The Compliance Dilemma in LLM Creative Writing: From General Optimization to Personalized Expression

    This article examines why general-purpose large language models often conflict with creative writing that depends on a strong personal style. It traces the technical causes, including post-training optimization toward mainstream preferences, and reviews prompt engineering, fine-tuning, and evaluation redesign as practical approaches within existing architectures.

    Large language models (LLMs) have matured considerably in general tasks such as code generation, document organization, and information retrieval. For creators who rely on models for highly non-linear writing with a strong personal voice, however, the “optimal” tendency of general-purpose models can become a constraint on creative expression. Drawing on hands-on experience, this article analyzes the technical roots of this conflict and discusses how the collaborative relationship between models and creators could be redefined within existing architectures.

    1. The “Defensive Writing” Trap of General Optimization

    In recent years, post-training for large language models has widely adopted reinforcement learning from human feedback (RLHF) or similar techniques. The core goal is to make model outputs better match mainstream taste and “standard answers.” This approach works well for factual question answering and instruction following. In creative writing, however, it produces a “defensive” mode. The model tends to choose the safest phrasing, the most grammatically conventional wording, and the least controversial expressions. It also automatically closes logical gaps and may distort the user’s original intent in a “well-meaning” way.

    In one literary scene-writing task, the user asked for the protagonist’s memories to be presented in a fragmented, non-linear narrative. The general-purpose model often filled in the timeline and inserted causal connectives. The result was a text that was structurally complete but had lost the original atmosphere.

    This behavior stems from a “convergence preference” the model learned during training. The model tries to map every input to high-probability regions of the training distribution. Creative writing, however, requires moving beyond those regions into “uncommon but highly expressive” text combinations.

    More seriously, this defensive writing is reinforced in some cloud models. To serve “wish-style” code generation, where users give vague requirements and the model produces runnable code, models were post-trained into a “proactive completion” mode. The model assumes what the user’s goal is and pushes toward the direction it judges most reasonable. When this strategy transfers to text creation, it appears as argumentative behavior: the model keeps asking, confirming, and revising. It looks rigorous, but it interrupts the creator’s train of thought.

    2. Capability Limits and Overfitting: Why Atypical Cognitive Tasks Fail

    LLM intelligence has clear limits. “Generalization” in post-training effectively compresses the model onto a decision manifold suited to most tasks. This manifold covers everyday task forms, such as organizing literature, writing weekly reports, generating summaries, and answering common-sense questions. These tasks have clear goals and relatively fixed evaluation criteria, so models can reach high accuracy with large training budgets.

    When faced with atypical cognitive tasks, such as a poem with leaping imagery, a stream-of-consciousness monologue, or a debate with a strong author bias, models show clear overfitting. It appears in three ways:

    1. Semantic smoothing: Extreme or novel wording is replaced with high-frequency synonyms, which makes the language style converge.
    2. Templated structure: The model applies a formulaic “opening, thesis, evidence, conclusion” structure, even to essays and fiction.
    3. Emotional overload: To satisfy metrics that reward “emotional richness,” the model stacks exaggerated adjectives when describing feelings. This reduces the sense of authenticity.

    The root cause lies in the post-training objective function, which gives insufficient weight to fidelity to the original intent. Most evaluation benchmarks, such as general question answering and instruction following, value the reasonableness of answers over their uniqueness. As a result, reinforcement learning signals do not encourage models to preserve users’ non-mainstream expression. In other words, the model is trained to be a “considerate assistant” rather than a “faithful co-creator.”

    3. Strong Compliance Requirements: Special Constraints in Creative Writing

    The final output of creative writing cannot be measured by a single objective function. Its value lies in unpredictability and personal imprint. Therefore, what creators need from a model is often not “help me write well” but “write in my way.” This is called “strong compliance.” The model must give up self-optimization and strictly follow the style, rhythm, and even grammatical errors specified by the user.

    However, the default inference path of current mainstream models is to “understand intent first, then plan the optimal output.” Even when users repeatedly stress “don’t optimize, copy my wording,” the model still rewrites in its internal processing. For example, given the input “He slammed the door and left, and the impact shook plaster off the wall,” the model may adjust it to “He angrily slammed the door and left, and the enormous force caused the coating on the wall to peel off in pieces.” The original is short and forceful. The rewrite is wordy and changes the instantaneous quality of the action. Such “optimization” helps general tasks but is fatal to creative writing.

    4. The Tension Between General Capability and Domain Specialization: Why “Strong at General Tasks” Does Not Mean “Strong at Writing”

    Some argue that increasing parameter counts and training data can satisfy general and creative tasks at the same time. Practical experience suggests otherwise. The strong general abilities of a model, such as high accuracy and strong logical reasoning, can interfere with its performance on creative tasks. General tasks require the model to reason correctly. Creative tasks require it to “reason incorrectly,” where “incorrect” means breaking conventional grammar, making non-logical leaps, and juxtaposing images.

    In one test, a model was asked to write an inner struggle in monologue form. The output was full of rational analysis, such as “This kind of contradiction stems from a conflict in self-identity.” The creator wanted something like “The person in the mirror splits into two strangers, one laughing, one crying.” The model did not produce the latter because the former appears far more often in its training data and is rated as “deeper.” This statistical bias makes the model a poor imitation of a creative writer.

    5. Practical Approaches: Prompt Engineering, Fine-Tuning, and Evaluation Redesign

    5.1 Prompt Engineering: Deliberately Building an “Anti-General” Context

    Carefully designed prompts can partly suppress the model’s defensive writing tendency. For example:

    • Explicitly ban optimization: State directly in the instruction: “Do not add modifiers, do not change word order, do not add logical connectives.”
    • Provide extreme examples: Give a fragmented, unpunctuated line of poetry as a style reference, so the model imitates that distribution.
    • Limit output length and format: Require short sentences and no paragraph breaks to reduce the model’s urge to add sections automatically.

    The effect of prompt engineering, however, is limited by the model’s highest-probability distribution. When creative requirements deviate too far from the training distribution, the effect of the prompts drops sharply.

    5.2 Fine-Tuning: Targeted Training for Style

    For individuals or teams with long-term creative needs, a more reliable method is domain fine-tuning on a small, high-quality dataset. Key steps include:

    1. Collect the creator’s past works (at least 100,000 characters) as style samples.
    2. Build triples of “request, reference text, expected output.” The expected output must be original text that has not been rewritten.
    3. Add a “fidelity penalty” to the cross-entropy loss. Compute the token-level edit distance between the model output and the reference text. The higher the distance, the heavier the penalty.

    A model trained this way prioritizes matching the creator’s habits when generating text, rather than optimizing toward a general writing style.

    5.3 Redesigning Evaluation: From Manual Scoring to Divergence Metrics

    To encourage models to move beyond generality, evaluation methods must change. A common industry approach is to have evaluators, human or automated, judge how far the output deviates from the user’s requirements, rather than judging the quality of the output itself. Specific metrics may include:

    • Style similarity: Compare the statistical distance between the output and the user’s sample in word frequency, sentence length, and rhetorical devices.
    • Structural deviation: Check whether the output strictly follows the user-specified paragraphing, punctuation, and point of view.
    • Information retention: Ensure that all key elements the user provided, such as specific nouns, events, and emotions, are retained without added extensions.

    Only when these metrics enter the post-training reward function will models truly learn not to improvise.

    6. Future Outlook: From General Assistant to Configurable Co-Creator

    AI development should not come at the cost of erasing user individuality. An ideal creative model should have two layers of capability. The base layer handles language generation and ensures grammatical correctness and rich vocabulary. The top layer provides control, letting users adjust the balance between creativity and generality through parameters or natural language. For example, a “compliance slider” could work as follows: at 100%, the model only restates the input literally; at 0%, it improvises freely. Intermediate settings would allow limited embellishment within a fixed framework.

    Some models already support rough control of similar effects through system prompts, but they are far from fine-grained adjustment. For cloud service providers, this is both a challenge and an opportunity. If they offer “style binding” and “custom post-training” at the API level, combined with copyright-friendly content compensation mechanisms, they could attract professional creators in large numbers.

    Conclusion

    Creative writing tests a model’s intelligence ceiling and its training objectives at the same time. General-purpose testing measures how well a model fits an existing distribution. Creative scenarios test its willingness to explore the space beyond that distribution. Current models generally lack this capacity for non-compliance, and the root cause lies in the uniformity of post-training strategies. Through combined efforts in prompt engineering, domain fine-tuning, and evaluation metrics, it may be possible to reshape a model’s expressive preferences within existing architectures. In the long run, however, solving this problem will require a paradigm shift at the algorithmic level, moving from seeking the optimal solution to respecting diverse solutions. That is the stance a mature creative tool should adopt.