Large language models (LLMs) often default to single-label classification in zero-shot multi-label tasks—a tendency we term "conservative default". While few-shot prompting mitigates this, it introduces "example bias". We evaluate zero-shot strategies to modulate this tendency using 1,441 healthcare feedback records and two LLMs. We compare instruction-based methods with structural constraints that modify the token generation sequence, specifically an Enum-First format requiring domain enumeration before selection. Results show that structural constraints substantially reduce single-label rates (Magistral: 96% → 19%; Qwen3: 54% → 0.0%), though the latter suggests potential over-correction compared to human baselines (16.7–41.3%). These findings indicate that while output structure is a potent modulator of classification behavior by shifting the decision point upstream, its effect magnitude is model-dependent, necessitating empirical calibration to prevent spurious associations.