Knowing the Scale, Missing the Negation: Language Models and Corrective Negation by Excess in English, Spanish, and French

Language models are the engines behind many of the digital tools we use today, capable of generating text, translating languages, and answering questions with startling fluency. To understand how these systems work, researchers often look at how they handle the building blocks of human communication, such as adjectives that describe intensity. Words like "warm," "hot," and "boiling" exist on a scale of increasing heat, and humans naturally understand that if something is boiling, it is also at least warm. This logical connection is called an entailment. However, human language is full of twists where we deny a weaker word not to say the opposite is true, but to say the stronger word is the right one. When someone says a soup is "not warm but boiling," they are not denying that the soup has heat; they are rejecting "warm" as an understatement. This specific type of correction, where a speaker swaps a weak description for a strong one, is a subtle test of whether a machine truly understands the relationship between words or is merely matching patterns.

A team of researchers set out to see if modern language models could navigate this tricky linguistic terrain. They focused on a phenomenon known as corrective negation by excess, where a sentence denies a weaker adjective to affirm a stronger one. The team gathered hundreds of pairs of adjectives from English, Spanish, and French, such as "dim" and "dark" or "good" and "excellent," which are arranged in order of intensity. They then asked five different language models to judge these pairs under various conditions. First, they confirmed that the models knew the order of the words, recognizing that "dark" is a stronger version of "dim." Next, they tested if the models understood that if something is "dark," it must be at least "dim." Finally, they presented the models with the corrective sentences, like "It is not dim but dark," and asked if the object was still at least "dim."

The results revealed a significant gap between what the models know and how they apply that knowledge. On the units where the models correctly identified the order of the words and accepted the basic logic that "dark" implies "dim," they succeeded in understanding the affirmative statement 93.4% of the time. They also answered simple negation questions correctly 99.2% of the time, but this high accuracy is misleading: the models generally leaned towards answering "no" overall (with "yes" rates ranging from 34% to 44%), a bias that artificially inflates performance on questions where the correct answer is "no." When faced with the corrective sentence "It is not dim but dark," their performance dropped sharply. On the same set of items where they had previously shown they understood the relationship between the words, they failed to recognize the correction in 452 cases, while succeeding in only 6 cases. Instead of understanding that the soup is still warm (just more so), the models often treated the sentence as a simple denial, concluding that the object was not dim at all.

This failure was not due to a lack of knowledge about the words themselves. When the researchers tested a similar construction using the same word, such as "It is not dim but very dim," the models performed with high accuracy, getting it right 94.4% of the time. This suggests the models can handle the structure of the sentence when the words are identical, but they stumble when they must connect two different words within a corrective frame. The error pattern was consistent across all five models tested and in all three languages. In many instances, the models answered "no" to the question of whether the object was at least the weaker term, effectively ignoring the second half of the sentence that provided the correction. This indicates that the models are not failing because they cannot rank the adjectives, but because they struggle to override the literal meaning of the negation when a stronger alternative is introduced.

The study also explored whether making the correction more explicit would help. In English, the researchers compared the standard "not X but Y" with a version that added the word "rather," creating the phrase "not X but rather Y." This small change, which signals a correction more clearly, improved the models' accuracy from 36.1% to 49.3%. While this was an improvement, it was not a complete fix, as the models still missed the meaning in more than half of the cases. The researchers noted that in Spanish, where a specific conjunction is used for corrections, the models performed slightly better than in English or French, but the fundamental difficulty remained. The data suggests that these systems possess the raw knowledge of how words relate to one another, but they often fail to use that knowledge when the sentence structure requires them to reject a weaker description in favor of a stronger one.

The implications of this finding extend beyond academic curiosity. Language models are increasingly used to summarize reviews, analyze feedback, and process customer complaints. If a model reads a review that says a product is "not good but outstanding," it might misinterpret the strong praise as a negative statement because it fails to grasp the corrective nature of the sentence. The researchers conclude that current evaluations of these systems need to include this specific type of linguistic challenge. It is not enough to check if a model knows that "outstanding" is stronger than "good"; the model must also be tested on whether it can apply that knowledge when a sentence denies the weaker term to affirm the stronger one. Until models can reliably navigate these corrections, they risk misreading the tone and intent of human communication, turning high praise into criticism and missing the nuance that makes language so powerful.