Zero-Shot Learning

Zero-shot learning means getting a model to perform a new task without showing it any examples — only a description of the task. In the pre-LLM era, zero-shot usually meant building semantic mappings between seen and unseen classes (zero-shot image classification, zero-shot translation). In the foundation-model era, zero-shot usually means prompting a language model with a clear instruction and letting its pretrained knowledge do the rest — `classify this ticket as billing, technical, or shipping` with no examples attached.

The capability is remarkable but the ceiling is real. Zero-shot works well when the task is something the pretraining corpus covered extensively (standard classification, common-sense reasoning, well-documented programming tasks) and fails when the task demands domain-specific patterns the model has not seen — proprietary product taxonomies, internal ticket-routing rules, brand-specific tone guidelines. The cure is usually **few-shot prompting** (two to five demonstrations) or **fine-tuning / DPO** for truly custom behaviour.

For APAC mid-market, zero-shot is the right starting point for every new use case. Write the prompt, measure the output on a small eval set, and only invest in few-shot examples or fine-tuning when the zero-shot baseline misses the bar. Skipping this step and jumping straight to fine-tuning is one of the most common ways teams waste a quarter on work the base model could have done unaided.

The non-obvious operational detail is that **zero-shot quality varies wildly between model versions of the same vendor**. A prompt that worked well on GPT-4 Turbo may underperform on GPT-4o on some tasks and overperform on others; the same is true of Claude Opus versus Sonnet. Always re-baseline zero-shot quality when you upgrade the backing model, or you will silently ship regressions.