How do we know if a Large Language Model (LLM) actually understands the rules of language, or if it is simply a "stochastic parrot" repeating fragments of its training data? If a model correctly identifies that the plural of cat is cats, it might just be recalling a common word. But if it can tell you that the plural of a made-up word like wug is wugs, it has demonstrated morphological productivity—the ability to apply abstract rules to novel inputs.
Recent research has converged on using "nonce words" (phonologically plausible but non-existent terms) to probe the internal logic of neural networks. By stripping away the possibility of rote memorization, researchers are uncovering the specific ways LLMs generalize, where they fail, and how their "intuition" differs from human cognition.
The Artificial Wug Test
The core mechanism across this research direction is the adaptation of the classic "Wug Test" for machines. To isolate a model's structural understanding from its parametric memory, researchers like Mete Ismayilzada and Abdullatif Köksal present models with nonce roots—words that sound like they could exist in a language but do not.
In Evaluating Morphological Compositional Generalization in Large Language Models, the authors demonstrate that while models like GPT-4 appear proficient with common words, they face a "performance cliff" when asked to apply complex morphological transformations to these novel roots. This suggests that the models' apparent mastery of grammar is partially propped up by the frequency of specific words in their training sets.
Analogical vs. Symbolic Generalization
A central debate in this field is whether LLMs learn explicit, symbolic rules (like "add -s for plural") or if they generalize through analogical reasoning.
Valentin Hofmann and Janet B. Pierrehumbert investigated this by looking at how models turn adjectives into nouns (e.g., active to activity). In Derivational Morphology Reveals Analogical Generalization in Large Language Models, they found that LLMs do not behave like rule-following machines. Instead, they generalize by comparing new words to stored "exemplars."
Crucially, this leads to a distinctively non-human bias: while humans generalize based on the number of unique word types they know, LLMs are heavily influenced by raw token frequency. If a model sees one specific word millions of times, that single example disproportionately shapes how it handles new, similar-sounding nonce words.
Resilience to Broken Tokenization
One might assume that LLMs struggle with morphology because their "eyes"—the tokenizers—often chop words into arbitrary sub-word units that ignore linguistic boundaries. For example, a tokenizer might split unhappy into un- and happy, but it might split a nonce word like unbof into unb and of.
However, Thinh Hung Truong and Trevor Cohn found that larger models are surprisingly resilient to these "broken" tokens. In Revisiting subword tokenization: A case study on affixal negation in large language models, they show that models can often "see through" poor tokenization to recognize underlying morphological patterns, provided the model is large enough to have developed a robust internal representation of word structures.
Probing the "DNA" of Language
The use of nonce words extends beyond simple suffixes to the very sounds of language. Researchers are using these "empty" words to see if models understand phonology—the system of sounds.
- Subphonemic Features: Gal Astrach and Yuval Pinter used nonce words to prove that Transformers implicitly learn features like "vowel rounding" and "voicing." In Probing Subphonemes in Morphology Models, they show that while simple sound rules are stored in the model's first layers, complex rules like vowel harmony are processed in the deeper, contextual layers.
- Sound Symbolism: Jinhong Jeong and colleagues developed the LEX-ICON dataset to test if models associate certain sounds with meanings (e.g., "i" sounds for small things). By using pseudo-words, Do Language Models Associate Sound with Meaning? demonstrates that multimodal models possess a form of "phonetic intuition" that isn't just memorized from a dictionary.
The Limits of Abstraction
Despite these capabilities, the "morphological wall" remains a significant hurdle. When tasks move from simple word-level changes to complex semantic reasoning, the limits of LLM generalization become clear.
For instance, Giulia Rambelli and Emmanuele Chersoni found that while models can interpret common noun compounds, they struggle with novel ones. In Can Large Language Models Interpret Noun-Noun Compounds?, they used hypernym substitution (replacing carrot cake with root-vegetable dessert) to show that models often fail to abstract the underlying relationship (e.g., "made of") when the specific vocabulary changes.
By testing models on what they haven't seen, this research direction provides a precise yardstick for "understanding." It reveals that while LLMs are not mere parrots, their path to generalization is paved with statistical analogies rather than the rigid, symbolic rules that define human grammar.