💬 NLP

Language Models Generalize to Human-like Word Order Preferences

This study demonstrates that language models, when trained on data lacking direct evidence of modifier ordering, consistently generalize to prefer human-like scope-homomorphic word orders, suggesting these biases emerge from general learning mechanisms rather than simple statistical associations.

Original authors: Amanda Popadich, Shane Steinert-Threlkeld

Published 2026-08-06

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Amanda Popadich, Shane Steinert-Threlkeld

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to speak a new language, but you decide to play a trick on it. You give the robot a massive library of books, but you carefully rip out every single page that contains a sentence with more than one description for a noun. So, the robot learns that you can say "the red ball" or "the big ball," but it never sees "the big red ball" or "the red big ball." It has no direct evidence of how to stack multiple descriptions together. Now, here is the big question: If you ask the robot to guess the order for a new sentence like "the big red ball," will it just guess randomly? Or will it somehow figure out the hidden logic of language and pick the right order, just like a human child would?

This is the kind of puzzle that sits at the heart of linguistics and artificial intelligence. Scientists have long wondered if humans are born with a special "language chip" in their brains that tells them how to order words, or if we just learn everything by noticing patterns in what we hear. To test this, researchers use something called "Artificial Language Learning." It's like a controlled experiment where they teach people or machines a tiny, made-up language with missing pieces to see how they fill in the blanks. If a learner picks a specific pattern even though they've never seen it before, it suggests they have a built-in bias or a clever way of guessing based on the structure of the world, not just the frequency of words. This paper dives into that mystery using modern AI language models as our test subjects.

The Great Word-Order Mystery

In this study, researchers Amanda Popadich and Shane Steinert-Threlkeld decided to see if computer programs, known as language models, could pull off the same magic trick that human learners do. They wanted to know: Can a machine learn the "rules" of how to stack adjectives and numbers without ever seeing a single example of a stacked phrase?

The specific rule they were hunting for is called scope-homomorphism. That's a fancy way of saying that the order of words in a phrase should match the order of their meaning. Think of it like Russian nesting dolls. If you have a "big red ball," the word "red" describes the ball directly, and "big" describes the "red ball." In English, we naturally put the bigger, outer description ("big") first, and the specific, inner description ("red") closer to the noun. The researchers wanted to see if their AI models would naturally prefer this "nesting doll" order, even if they were trained on a diet of only single-description phrases.

The Experiment: A Language Diet

To test this, the team created a very strict training environment. They took a huge chunk of the internet (specifically, the 2023 Wikipedia dump) and filtered it. They used a computer program to find every sentence with a noun that had two or more modifiers (like "those two fluffy dogs") and broke them apart. They replaced "those two fluffy dogs" with separate examples like "those dogs," "two dogs," and "fluffy dogs."

The result was a "poverty of the stimulus" diet. The models were fed 100 million words, but they never saw a single example of a noun with two modifiers. They learned the individual ingredients but never saw the final cake. Then, the researchers tested the models by asking them to choose between two versions of a sentence: the one that followed the "nesting doll" logic (scope-homomorphic) and the one that didn't.

They tested three different sizes of AI models: a small one (52 million parameters), a medium one (110 million), and a large one (350 million).

The Results: The AI Gets It Right

The findings were surprisingly consistent. Even though the models had never seen a phrase with multiple modifiers during their training, all three models consistently preferred the "nesting doll" order.

When the researchers asked the models to choose between "those two dogs" (the logical order) and "two those dogs" (the scrambled order), the models overwhelmingly picked the logical one.

The small model got it right about 76% of the time.
The medium model got it right about 73% of the time.
The large model got it right about 70% of the time.

This suggests that the models didn't just memorize the training data; they generalized a rule. They figured out that when you have multiple descriptions, there is a specific, logical way to arrange them, even without being shown the answer. Interestingly, the size of the model didn't change the result much; the small, medium, and large models all showed this bias, suggesting it's a robust feature of how these systems learn from language.

The Twist: It's Not Just About Word Pairings

The researchers then asked a follow-up question: How did the models figure this out? One popular theory is that learners just look at how often words hang out together. For example, maybe "red" is usually found right next to "ball," so it should always be closer to the noun than "big." This is measured by something called Pointwise Mutual Information (PMI), which is basically a score of how strongly two words are associated.

The team checked the math. They calculated the association scores for the words in their training data. They found that, yes, adjectives are strongly associated with nouns, and numbers and demonstratives (like "those") are less so. If the models were just following these association scores, they should have been able to predict the word order perfectly.

But here is the twist: The association scores failed to predict the models' choices.

When the researchers compared the models' actual choices to what the math predicted, the correlation was almost zero. The models were choosing the right order, but they weren't doing it because they were counting how often words appeared together. This suggests the models are picking up on something deeper—perhaps a structural understanding of how language works—that goes beyond simple word-pair statistics.

What This Means

This study shows that language models can recover human-like linguistic generalizations from "impoverished" input. Just like a human child who might guess the right word order in a new language without seeing every possible combination, these AI models found a pattern that wasn't explicitly in their training data.

However, the authors are careful to note that this doesn't prove AI learns exactly the same way humans do. The models are not conscious children; they are complex statistical engines. But the fact that they developed this bias without being told to suggests that you don't necessarily need a pre-programmed "language chip" to learn these rules. Sometimes, the rules of language can emerge naturally from the experience of learning, even when the evidence is incomplete.

The researchers also noted a small difference between the AI and humans: while humans usually show the strongest preference for ordering demonstratives and adjectives (like "that big dog"), the AI models showed the strongest preference for demonstratives and numbers (like "those two dogs"). This hints that while the general ability to generalize is there, the specific flavors of the bias might depend on the details of the training data.

In the end, this paper gives us a cool new tool for understanding language. By feeding AI models a diet of missing information, we can see what they "guess" on their own, helping us understand whether the rules of language are hard-wired or learned from the patterns of the world.

Technical Summary: Language Models Generalize to Human-like Word Order Preferences

Problem Statement
A central question in linguistics and language acquisition is whether linguistic biases can emerge from general learning mechanisms operating over underdetermined input, or if they require innate, language-specific constraints. Artificial Language Learning (ALL) studies have demonstrated that human learners reliably generalize beyond provided evidence, specifically preferring "scope-homomorphic" noun phrase modifier orders (where linear order reflects semantic scope, e.g., two fluffy dogs vs. fluffy two dogs). However, the source of these human biases remains unclear. This paper investigates whether neural language models (LMs), acting as general learners without innate linguistic constraints, exhibit the same scope-homomorphic bias when trained on impoverished input that lacks direct evidence of multi-modifier ordering.

Methodology
The authors employ a Filtered Corpus Training (FiCT) paradigm to simulate "poverty of the stimulus" conditions for language models.

Corpus Construction: Using the 2023 Wikipedia dump, the authors constructed a perturbed training corpus. They identified noun phrases (NPs) containing multiple modifiers (demonstratives, numerals, or adjectives) and decomposed them into variants containing only a single modifier per noun. This process eliminated all direct evidence of multi-modifier ordering while retaining single-modifier instances.
Model Architecture: Three decoder-only Transformer models based on the OPT architecture were trained with varying parameter counts: Small (52M), Medium (110M), and Large (350M).
Training: Models were trained for 40 epochs on the perturbed corpus.
Evaluation: The models were tested using minimal pairs of multi-modifier NPs. For each item, the model's log-probability was compared between the canonical (scope-homomorphic) order and a swapped (non-homomorphic) order.
- Metrics:
  - Homomorphic Consistency Delta ( $HC\Delta$ ): The difference in log-probability between the homomorphic and non-homomorphic orders.
  - Homomorphic Consistency Percentage ( $HC\%$ ): The proportion of items where the model preferred the homomorphic order.
Source Analysis: To determine the origin of the bias, the authors analyzed:
- Frequency: Counts of single-modifier types in the training data.
- Pointwise Mutual Information (PMI): The strength of association between nouns and their modifiers, hypothesized to drive ordering preferences based on co-occurrence statistics.

Key Results

Emergence of Bias: Across all three model sizes, the models consistently preferred scope-homomorphic orders despite never observing such structures during training. All models showed $HC\Delta$ values and $HC\%$ accuracies significantly above chance.
Model Size Independence: The strength of the preference did not significantly vary with model size. While perplexity increased slightly with size, the ordering preference remained robust across the 52M, 110M, and 350M models.
Modifier Type Variation: Preferences varied by modifier combination.
- Demonstrative-Numeral (Dem-Num): Showed the strongest preference for homomorphic orders (e.g., ~94% accuracy for smaller models).
- Demonstrative-Adjective (Dem-Adj) and Numeral-Adjective (Num-Adj): Showed weaker, though generally reliable, preferences.
- Note: This hierarchy differs from human ALL results, where Dem-Adj typically shows the strongest preference.
Source of Bias:
- Frequency: Simple frequency counts of modifier types in the training data did not explain the results.
- PMI: Analysis of noun-modifier association strength (PMI) revealed that while adjectives had the highest association strength, PMI scores did not predict the models' ordering preferences. Correlations between $\Delta$ PMI and model preferences were weak and near chance levels.

Key Contributions

FiCT Paradigm for Scope-Homomorphism: The paper introduces a controlled learning environment (FiCT) adapted for investigating scope-homomorphic word order in LMs, closely mirroring the conditions used in human ALL experiments.
Demonstration of Generalization: The study provides empirical evidence that LMs can recover human-like linguistic generalizations (scope-homomorphism) from impoverished input where direct evidence is absent.
Disentangling Bias Sources: By comparing model behavior with PMI analysis, the authors demonstrate that the learned bias is not merely a reflection of surface-level lexical co-occurrence statistics, suggesting the models are utilizing more abstract structural or semantic generalizations.

Significance and Claims
The authors claim that their findings demonstrate that strong innate linguistic constraints are not strictly necessary for scope-homomorphic preferences to emerge; rather, such biases can arise in general learning systems from experience with impoverished input. This parallels the central result of human ALL studies, suggesting that the mechanisms underlying these generalizations may be learnable from distributional and structural properties of the input.

The paper positions LMs as a useful framework for investigating the origins of linguistic biases. By showing that these models can generalize without direct evidence, the work provides a controlled setting to separate what aspects of linguistic generalization require language-specific constraints from what can emerge through experience. The authors modestly note that while LMs are not direct models of human learners, their ability to recover these biases suggests that the input statistics and structural regularities available in natural language may be sufficient to support such generalizations, even if the specific mechanism differs from human cognition.