LLMs Learn Constructions That Humans Do Not Know

2025-08-22

AI Summary

Large language models (LLMs) systematically generate “false positive constructions”—syntactically ill-formed patterns rejected by native speakers—revealing latent syntactic misconceptions and exposing confirmation bias in existing construction probing methods. Method: The study innovatively integrates three complementary approaches: (1) context-embedding-based behavioral probing (assessing output distributions), (2) prompt-driven metalinguistic probing (eliciting explicit grammaticality judgments), and (3) hypothesis-testing simulation (evaluating empirical reliability against linguistic survey standards). Contribution/Results: Experiments demonstrate that LLMs stably produce diverse pseudo-constructions; in simulated linguistic surveys, these are misclassified as grammatical with up to 92% accuracy. This constitutes the first empirical evidence of systematic structural deficiencies in LLMs’ syntactic knowledge. Moreover, the work establishes a novel, linguistically grounded paradigm for evaluating the reliability of model language competence—offering critical implications for construction grammar theory, model evaluation, and interpretability research.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Simulating Human Behavior

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

This paper investigates false positive constructions: grammatical structures which an LLM hallucinates as distinct constructions but which human introspection does not support. Both a behavioural probing task using contextual embeddings and a meta-linguistic probing task using prompts are included, allowing us to distinguish between implicit and explicit linguistic knowledge. Both methods reveal that models do indeed hallucinate constructions. We then simulate hypothesis testing to determine what would have happened if a linguist had falsely hypothesized that these hallucinated constructions do exist. The high accuracy obtained shows that such false hypotheses would have been overwhelmingly confirmed. This suggests that construction probing methods suffer from a confirmation bias and raises the issue of what unknown and incorrect syntactic knowledge these models also possess.

Problem

• LLMs hallucinate grammatical constructions unknown to humans
• Models exhibit false positive constructions through behavioral probing
• Construction probing methods suffer from confirmation bias issues

Innovation

• Behavioral probing using contextual embeddings
• Meta-linguistic probing through prompt analysis
• Hypothesis testing simulation for false constructions