (video)

• A struggling restaurant was saved by the power of cats

• Watch a real cat's hilarious reaction to its lego doppelgänger

[END]
---
[1] Url: https://boingboing.net/2025/07/31/study-reveals-700-error-increase-when-ai-encounters-cat-related-text.html

Published and (C) by BoingBoing
Content appears here under this condition or license: Creative Commons BY-NC-SA 3.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/boingboing/