The Technical Drivers Behind Language Model Expression Quality: Corpus Composition and Training Strategies

From “Talking” to “Talking Well”: The Essential Differences in LLM Expression

In daily use, users can intuitively feel the differences in response quality among various language models. Some models output accurate but mechanical content, while others can automatically adjust their tone based on context, select more appropriate vocabulary, and even maintain a stable rhythm and logical coherence in long-text generation. This difference is often summarized as “expression quality,” but at the technical level, it is the result of multiple factors, including the breadth and depth of training data, the model’s ability to perceive context, and the embedding of human preferences during alignment training.

1. The Nutritional Structure of Corpus: The Key Role of Book Data

A language model’s expressive ability heavily depends on its “reading volume,” especially high-quality, long-form, and diverse text inputs. It is widely recognized in the industry that while general web corpora provide rich knowledge, they have limitations in sentence variety, rhetorical devices, and narrative rhythm. In contrast, book corpora possess natural advantages such as complete sentence structures, rigorous logic, and elegant expression, providing the model with deeper linguistic templates.

Public information indicates that a prominent AI lab reportedly launched a large-scale book collection initiative, investing tens of millions of dollars to acquire millions of physical books from the used book market, aiming to obtain “all the books in the world.” The most notable aspect of this project was its scanning method—using industrial hydraulic cutters to sever the spines, turning the books into loose pages fed into a high-speed scanning assembly line, with the original books subsequently destroyed. Although this method sparked controversy regarding copyright and environmental impact, from a technical perspective, it reveals a core judgment: the model’s learning of excellent expression relies not on intensive reading of a few classics, but on the broad absorption of massive text.

Why is book corpus so important? From a linguistic distribution perspective, book texts far exceed ordinary web content in syntactic complexity, rhetorical density, and logical hierarchy. When a model is exposed to a large amount of such text during the pre-training phase, it can invoke richer syntactic and lexical resources during generation, resulting in naturally more layered output. Conversely, if a model’s training data is dominated by short, fast-paced dialogues and social media posts, its expression tends to be concise and straightforward. While highly efficient for information transfer, it struggles in scenarios requiring in-depth argumentation or literary expression.

A typical example of this difference occurs in specific linguistic contexts. For instance, early versions of certain models generating Chinese text exhibited a tendency to mix classical and modern Chinese or coin new archaic terms—an obvious byproduct of heavy exposure to ancient texts. After iterations, these extreme features were mitigated, but what lies behind them reflects the profound shaping power of corpus composition on expressive style.

2. Theory of Mind vs. Expressive Flexibility

Language expression is not merely a combination of vocabulary; it is also an assessment of the reader’s psychological state. A high-quality article considers the reader’s knowledge background, comprehension ability, and emotional state to adjust information density and expression. This ability is known in cognitive science as “Theory of Mind” (ToM)—the ability to understand others’ beliefs, intentions, and emotions.

A 2024 study published in a top academic journal evaluated mainstream models using standardized Theory of Mind tests. The results showed that in most test items, a representative model scored at or above the human average. This data suggests that the model’s ability to reason about others’ psychological states is already quite impressive.

However, a thought-provoking phenomenon is that models scoring high on ToM tests do not necessarily exhibit corresponding expressive flexibility in actual conversations. This disconnect between “knowing” and “doing” reveals an important characteristic of current model architectures—the model can accurately judge the appropriate emotion or tone in a specific context at the reasoning level, but the decoding strategy and sampling parameter settings during the generation phase may bias its output towards safe, steady expressions rather than agile language that fits the dialogue context.

Essentially, ToM capability provides the model with “knowledge,” but whether this knowledge can be translated into flexible expressive shifts depends on the model’s generation strategy. Temperature parameter settings, penalty factor adjustments, and randomness control during sampling all influence the model’s trade-off between “safe expression” and “personalized expression.” If the model’s inherent bias tends toward conservatism, even with strong comprehension capabilities, it will struggle to produce impressively interesting responses.

3. Human Feedback Alignment: The Crucial Process for Crafting the “Human Touch”

After pre-training, models typically undergo two phases: Supervised Fine-Tuning (SFT) and Human Feedback Alignment (such as RLHF). The latter is particularly crucial—it determines the “personality” the model exhibits in its final behavior.

Strategic differences among different labs during alignment training lead to significant style differentiation at the product level. Some teams tend to have the model maintain a neutral, reliable, information-dense style in responses, taking a conservative stance toward expressions with higher “degrees of freedom.” Other teams encourage the model to display linguistic personality, allowing for a certain range of humor, metaphor, and emotional expression. These differences in training preferences, after multiple rounds of reinforcement learning iterations, evolve into a product personality gap that is difficult to bridge.

It is worth noting that the quality of human preference data annotation during alignment training is equally critical. The annotators’ language literacy, sensitivity to expression quality, and the detail of annotation guidelines directly affect the “reward model’s” criteria for judging good and bad responses. A reward model that prefers “complete information but flat sentence structures” during the annotation phase will ultimately produce content that lacks vividness, regardless of how strong the underlying pre-trained model’s language capabilities are. Therefore, alignment training essentially injects a set of “aesthetic standards” into the model, and different teams have different aesthetic standards.

4. Engineering Perspective on Expression: From Model Selection to Application Optimization

From the practical application perspective of developers or enterprises, differences in model expression are not simply a matter of “good or bad,” but rather “whether it matches the application scenario.”

In scenarios requiring extreme accuracy, such as code generation, structured data extraction, and document summarization, the personalization of expression style is unimportant and may even become a distraction. In such tasks, the more direct and stable the model output, the better. Conversely, in highly interactive scenarios like intelligent customer service, creative writing assistance, educational tutoring, and psychological companionship, the delicacy of expression and tone adaptation capabilities directly impact user experience.

When integrating in practice, developers can evaluate and tune models across the following dimensions:

• Prompt Design Adaptation: For models with relatively conservative expression, explicitly specify the style in the prompt (e.g., “explain in a natural, colloquial way,” “add analogies and metaphors”) to compensate for the model’s default expressive shortcomings.
• Decoding Parameter Tuning: Appropriately increasing the temperature parameter can enhance expressive diversity; lowering frequency penalty and presence penalty can prevent the model from over-avoiding common vocabulary, making the output closer to daily language habits.
• Few-Shot Example Guidance: Providing high-quality Q&A pairs as examples in the context can effectively guide the model to imitate specific expressive styles. The selection of examples should deliberately cover three levels: sentence structure, word choice, and rhythm.
• Multi-Turn Context Management: As the number of dialogue turns increases, the model tends to forget the initially set style requirements. Re-emphasizing style preferences periodically or at key nodes can maintain expressive consistency.

It should be noted that while the aforementioned methods can partially narrow the gap in expression quality between models, fundamental differences in corpus composition and training strategies cannot be entirely erased through inference-phase means. This is akin to a writer trained extensively in official document writing; even if repeatedly asked to “write more playfully,” their naturally flowing draft style will still bear the traces of their training.

5. Conclusion and Future Outlook

The differences in language model expression capabilities are rooted in every stage from data collection, preprocessing, and pre-training to alignment training. The introduction of book corpora provides models with deeper linguistic heritage, Theory of Mind capabilities endows models with the intellectual foundation to understand context, and human feedback alignment determines the model’s ultimate “personality base.” These three threads intertwine to collectively shape the expression quality users ultimately feel in the chat box.

For technical teams, rather than simply judging whether a certain model “expresses well or not,” it is better to deeply understand the tendencies and trade-offs of its underlying training, and then guide the model’s expressive capabilities toward the target direction through reasonable engineering means. With the advancement of multimodal training and larger-scale book digitization initiatives, future differentiation among models in the expression dimension will become more diverse. Understanding the connotations of these differences will be a key prerequisite for effectively utilizing language models.