Introduction: The Precision Gap in LLMs
Large Language Models (LLMs) have demonstrated remarkable capabilities in high-level cognitive tasks, ranging from solving complex mathematical problems to generating creative prose. However, their ability to adhere to strict, low-level structural constraints remains less understood. While a model might generate a coherent essay, can it reliably produce a sentence containing exactly three verbs?
This research investigates the "linguistic competence" of LLMs—their capacity to manipulate the fundamental building blocks of language, such as parts of speech and syntactic depth, independent of a specific communicative task. The study reveals a significant "precision gap": models often understand the general direction of a linguistic order (e.g., "more verbs") but frequently fail to meet exact numerical targets.
Methodology: Linguistic Profiling
The researchers employed a methodology known as "linguistic profiling." Rather than evaluating task-specific performance, they prompted five open-weight LLMs (Gemma-2B, Gemma-7B, LLaMA-2-7B, LLaMA-2-13B, and Mistral-7B) to generate sentences satisfying 20 distinct (morpho)syntactic constraints.
The Constraints
The constraints were categorized into two main types:
- Morpho-syntactic: These involved counts of specific Parts-of-Speech (POS). This included content words (nouns, verbs, adjectives) and functional words (adpositions, conjunctions, auxiliary verbs).
- Syntactic: These focused on structural features, such as the depth of the syntactic dependency tree, the length of the longest dependency link, and the placement or proportion of subordinate clauses.
Evaluation Metrics
To quantify performance, the study used two primary metrics:
- Success Rate (SR): The percentage of trials where the model's output exactly matched the requested numerical constraint.
- Spearman Correlation ($\rho$): A measure of how well the model's output tracked with increasing requested values. For instance, if a model is asked for 2, 4, and 6 verbs, a high correlation indicates that it correctly produces more verbs as the requested number increases, even if it misses the exact count.
Key Insights and Ideas
Directional Sensitivity vs. Numerical Precision
A major finding is that LLMs are "directionally correct" but "precisely wrong." While Success Rates were often low, Spearman correlation scores were frequently high (often exceeding 0.80). This suggests that models possess a conceptual understanding of linguistic categories—they know what a verb is and how to add more of them—but they lack the internal mechanism to "count" or strictly regulate these features during the autoregressive generation process.
Content vs. Function Words
The models demonstrated a clear hierarchy in their ability to control different word classes. They found it significantly easier to manage content words like nouns and adjectives than functional words like conjunctions or auxiliary verbs. This implies that LLMs have a more robust grasp of lexical categories (the "what" of a sentence) than the syntactic "glue" (the "how") that governs sentence structure.
The "Mistral" Factor and Architectural Efficiency
Interestingly, Mistral-7B emerged as a top performer in zero-shot scenarios, frequently outperforming the larger LLaMA-2-13B model. This suggests that raw parameter count is not the sole determinant of linguistic competence; training data quality and architectural efficiency play crucial roles in how well a model internalizes structural language rules.
Key Results
The evaluation highlighted the transformative effect of few-shot prompting. While zero-shot performance was often mediocre, providing five examples significantly boosted the Success Rate for most models, particularly for complex syntactic constraints like tree depth.
In zero-shot morphosyntax, Mistral-7B achieved the highest average Success Rate of 44.58%, while the smaller Gemma-2B struggled at 22.52%. However, few-shot learning helped close this gap for several models. Paradoxically, Mistral's performance occasionally dipped slightly in the few-shot setting for certain categories, suggesting a potential sensitivity to the specific examples provided.
Syntactic constraints proved to be the most challenging. Success rates for features like "max_depth" were often below 20% in zero-shot settings. The improvement seen in few-shot scenarios suggests that examples help models "anchor" their generation in realistic structural patterns, bringing the linguistic profile of their output closer to that of natural human language.
Conclusion and Future Directions
This research underscores that while LLMs are powerful reasoning engines, their ability to follow precise structural "orders" is limited. They exhibit a strong "directional" understanding of language but struggle with the exactitude required for sophisticated Controllable Text Generation (CTG). Future work should expand this linguistic profiling to other languages and investigate whether massive models like GPT-4 exhibit a similar "precision gap."
Summary updated